A conta da IA: financiamento circular, dívida fora do balanço e o que 2008 já ensinou

Resumo

Entre 2024 e 2026, as maiores empresas de tecnologia dos Estados Unidos passaram a investir em data centers de IA num ritmo que o Banco da Inglaterra descreveu como "unprecedented historically" [1]. O capex de Microsoft, Alphabet, Amazon e Meta foi de US$ 228 bilhões em 2024 para US$ 376 bilhões em 2025, e as previsões para 2026 somam até US$ 745 bilhões. Este artigo examina como esse investimento está sendo pago e quem fica com o risco se a receita demorar a chegar.

A tese é que o ciclo da IA reúne os sinais clássicos de uma bolha de crédito, mesmo com uma tecnologia real por trás. Fornecedores investem nos clientes que compram deles: o FMI contou US$ 1,4 trilhão em compromissos de receita futura em 14 acordos desse tipo desde 2023. Parte da dívida fica fora dos balanços. A Meta construiu um campus de US$ 27 bilhões financiado por títulos de US$ 27,3 bilhões emitidos pela sócia majoritária de uma joint venture que ela não consolida, e as maiores empresas do setor somam mais de US$ 1 trilhão em aluguéis assinados que ainda não começaram. O colateral envelhece rápido: a NVIDIA lança produtos novos todo ano, enquanto as empresas depreciam servidores em cinco ou seis anos. E o risco se concentra em poucos nomes, como OpenAI, Oracle e CoreWeave, e se espalha para fundos, seguradoras e investidores individuais.

O artigo usa a história como evidência de mecanismos, não como previsão: a mania ferroviária britânica e o pânico americano de 1873, a fibra óptica de 2000, a Enron e a crise de 2008. Mostra também o que pode desmentir a tese, como a demanda de nuvem crescendo fora dos laboratórios de IA, balanços ainda fortes e o aluguel de GPUs antigas em alta em 2026, e termina com cenários e uma lista de indicadores para acompanhar. Uma seção trata do Brasil.

O tamanho da aposta

Em 2024, Microsoft, Alphabet, Amazon e Meta gastaram juntas cerca de US$ 228 bilhões em capex, as compras de prédios, máquinas e equipamentos registradas no fluxo de caixa. Em 2025, foram US$ 376 bilhões, 65% a mais. Só no primeiro semestre de 2026, US$ 295 bilhões, quase 80% do ano anterior inteiro. As contas são minhas, sobre os números enviados pelas empresas à SEC, com a Microsoft convertida para o ano-calendário a partir dos relatórios trimestrais [2], [3], [4], [5], [6], [7], [8], [9]. Somando a Oracle, cujo capex mais que dobrou no ano fiscal encerrado em maio de 2026, o ritmo anual das cinco passa de US$ 430 bilhões [10].

A tabela mostra o capex de cada empresa, em bilhões de dólares, e a previsão mais recente para 2026. Microsoft e Oracle usam anos fiscais que não coincidem com o calendário, e as previsões usam definições diferentes: a da Microsoft inclui leases financeiros, e a da Meta, o principal pago nesses leases [11], [12], [13], [14], [15], [16].

Somadas, as previsões de Microsoft, Alphabet, Amazon e Meta para 2026 ficam entre US$ 720 bilhões e US$ 745 bilhões, quase o dobro de 2025. A agência de rating Moody's, numa nota de julho de 2026 citada pela CNBC, projetou capex de US$ 785 bilhões em 2026 para seis empresas (as quatro, mais Oracle e CoreWeave) e de cerca de US$ 1 trilhão em 2027 [17]. A Oracle, sozinha, gastou US$ 28,5 bilhões no trimestre de junho a agosto de 2026, mais que em todo o ano fiscal de 2025 [18].

O que o dinheiro compra

A maior parte do gasto vai para equipamento de vida curta. A Alphabet informou que, em 2025, "approximately 60% of that investment in servers and 40% in data centers and networking equipment" [19]. A Microsoft disse, sobre o trimestre de abril a junho de 2026: "Roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs" [13]. É esse equipamento que se desvaloriza no ritmo descrito na seção "A física da depreciação".

Parte do aumento é preço, não quantidade. Em abril de 2026, a Microsoft previu cerca de US$ 190 bilhões de capex para o ano, incluindo "approximately $25 billion from the impact of higher component pricing" [20]. A Meta elevou a previsão no mesmo mês e explicou que a revisão "reflects our expectations for higher component pricing this year" [21]. A Amazon subiu a sua para US$ 220 bilhões em julho, e o presidente-executivo, Andy Jassy, atribuiu a alta ao preço da memória [15]. A IEA registra uma falta de memória de alta largura de banda, a HBM usada nos chips de IA, que deve "persist through at least the end of 2027" [22]. Num ciclo de investimento, quem controla o componente escasso tende a ficar com a margem, e em 2026 o componente escasso é a memória.

As empresas defendem a conta com a demanda. A Microsoft disse estar "confident in the return on these investments given higher demand signals and increasing product usage" [20]. Jassy foi mais longe: "But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027 too," e "In fact, the demand we already have for 2028 is striking." O backlog da AWS, os contratos assinados e ainda não entregues, chegou a US$ 496 bilhões [15]. Esses números são o melhor argumento contra a tese deste artigo, e voltam na seção "O que pode desmentir a tese". Mas os maiores contratos de nuvem assinados em 2025 e 2026 têm do outro lado laboratórios de IA que ainda queimam caixa e dependem de novas rodadas de financiamento, como mostra a seção "O círculo".

O tamanho na economia

Pelas minhas contas, o capex das quatro maiores em 2025 equivaleu a cerca de 1,2% do PIB nominal dos Estados Unidos; as previsões para 2026 levam a algo entre 2,2% e 2,3% [23]. Nem todo esse dinheiro é gasto no país, porque há data centers no exterior, e boa parte do equipamento é importada, como mostra a seção sobre macroeconomia. Ainda assim, a ordem de grandeza impressiona. Segundo a IEA, "Capital expenditure of just five technology companies is now larger than global investment in oil and natural gas production." [22]. Em proporção da economia, o ciclo ainda está longe da mania ferroviária britânica, quando o investimento em ferrovias chegou a 7,3% do PIB em 1847 [24]. O que distingue o ciclo atual é a concentração: quatro ou cinco empresas respondem pela maior parte do gasto, e as maiores empresas de tecnologia, entre elas as quatro que mais investem, somam mais de um terço do S&P 500, como mostra a seção sobre macroeconomia.

O caixa já não cobre tudo

Até pouco tempo atrás, as grandes de tecnologia pagavam o investimento com o caixa que geravam, e sobrava dinheiro. A folga diminuiu. A tabela compara o caixa gerado pelas operações com o capex, em bilhões de dólares, pelas minhas contas sobre os números enviados à SEC:

As próprias empresas registram a virada. A Amazon informou em julho de 2026: "Free cash flow decreased to an outflow of $7.6 billion for the trailing twelve months, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment, net of proceeds from sales and incentives. This increase primarily reflects investments in artificial intelligence." [25]. A Alphabet, no mesmo mês: "We had negative free cash flow of $5.9 billion in the second quarter, driven by our investments in CapEx." [14].

Quando o capex passa do caixa, a diferença vem de dívida, de emissão de ações ou de contratos de aluguel de longo prazo. A Moody's resumiu a mudança, segundo a CNBC: "Previously, these companies relied on asset-light structures centered on software, intellectual property, and scalable cloud services that required modest capital investment," e "The transition from asset-light to asset-heavy models requires unprecedented levels of investment and capital raising." A agência estimou a dívida direta das seis empresas em cerca de US$ 460 bilhões e seus compromissos de aluguel em US$ 1,2 trilhão, dos quais mais de US$ 820 bilhões em contratos que ainda não começaram, e disse tratar esses compromissos como equivalentes a dívida [17]. A seção "Fora do balanço" mostra de onde vêm esses números.

Nada disso, sozinho, prova uma bolha. Empresas lucrativas investindo pesado numa tecnologia nova é exatamente o que se espera de uma revolução tecnológica. A literatura sobre bolhas manda olhar para outra coisa: como o investimento é pago e quem fica com o risco se a receita não vier. É o assunto da próxima seção.

Como se reconhece uma bolha

"Bolha" virou palavra de briga. Quem acha que a IA está cara usa o termo como xingamento; quem discorda responde que a tecnologia é real. As duas coisas podem ser verdade ao mesmo tempo. Charles Kindleberger e Robert Aliber, autores do livro de referência sobre crises financeiras, separam os termos com mais cuidado: "The term mania describes the frenzied pattern of purchases, often an increase in prices accompanied by an increase in trading volumes; individuals are eager to buy before the prices increase further. The term bubble suggests that when the prices stop increasing, they are likely—indeed almost certain—to decline." [26]. Na mesma obra há uma ressalva que convém manter à vista durante o artigo inteiro: "Virtually every mania is associated with a robust economic expansion, but only a few economic expansions are associated with a mania." [26].

O roteiro que eles descrevem começa por um choque real. Kindleberger e Aliber resumem assim o modelo de Hyman Minsky: "Minsky argued that the events that lead to a crisis start with a ‘displacement’, some exogenous, outside shock to the macroeconomic system. If the shock was sufficiently large and pervasive, the economic outlook and the anticipated profit opportunities would improve in at least one important sector of the economy." [27]. Depois vem a euforia, em que "investment spending surges because credit is plentiful" e o dinheiro parece de graça ("money always seems free in manias"). Em seguida, uma pausa na alta, a venda forçada de quem comprou com dinheiro emprestado e, às vezes, o pânico [26]. A tese central do livro não é psicológica. É uma tese sobre crédito: "the cycle of manias and panics results from the pro-cyclical changes in the supply of credit" [26].

Minsky deu a essa história um vocabulário que serve para ler balanços. Ele divide quem toma dinheiro em três tipos. As unidades hedge "can fulfill all of their contractual payment obligations by their cash flows". As especulativas pagam os juros com a receita, mas precisam rolar o principal: "Such units need to "roll over" their liabilities". Nas unidades Ponzi, o caixa das operações não cobre nem os juros, e elas "can sell assets or borrow" para continuar de pé [28]. O segundo teorema da hipótese de instabilidade financeira é a frase que mais interessa aqui: "over periods of prolonged prosperity, the economy transits from financial relations that make for a stable system to financial relations that make for an unstable system" [28].

Aplicada ao ciclo da IA, a taxonomia separa empresas que o debate costuma misturar. As grandes plataformas que pagam o investimento com o caixa das próprias operações são unidades hedge pelo critério de Minsky, mesmo quando emitem dívida. As empresas de nuvem especializadas em GPU, que compram chips com empréstimos a serem pagos por contratos futuros e refinanciamentos, estão mais perto do perfil especulativo. E uma empresa cujos compromissos de compra somam várias vezes a própria receita, e que depende de novas rodadas de capital para honrá-los, entra no território que Minsky chamou de Ponzi. O termo é técnico e não acusa ninguém de fraude. Descreve uma estrutura que só continua de pé enquanto o financiamento continua chegando. As seções "O círculo" e "Os elos mais fracos" mostram onde cada empresa está.

Minsky também apontou quem empurra o sistema de um regime para o outro: "bankers (using the term generically for all intermediaries in finance), whether they be brokers or dealers, are merchants of debt who strive to innovate in the assets they acquire and the liabilities they market." [28]. Em 2007, a inovação estava nos veículos de securitização. No ciclo da IA, está nas sociedades de propósito específico que constroem data centers, nos empréstimos lastreados em GPU e nos fundos de crédito privado que compram essa dívida. A forma muda; o papel é o mesmo.

A economista Carlota Perez acrescenta a peça que falta para entender bolhas de tecnologia. Na leitura dela, cada revolução tecnológica tem um período de instalação, de 20 a 30 anos, comandado pelo capital financeiro e encerrado por uma grande bolha. Depois de um ponto de virada vem o período de implantação, comandado pelo capital produtivo [29]. A bolha não é acidente: "A major technology bubble (MTB) is not an accidental event. It regularly occurs midway along the process of assimilation of each technological revolution." [30].

O ponto mais incômodo da teoria de Perez é que a bolha cumpre uma função. "It is the switch to short-term gains during the bubble that attracts the necessary capital to be poured into the infrastructural networks of each revolution." [30]. As ferrovias e a fibra óptica foram construídas em excesso com dinheiro que esperava retorno rápido e acabaram usadas por outros, mais tarde. Ela descreve o espírito do auge em duas frases: "Soon there is more capital wanting a piece of the action than projects looking for funds." e "The illusion is that there are high profits to be had with very low risk." [30]. E dá dois critérios para reconhecer uma grande bolha tecnológica: "The two defining characteristics of these major technology bubbles are: (a) their concentration on the new technologies –especially the new infrastructural networks– and (b) their decoupling from the real economy." [30].

Perez separa esse tipo de bolha das bolhas de crédito descritas por Minsky: "They are the result of opportunity pull rather than of easy credit push. But they are indeed bubbles." [30]. Essa distinção é o eixo deste artigo. O ciclo da IA começou como uma bolha de oportunidade, financiada sobretudo com caixa e com ações. A partir de 2025, ganhou as marcas de uma bolha de crédito. O FMI registrou que as hyperscalers levantaram mais de US$ 100 bilhões em títulos de dívida desde janeiro de 2025, "supplemented by leveraged loans and intercorporate arrangements that are, at times, part of circular financing structures" [31]. É essa mistura que torna 2008 relevante.

Robert Shiller, que estudou bolhas a vida inteira, prefere uma definição comportamental: "A situation in which news of price increases spurs investor enthusiasm which spreads by psychological contagion from person to person, in the process amplifying stories that might justify the price increase and bringing in a larger and larger class of investors, who, despite doubts about the real value of the investment, are drawn to it partly through envy of others’ successes and partly through a gambler’s excitement." [32]. O indicador que leva o nome dele, o CAPE, divide o preço do índice S&P pela média de dez anos dos lucros reais. Na série de Shiller atualizada em 5 de outubro de 2026, o CAPE chegou a 44,20 em dezembro de 1999, o recorde, e a 27,55 em maio de 2007. Em agosto de 2026 marcou 40,99, a leitura mais alta desde 2000, e a prévia de outubro de 2026 está em 40,71. A média desde 1881 é 17,80 [33].

Reconhecer uma bolha cedo não significa acertar quando ela estoura. Em 5 de dezembro de 1996, o presidente do Federal Reserve, Alan Greenspan, perguntou em público: "But how do we know when irrational exuberance has unduly escalated asset values, which then become subject to unexpected and prolonged contractions as they have in Japan over the past decade?" [34]. O CAPE estava em 27,72 naquele mês e subiu por mais três anos, até 44,20 em dezembro de 1999 [33].

O próprio Shiller avisa que o indicador não marca data e que bolhas nem sempre estouram de uma vez: "Though the abrupt ends of stock market booms in 1929, 2000 and 2007 might seem consonant with such a metaphor, these booms were reflated again before long" [32]. Andrew Odlyzko, que estudou as manias ferroviárias, documenta um caso que desmente a regra: a mania ferroviária britânica dos anos 1830, especulativa e lucrativa para quem investiu. A conclusão dele serve de freio para este texto: "The example of the railway mania of the 1830s serves as a useful antidote to claims that bubbles are easy to detect or that all large and quick jumps in asset valuations are irrational." [35].

Daqui em diante, o artigo usa seis marcadores, todos tirados dessa literatura e todos verificáveis com dados públicos:

Nenhum marcador, sozinho, prova bolha. A tese deste artigo é que, em 2026, os seis aparecem ao mesmo tempo, e que três deles (crédito, inovação nos passivos e receita circular) são justamente os que separaram uma correção de mercado de uma crise financeira nos episódios que vêm a seguir.

Ferrovias: a infraestrutura fica, o investidor paga

A comparação mais citada para a IA é a internet de 2000, mas a mais útil talvez seja mais velha. Na Grã-Bretanha dos anos 1840, as ferrovias eram a tecnologia que ia mudar tudo, e mudaram. Cada linha precisava de autorização do Parlamento, o que deixou um registro anual da euforia. Foram 91 milhas aprovadas em 1843, 805 em 1844, 2.700 em 1845 e um recorde de 4.538 em 1846. Em 1847, 1.354; em 1850, só 8 [24].

O tamanho do investimento é o dado que mais conversa com 2026. Segundo Odlyzko, o gasto com ferrovias "peaked in 1847, at a level of about £44 million (in contrast to about £50 million for the national government budget)". Pela conta dele, "railway investments of £44 million in 1847 came to 7.3% of a GDP of £600 million, which is comparable to $1.1 trillion for the U.S. today", e o "today" do texto é 2010. Na tabela do mesmo trabalho, as ferrovias respondem por £43,9 milhões dos £88 milhões investidos no país em 1847, quase metade de todo o investimento britânico [24]. A comparação mais forte do autor é com o orçamento militar: "At its height, individual capitalists, in pursuit of private profit, were plowing more than twice as much into the construction of a public infrastructure as their nation was spending on the military." [24].

O investimento não parou quando as ações caíram. Odlyzko explica por quê: "Even in 1850, when investors felt crushed by the stock market decline, investments were high, higher than they had ever been before 1845. The reason was not that any shareholders were enthusiastic about the spending. Rather, so much money had been spent that it was felt better to go on and complete the lines and earn small returns than abandon them completely and get nothing." [24]. É a lógica do custo afundado, e ela importa para a IA. Um campus de data center pela metade não vale nada, e por isso o capex tende a continuar alto por algum tempo depois de a confiança acabar.

Para quem pôs dinheiro, o resultado foi ruim. "The best known index of railway shares prices declined from 167.9 in the summer of 1845 to 60.5 at the end of 1849, a drop of 64%." [36]. A Economist resumiu o paradoxo em 1855, numa frase que Odlyzko reproduz: "[m]echanically or scientifically, the railways, with all their multiplied conveniences and contrivances, are an honour to our age and country: commercially, they are great failures." [24]. Charlotte Brontë, que tinha ações de uma das linhas, escreveu em 1849, também segundo Odlyzko: "The original price of Shares in this Railway was £50. At one time they rose to 120; and for some years gave a dividend of 10 per cent; they are now down at 20, and it is doubtful whether any dividend will be declared this half-year." [24].

A explicação popular põe a culpa em fraudadores, como George Hudson, que maquiava balanços. Odlyzko discorda: "accounting fraud was not the fundamental cause of the investment disaster of the Railway Mania. Distorted financials helped mask for a while the disappointing returns that new railways produced for investors, but were not the reason for the lack of profits. The main problem was the headlong expansion during the Mania. It tripled British railway mileage, and the new lines cost more than expected and brought revenues lower than expected." [36]. A receita anual por milha caiu de £3.470 em 1845 para £2.080 em 1850 [36]. A lição vale para hoje: o problema de uma bolha de infraestrutura costuma ser aritmético, não moral. Capacidade demais, construída cara demais, para uma demanda que cresce, mas mais devagar do que o prometido.

Kindleberger descreve como o perfil de quem investe muda ao longo do ciclo: "In early railway investment, for example, initial investors were persons doing business along the rights of way who sought benefits from the railroad for their other concerns. They were followed by a second group of investors interested in the profits the railroad would earn, and by a third group, made up of speculators who, seeing the rise in the railroad’s shares, borrowed money or paid for the initial instalments with no intention of completing the purchase, to make a profit on resale." [37]. Na IA, o primeiro grupo são as empresas que constroem capacidade para os próprios produtos. O segundo são os que constroem para vender ou alugar capacidade a terceiros. O terceiro aparece quando o dinheiro passa a entrar porque o preço sobe, e não porque o fluxo de caixa justifica.

Perez registra um detalhe da mania ferroviária que lembra o Ponzi de Minsky: "During the railway mania in the 1840s, though, companies without profits did pay dividends out of the new capital they attracted in order to keep previous investors in and new investors coming." [30]. Quando o dinheiro novo paga quem chegou antes, a estrutura depende de o dinheiro novo continuar entrando.

Odlyzko liga a mania vitoriana à de 2000 de modo explícito. Para ele, os dois maiores desastres de investimento real da bolha da internet foram a fibra de longa distância nos EUA e os leilões europeus de espectro 3G, "each of which cost on the order of $100 billion to €100 billion" [24]. Nos dois casos, "there was hard quantitative evidence that once those went into service, there would not be enough demand to provide the expected profits." [24]. E a frase que ele usa para os dois episódios serve para o terceiro: "society decided to adopt the new offerings at its own speed, slower than that expected by inventors, promoters, and investors, and financial ruin followed in both cases." [24].

Há um contraponto, e é do mesmo autor. A mania ferroviária dos anos 1830, menor e anterior, deu lucro a quem investiu. Odlyzko também separa infraestrutura financiada com ações de crises de dívida: "One could argue that the infrastructure construction funded by equity is radically different from the much more frequent, far more dramatic and attention-grabbing financial crises that are associated with large accumulations of debt, typically for consumption." [35]. Essa distinção decide o tamanho do estrago. Quando a conta fica com acionistas, a perda é dolorosa, mas não se espalha como numa crise bancária. Quando fica com credores alavancados, se espalha.

1873: quando a ferrovia é paga com dívida

Do outro lado do Atlântico, a história teve outro desfecho, porque o dinheiro era outro. A crise americana de 1873, nas palavras de historiadores do Federal Reserve, "arose from investments in railroads. Railroads had expanded rapidly in the nineteenth century, and investors in many early projects had earned high returns. As the Gilded Age progressed, investment in railroads continued, but new projects outpaced demand for new capacity, and returns on railroad investments declined." [38]. Quando quebras na bolsa de Viena levaram investidores europeus a vender títulos de ferrovias americanas, as empresas ficaram sem dinheiro para operar e para refinanciar dívidas. O banco Jay Cooke & Co., muito exposto à Northern Pacific, quebrou em 18 de setembro de 1873. Dois dias depois, a Bolsa de Nova York fechou pela primeira vez na história e ficou dez dias sem pregão, e pelo menos cem bancos quebraram no país [38].

A diferença em relação à mania britânica estava no passivo. "Railroads were built mainly with borrowed money", escrevem Kay Giesecke, Francis Longstaff, Stephen Schaefer e Ilya Strebulaev num estudo de 150 anos de calotes de títulos corporativos. Entre 1873 e 1875, os calotes somaram "36 percent of the par value of the entire corporate bond market"; o pior triênio da Grande Depressão ficou em 12,88% [39]. A contração que veio depois durou 65 meses, de outubro de 1873 a março de 1879, a mais longa da cronologia do NBER [40]. Os próprios autores fazem a ponte com os ciclos recentes: os booms de tecnologia financiados com dívida no século 19 foram seguidos de crises "closely resembling patterns in the telecommunications and structured-credit-products cycles of the past decade." [39].

Também aqui há contraponto. O historiador econômico Scott Mixon mostra que, antes de 1873, ações, opções e títulos já precificavam o risco. Os investidores "were not caught up in an irrational, speculative mania", e a crise começou pelas ferrovias mais alavancadas [41]. Não faltou aviso. O problema foi a dívida concentrada nos elos mais fracos, que é justamente a pergunta da seção "Os elos mais fracos".

A comparação de durações resume o argumento deste artigo. Segundo o NBER, a recessão que seguiu o estouro da internet, financiada sobretudo com ações, durou 8 meses, de março a novembro de 2001. A de 2007–2009, uma crise de crédito, durou 18 meses. A de 1873, uma crise de títulos de ferrovia, durou 65 [40]. O tamanho do estrago depende menos da tecnologia do que de quem está segurando a dívida.

As ferrovias deixam duas lições para a IA. A infraestrutura sobrevive à bolha: as linhas da mania foram terminadas e continuaram em uso depois que as ações afundaram, e os prédios, subestações e linhas de transmissão dos data centers de hoje também vão servir a alguém. Mas uma ferrovia serve por décadas, e o chip que justifica o investimento dura poucos anos, como mostra a seção sobre depreciação. E até 2024 o ciclo da IA era pago sobretudo com caixa e ações, como as manias britânicas; desde 2025 está migrando para a dívida, como a expansão ferroviária americana que terminou no pânico de 1873.

Telecom, 1996–2002: fibra apagada e receita de ida e volta

Se as ferrovias mostram a aritmética de uma bolha de infraestrutura, o telecom do fim dos anos 1990 mostra a contabilidade. Foi o último ciclo em que uma tecnologia de rede atraiu capital nessa escala, e quase todos os mecanismos discutidos neste artigo apareceram ali: demanda medida pela capacidade, fornecedores financiando clientes, vendas cruzadas entre concorrentes e uma melhora técnica que multiplicou a oferta mais rápido do que o mercado conseguia usar.

O investimento foi grande e rápido. Economistas do Federal Reserve de Richmond calcularam que, entre o primeiro trimestre de 1996 e o quarto de 2000, "investment in communications equipment grew from approximately $62 billion per year to over $135 billion per year in constant 1996 dollars", um crescimento médio de quase 18% ao ano. A queda foi igualmente rápida: no quarto trimestre de 2001, o número ficou abaixo de US$ 93 bilhões, "only 69 percent of the same figure one year earlier", e a participação dos equipamentos de comunicação no investimento privado caiu de quase 7% em 2000 para 4,8% no fim de 2002 [42].

O mito que justificava o investimento

A história que sustentava esse capex era uma estatística. Andrew Odlyzko a descreve assim: "The most popular and extremely misleading myths of the dot-com and telecom bubbles was that “Internet traffic doubles every 100 days” (or 3 months, or 4 months)." [43]. O número real, segundo ele, era bem menor e mesmo assim enorme: o tráfego estava "close to doubling each year since 1997". A fonte principal do mito tinha nome: "At its base, most of the support for the myth of astronomical growth rates in Internet traffic was coming from WorldCom." [43]. Numa apresentação de março de 2000, Bernie Ebbers, presidente da WorldCom, disse que, por causa da demanda, o capex da empresa em 2003 "would have to exceed $100 billion" [43].

O detalhe que mais interessa para 2026 é a confusão que estava por trás do mito: "most of the WorldCom claims about astronomical growth rates of their network were about network capacity, not traffic." [43]. Capacidade instalada é uma decisão de quem investe. Tráfego é uma decisão de quem usa. Quando a primeira passa a ser apresentada como se fosse a segunda, o mercado perde o termômetro. A seção "O círculo" mostra que boa parte dos números de demanda da IA, como gigawatts contratados e carteiras de contratos, mede compromissos de capacidade, não uso.

Fibra demais, e cada fio carregando mais

O último levantamento anual da FCC sobre fibra, com dados até o fim de 1998, já mostrava o desenho. As operadoras antigas usavam boa parte da fibra que tinham: 50% na AT&T, 85% na Sprint. As novas, que construíam redes inteiras do zero, quase não acendiam a delas. A Qwest tinha 2% da fibra acesa em 1998, a Williams 1%, a GST 2% e a NEON 5%. Só naquele ano, a fibra instalada pelas operadoras de longa distância cresceu mais de 30%, para mais de 3,6 milhões de milhas de fibra [44]. O relatório da FCC explicava a lógica do excesso numa frase: "Although up-front costs for fiber deployment are high, a significant portion of the total investment can be deferred until actual demand materializes." [44]. Os economistas de Richmond acrescentam que as quatro grandes de 1996 tinham 72% da fibra de longa distância dos EUA e, em 1999, só 30%, enquanto a instalação anual de fibra se multiplicava por mais de quatro [42].

Mark Doms, do Federal Reserve de São Francisco, mediu a corrida de outro jeito: "in 1996 there were 143,000 route miles of fiber in the country, but by 2000 this had increased to 320,000." Em 1990, quatro empresas tinham redes nacionais de fibra; em 2000, dez. "Chicago and Minneapolis are each served by twenty long-haul firms", contra cerca de cinco em 1990 [45].

E cada fio passou a carregar muito mais. "In 1996, a strand of fiber transmitted data at approximately 2.5 gigabits per second (Gbps). By 2000, the capacity of the same fiber could reach 100 Gbps.", graças à multiplexação por comprimento de onda; a capacidade de uma fibra dobrou "approximately every nine months between 1996 and 2000", mais rápido que a lei de Moore [42]. Os autores listam três causas do excesso: o exagero típico de uma indústria nova, uma capacidade por fio que cresceu mais do que o previsto e, "perhaps most importantly", uma demanda que "had not grown as fast as many had forecast: the pessimists turned out to be right." [42].

Esse é o paralelo técnico mais direto com a IA. Quando a eficiência sobe mais rápido que a demanda, a capacidade física construída para a demanda prevista sobra, mesmo que a demanda continue crescendo. Na IA, a eficiência vem de chips novos a cada ano e de modelos que entregam o mesmo resultado com menos computação. A seção sobre depreciação trata disso com números.

O preço da conta

O índice de ações de telecom da Nasdaq foi de 198 em abril de 1997 a 1.230,06 em 10 de março de 2000 e caiu 62% até o fim daquele ano. Em 16 de maio de 2003 estava em 136. Só as empresas de telecom do S&P perderam cerca de US$ 700 bilhões em valor de mercado entre 2000 e 2002, "more than 3.5 percent of the entire value of U.S. corporate equities at the stock market peak in 2000" [42]. O índice Nasdaq Composite fechou em 5.048,62 no mesmo 10 de março de 2000 e em 1.114,11 em 9 de outubro de 2002, uma queda de 77,9%. Só voltou a fechar acima do pico de 2000 em 23 de abril de 2015, mais de quinze anos depois [46].

O Banco de Compensações Internacionais (BIS) explicou em 2002 por que os bancos saíram relativamente inteiros: "a considerable amount of the financing for the most risky elements of the late 1990s boom was obtained from outside the banking system. This is perhaps best illustrated by the fact that firms in the technology and telecommunications sectors relied heavily on vendor financing, venture capital and the equity and bond markets for their financial needs." [47]. É a mesma composição de financiamento que aparece hoje, com outros nomes: crédito de fornecedor, capital de risco, ações e títulos.

O fornecedor que empresta para o cliente comprar dele

Os balanços dos fabricantes de equipamentos de 2001 registram o mecanismo que voltaria em 2025. A Lucent explicava no relatório anual: "Some of our customers worldwide are requiring their suppliers to arrange or provide long-term financing for them as a condition of obtaining or bidding on infrastructure projects. These projects may require financing in amounts ranging from modest sums to more than a billion dollars." [48]. Os compromissos totais de financiamento a clientes da Lucent eram de US$ 8,1 bilhões em 30 de setembro de 2000. No ano fiscal de 2001, a empresa lançou US$ 2,2 bilhões em provisões para créditos duvidosos e financiamentos a clientes, cerca de US$ 1,3 bilhão em três projetos, entre eles a Winstar, que pediu recuperação judicial em abril de 2001 [48]. Entre os fatores de risco, uma frase que valeria para qualquer fornecedor que hoje investe nos próprios clientes: "We are vulnerable to downturns in our customers' businesses as a result of our customer financing arrangements." [48].

A Nortel contou a mesma história. "As a result of the competitive environment during 1999 and 2000, we provided significant amounts of medium-term and long-term customer financing" [49]. O total de financiamento a clientes caiu de US$ 5,168 bilhões no fim de 2000 para US$ 2,075 bilhões no fim de 2001, porque "our ability to place customer financing with third-party lenders was significantly reduced during 2001". O prejuízo líquido da Nortel em 2001 foi de US$ 27,3 bilhões [49]. A Cisco tratava o crédito como arma de venda: "We believe customer financing is a competitive factor in obtaining business, particularly in supplying customers involved in significant infrastructure projects." Seus empréstimos podiam incluir "additional funds for soft costs associated with network installation and integration of our products and for working capital purposes" [50].

Em outubro de 2026, economistas do BIS usaram exatamente esse episódio para explicar o risco das relações circulares na IA: "upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors' equipment [...]. This meant that part of the equipment vendors' reported sales was being funded by the vendors themselves." [51].

Receita de ida e volta

Em alguns casos, a circularidade foi além do financiamento e virou receita fabricada. Em 2004, a SEC acusou a Qwest de ter reconhecido de forma fraudulenta mais de US$ 3,8 bilhões em receitas entre 1999 e 2002. Parte delas veio da venda de direitos de uso de fibra (IRUs) e, quando a demanda caiu, de trocas: "When the demand for IRUs declined, Qwest engaged in IRU "swaps" whereby Qwest bought IRUs from other companies in exchange for agreements from those companies to buy IRUs from Qwest." [52]. Segundo a SEC, o recurso virou rotina a ponto de funcionários compararem a prática a um vício e chamarem as vendas não recorrentes de "heroin". A Qwest fez acordo sem admitir nem negar as acusações e pagou multa de US$ 250 milhões [52].

A Global Crossing é o caso mais didático, porque a SEC mediu a circularidade. No primeiro semestre de 2001, as vendas de capacidade da empresa passaram a depender de compras simultâneas das mesmas operadoras: "GX’s ability to sell capacity often became linked to, and in some cases dependent on, GX’s simultaneous purchase of a nearly equal dollar amount of capacity." No segundo trimestre de 2001, essas "reciprocal transactions" foram US$ 516 milhões dos US$ 567 milhões em vendas de IRUs, ou 91%, e sem elas o "Recurring Adjusted EBITDA" da empresa "would have gone from the reported $472 million to a negative $43 million." [53]. A Global Crossing pediu recuperação judicial em 28 de janeiro de 2002. O caso foi de divulgação, encerrado por acordo, e a SEC não acusou fraude [53].

A WorldCom, a fonte do mito do tráfego, acabou no caso mais simples. Em junho de 2002, a SEC a acusou de uma fraude contábil de "more than $3.8 billion", feita ao lançar como investimento, e não como despesa, os valores pagos a outras operadoras pelo uso de suas redes [54]. A empresa pediu recuperação judicial em 21 de julho de 2002 [55].

O contraponto

Nem todo mundo vê bolha no telecom. Os economistas de Richmond, escrevendo em 2003, chamaram o ciclo de "in large part—a rational response to the changing fundamentals of technology and regulatory environment", com grandes erros de previsão num ambiente de incerteza: "our explanation for the telecommunications boom and bust does not involve fraud, irrationality, or a bubble. To be sure, as the bust became apparent, fraud did occur. But it is not clear that fraud played an important role in the boom and the early stages of the bust." [42]. O argumento é sério, e vale para a IA: dá para errar a previsão de demanda sem fraude nenhuma. A fraude aparece depois, quando os números precisam fechar.

Doms acrescenta um argumento técnico que vale guardar para os data centers: "the marginal cost of fiber is small relative to the cost of laying a new fiber route; in other words, it would be much cheaper to lay additional strands of fiber today while trenches are open than to dig up the trenches again several years from now and lay some additional fiber lines. So, although the utilization rate is low, that does not imply that firms were irrational in the amount of fiber they installed." Mesmo assim, ele reconhece que o investimento em tecnologia de 1999 e 2000 ficou muito acima do que um modelo previa, porque "expectations were too high, especially in two sectors of the economy, telecommunications services and the dot-com sector." [45].

O telecom deixa quatro sinais para procurar em 2026, e nenhum deles exige supor má-fé. Demanda medida por capacidade contratada. Vendas financiadas pelo próprio vendedor. Contratos cruzados entre empresas que são clientes e fornecedoras umas das outras. E uma tecnologia que multiplica a oferta mais rápido do que o uso. As próximas seções mostram que os quatro estão presentes no ciclo da IA, e em escala maior.

Enron: a regra que nasceu de um escândalo

A Enron entra neste artigo por dois motivos que não têm a ver com a fraude em si. O primeiro é o motivo econômico que levou a empresa a tirar dívida do balanço, porque ele é o mesmo de qualquer empresa que precisa investir muito antes de receber. O segundo é que as regras contábeis que hoje decidem se um data center aparece ou não no balanço de uma big tech nasceram da investigação desse caso.

O relatório do comitê especial do conselho da Enron, conhecido como Relatório Powers, descreve o problema sem adjetivos: "Much of this growth involved large initial capital investments that were not expected to generate significant earnings or cash flow in the short term. While Enron believed these investments would be beneficial over a period of time, they placed immediate pressure on Enron's balance sheet. Enron already had a substantial debt load. Funding the new investments by issuing additional debt was unattractive because cash flow in the early years would be insufficient to service that debt and would place pressure on Enron's credit ratings. Maintaining Enron's credit ratings at investment grade was vital to the conduct of its energy trading business." [56].

A solução foi trazer sócios de fora para entidades separadas, que podiam tomar empréstimos em nome próprio, "although in many cases a guaranty or other form of credit support was required from Enron". E o relatório explica por que a administração preferia essa forma: "Enron management preferred the latter treatment--known as "off-balance-sheet"--because it would enable Enron to present itself more attractively as measured by the ratios favored by Wall Street analysts and rating agencies." [56]. Investimento grande, retorno distante e um rating a proteger: o motivo é idêntico ao de um campus de data center de US$ 27 bilhões, e a seção "Fora do balanço" mostra como ele é resolvido hoje dentro das regras.

A regra dos 3%

Antes de 2003, a fronteira entre dentro e fora do balanço era quase aritmética. O Relatório Powers resume: uma empresa podia tratar uma entidade de propósito específico (SPE) como independente se "(1) an owner independent of the company must make a substantive equity investment of at least 3% of the SPE's assets, and that 3% must remain at risk throughout the transaction; and (2) the independent owner must exercise control of the SPE." [56]. A equipe da SEC descreveu depois o efeito prático: dando a um terceiro com 3% de capital o controle de voto da entidade, "a sponsor could generally avoid consolidation, despite the fact that the activities of the SPE could not be substantively changed by the "controlling" investor." [57].

Algumas estruturas da Enron nem cumpriam a regra; outras cumpriam no papel e não na substância. As mais conhecidas, os veículos "Raptor", eram proteções contra a queda de investimentos da própria Enron, capitalizadas com ações da Enron. O Relatório Powers aponta o defeito: "If the value of the investments fell at the same time as the value of Enron stock fell, the SPEs would be unable to meet their obligations and the "hedges" would fail. This is precisely what happened in late 2000 and early 2001." [56]. No segundo semestre de 2000, os "lucros" vindos dos Raptors responderam por mais de 80% do lucro antes de impostos que a Enron divulgou [56]. Uma garantia que perde valor exatamente quando é necessária tem nome técnico, risco de correlação adversa (wrong-way risk), e o nome volta na seção sobre 2008 e na seção sobre o círculo da IA.

Receita que dá a volta

O mecanismo da Enron mais parecido com o tema deste artigo não foram os Raptors, mas os "prepays". Eram operações de três partes, com uma afiliada da Enron, um banco e um veículo criado pelo banco, em que os pagamentos davam a volta completa. A equipe da SEC descreve o resultado: "The circular nature of delivery and payments with respect to the commodities had the effect of eliminating any material risk or any potential gain with respect to the changes in the price of the underlying commodity. Each party's apparent assumption of price risk was illusory. With the elimination of price risk, "prepays" were effectively debt." [57]. A Enron lançava o dinheiro como caixa operacional e a obrigação como passivo de gestão de risco, não como dívida: "these prepay transactions allowed Enron to hide debt, lower key financial ratios followed by analysts, and provide the illusion of cash flow from operations." [57].

A mesma passagem dá a ferramenta analítica que este artigo usa para ler os acordos da IA, o chamado problema da unidade de conta: "The economics of a transaction may look quite different depending on how broadly or narrowly one defines the boundaries of the transaction. That is, a particular contract may appear to have certain economic characteristics when viewed in isolation—and may be given a certain accounting treatment that corresponds to those economic characteristics—but if understood as a piece of a larger agreed-upon transaction may actually have quite different economics, and be properly accorded different accounting treatment." [57]. Traduzindo: para entender um acordo circular, é preciso olhar o laço inteiro, e não cada perna isolada. Isso vale como método e não como acusação. Os acordos de IA discutidos adiante são públicos e não há alegação de fraude contra nenhum deles.

Quanto estava escondido

O tamanho do que ficava fora dos números é o dado mais assustador do caso. Citando o examinador nomeado pelo tribunal de falências, a equipe da SEC registrou: "For 2000, barring the use of these techniques, Enron's reported debt would have been $22.1 billion rather than $10.2 billion." E mais: em 19 de novembro de 2001, a Enron divulgou cerca de US$ 13 bilhões de dívida no balanço, mas no mesmo dia disse aos bancos que a dívida era de cerca de US$ 38 bilhões [57]. O Relatório Powers lembra que boa parte disso aparecia nas notas explicativas, mas de forma inútil: "these disclosures were obtuse, did not communicate the essence of the transactions completely or clearly, and failed to convey the substance of what was going on between Enron and the partnerships." [56]. Informação disponível e informação compreensível são coisas diferentes.

A resposta: entidades de interesse variável

A resposta do órgão que define as normas contábeis americanas, o FASB, veio em janeiro de 2003 com a Interpretação nº 46, revista em dezembro do mesmo ano [57]. A ideia central é que o controle por voto não basta para decidir quem consolida uma entidade: "The voting interest approach is not effective in identifying controlling financial interests in entities that are not controllable through voting interests or in which the equity investors do not bear the residual economic risks." [58]. Entidades com capital próprio insuficiente para se financiar sem apoio de terceiros passaram a ser chamadas de entidades de interesse variável (VIEs), e quem as consolida é o "primary beneficiary": "the party that absorbs a majority of the entity's expected losses, receives a majority of its expected residual returns, or both" [58].

As regras não acabaram com as estruturas. A equipe da SEC observou, já em 2005, que "many arrangements with potential VIEs were restructured such that the entity either would not be considered a VIE or such that no party would be required to consolidate the VIE." [57]. E uma exceção sobreviveu à reforma, a dos veículos de securitização qualificados (QSPEs): "Transferors to qualifying special-purpose entities [...] do not consolidate those entities." [59]. Foi por essa porta que passou boa parte da alavancagem de 2007.

O mesmo relatório da SEC mediu outra zona cinzenta, a dos arrendamentos. Com regras de "tudo ou nada", contratos eram desenhados para ficar logo abaixo das linhas que exigiam reconhecimento no balanço, e a equipe estimou em cerca de US$ 1,25 trilhão as obrigações futuras de arrendamentos operacionais das empresas americanas que não apareciam no balanço, só nas notas [57]. O diagnóstico ficou famoso: "lease structuring to meet various accounting, tax, and other goals, has become an industry unto itself in the last 30 years." [57].

A Enron deixa três lições para a leitura dos data centers de IA, todas sobre estrutura e nenhuma sobre fraude. O motivo para tirar investimento do balanço continua sendo proteger índices e rating. As regras que decidem o que entra no balanço dependem de julgamentos (quem tem o poder, quem absorve as perdas, quanto dura o contrato) e podem ser desenhadas em volta. E garantias cujo valor cai junto com o risco que deveriam cobrir falham justamente quando são chamadas.

2008: o que sai do balanço volta na crise

A comissão do Congresso americano que investigou a crise de 2008, a Financial Crisis Inquiry Commission (FCIC), abriu as conclusões com uma frase que deveria incomodar qualquer um que diga que crises são imprevisíveis: "We conclude this financial crisis was avoidable. The crisis was the result of human action and inaction, not of Mother Nature or computer models gone haywire." [60]. Este artigo usa 2008 como evidência no sentido estrito. Não para dizer que a IA vai produzir uma crise do mesmo tamanho, mas para mostrar quais mecanismos transformaram perdas num setor em perdas no sistema, e procurar esses mecanismos no ciclo atual.

O primeiro mecanismo era a alavancagem que não aparecia. Segundo a FCIC, os cinco grandes bancos de investimento operavam com "extraordinarily thin capital. By one measure, their leverage ratios were as high as 40 to 1, meaning for every $40 in assets, there was only $1 in capital to cover losses. Less than a 3% drop in asset values could wipe out a firm." [60]. E parte dela estava fora da vista: "the leverage was often hidden—in derivatives positions, in off-balance-sheet entities, and through "window dressing" of financial reports available to the investing public." [60]. A comissão resumiu o sistema numa frase: "We had a 21st-century financial system with 19th-century safeguards." [60].

Da Enron ao papel comercial

A FCIC liga diretamente a reforma pós-Enron à alavancagem de 2007. Bancos colocavam empréstimos em programas de papel comercial lastreado em ativos (ABCP), fora do balanço, e assim deixavam de reservar capital para eles. Quando o FASB, depois da Enron, dificultou essa saída, o mercado travou, e os reguladores bancários reagiram: em 2004 permitiram que esses ativos ficassem fora do cálculo de capital, com uma cobrança de "up to 0.8%" sobre as linhas de liquidez dadas aos programas. "Growth in this market resumed." [60]. A lição é desconfortável: a regra contábil apertou, a regra prudencial afrouxou, e a estrutura continuou crescendo.

O mercado de ABCP chegou a US$ 1.225,7 bilhões na semana encerrada em 8 de agosto de 2007, segundo os dados do Federal Reserve. No fim daquele ano estava em US$ 774,5 bilhões [61]. Economistas do próprio Fed descreveram o que aconteceu como uma corrida bancária sem bancos: "more than 100 programs (one-third of all ABCP programs) were in a run within weeks of the onset of the turmoil and the odds of subsequently leaving the run state were very low." [62]. Os veículos estruturados de investimento (SIVs), uma variante desses programas que investia sobretudo em títulos com nota AAA, eram 36, com quase US$ 400 bilhões em ativos, alavancados "an average of just under 14-to-1". Investidores de um deles, o Sigma, perderam mais de 95%. "As of fall 2010, not a single SIV remained in its original form." [60].

O que estava fora volta para dentro

O ponto central de 2008 para este artigo é o que acontece com compromissos que não estão no balanço quando a crise chega. O Citigroup tinha vendido "liquidity puts", promessas de comprar papel comercial emitido por CDOs se ninguém mais quisesse, sobre US$ 25 bilhões. Antes de 2004, essas promessas não exigiam capital nenhum; depois, 0,16%. Em 2007, a FCIC conta, os eventos iriam "catapult the entire $25 billion in commercial paper straight onto the bank's balance sheet, requiring it to come up with $25 billion in cash as well as more capital to satisfy bank regulators." [60].

Com os SIVs que o banco administrava, aconteceu o mesmo, por pressão e não por contrato. Em 13 de dezembro de 2007, o Citi anunciou uma linha de apoio aos veículos e informou: "As a result of this commitment, Citi will consolidate the SIVs' assets and liabilities onto its balance sheet under applicable accounting rules." Os ativos dos SIVs tinham caído de US$ 87 bilhões em agosto para US$ 49 bilhões [63]. Em novembro de 2008, o banco trouxe os últimos US$ 17,4 bilhões para dentro [60]. William Dudley, do Fed de Nova York, admitiu à FCIC que pouca gente entendia "the backstops that banks were providing SIV conduits either explicitly or implicitly." [60]. Na regulação bancária, isso ganhou depois o nome de risco de intervenção (step-in risk): o patrocinador não é obrigado a socorrer o veículo, mas socorre, porque a alternativa custa mais à reputação e ao negócio.

Chamadas de margem antes do prejuízo

O caso da AIG mostra outro mecanismo: perdas que chegam antes do prejuízo. A subsidiária de produtos financeiros da seguradora tinha vendido proteção contra calote, por meio de credit default swaps, sobre US$ 79 bilhões em CDOs [60]. Os contratos exigiam garantias em dinheiro se o valor de mercado dos títulos caísse ou se o rating da AIG fosse rebaixado: "collateral calls could be triggered even if there were no actual cash losses in, for example, the super-senior tranches of CDOs upon which the protection had been written." [60]. Em 12 de setembro de 2008, as chamadas somavam US$ 23,4 bilhões, e a AIG já tinha pago US$ 18,9 bilhões. O governo acabou comprometendo "more than $180 billion" para evitar que a quebra da AIG se espalhasse [60]. Quando o colateral é marcado a mercado e o contrato tem gatilhos de rating, a crise de liquidez chega antes de qualquer perda real. No ciclo da IA, o paralelo mais próximo são os empréstimos com GPUs como garantia, cujo tamanho depende do valor contábil dos chips, como mostra a seção "Os elos mais fracos".

O selo de qualidade

As agências de rating foram, para a FCIC, "essential cogs in the wheel of financial destruction". Entre 2000 e 2007, a Moody's deu nota máxima a quase 45 mil títulos ligados a hipotecas, e "83% of the mortgage securities rated triple-A that year ultimately were downgraded" (o ano é 2006) [60]. O problema não era só de método. Muitos daqueles títulos tinham nota alta porque a estrutura transferia para eles a qualidade de crédito de outra parte. A seção "Fora do balanço" mostra que o maior título de dívida já emitido para um data center tem rating alto pelo mesmo motivo: a nota é da garantia, não do prédio.

Maquiagem de fim de trimestre

O Lehman Brothers acrescentou um mecanismo que fecha a lista. Segundo o examinador do processo de falência, Anton Valukas, o banco usou operações chamadas internamente de "Repo 105" para tirar temporariamente do balanço cerca de US$ 50 bilhões em ativos no fim do primeiro e do segundo trimestres de 2008. Com elas, a alavancagem líquida divulgada no fim do segundo trimestre foi 12,1, e não 13,9 [64]. Os próprios funcionários do Lehman chamavam a prática de "accounting gimmick" [64]. O relatório não decidiu se as operações violavam tecnicamente a norma contábil; o ponto é que um índice observado de perto pelo mercado virou alvo de engenharia.

Risco de correlação adversa

O conceito que liga todos esses casos é o wrong-way risk. O Comitê de Basileia define a versão específica assim: "Specific wrong-way risk arises when the exposure to a particular counterparty is positively correlated with the probability of default of the counterparty due to the nature of the transactions with the counterparty." [65]. Na revisão pós-crise, o Comitê registrou que as seguradoras de crédito (monolines) eram contrapartes cuja qualidade de crédito andava junto com a exposição, e que essas operações "contributed to the losses during the crisis". O exemplo de manual que ele dá é o de uma contraparte que oferece as próprias ações como garantia [66], exatamente o defeito dos Raptors da Enron. Em outubro de 2026, o FMI aplicou o mesmo conceito aos acordos circulares da IA, como mostra a seção seguinte.

A reforma de 2009

O FASB respondeu em junho de 2009 com duas normas, a FAS 166 e a FAS 167. Elas acabaram com a exceção dos veículos de securitização qualificados e trocaram o teste quantitativo de consolidação por um qualitativo. Desde então, uma empresa consolida uma entidade de interesse variável se tem "a. The power to direct the activities of a VIE that most significantly impact the VIE's economic performance b. The obligation to absorb losses of the VIE that could potentially be significant to the VIE or the right to receive benefits from the VIE that could potentially be significant to the VIE." [67]. As estruturas existentes não foram poupadas, e o efeito foi medido: na semana encerrada em 31 de março de 2010, os bancos comerciais americanos tinham trazido para o balanço US$ 377,8 bilhões em ativos e passivos de veículos antes fora dele, a maior parte em securitizações de cartão de crédito [68]. É sob essa regra, o teste de "poder" e de "obrigação de absorver perdas", que as joint ventures de data center de hoje ficam fora do balanço.

O shadow banking de então e o crédito privado de agora

Em 2008, a alavancagem estava no que se chamou de sistema bancário paralelo. Hoje, a expansão mais rápida do crédito fora dos bancos está nos fundos de crédito privado. O FMI estimou o setor em cerca de US$ 2,1 trilhões em 2023 e listou as vulnerabilidades: "relatively fragile borrowers, increased exposure of pensions and insurers to the asset class, a growing share of semiliquid investment vehicles, multiple layers of leverage, stale valuations, and unclear interconnections between participants." [69]. Em maio de 2026, o Federal Reserve registrou US$ 1,4 trilhão em empréstimos de crédito privado nos EUA e um fato inédito: no primeiro trimestre de 2026, os pedidos de resgate aceitos nos fundos semilíquidos de crédito (BDCs perpétuos) superaram as entradas, "the first time this has occurred since these vehicles were created" [70]. O mesmo relatório anota que a alavancagem das maiores seguradoras de vida estava "well into the upper quartile of its historical distribution" [70].

Duas pontes ligam esse mercado à IA. A primeira é o destino do dinheiro: economistas do BIS calcularam que a participação de empresas de tecnologia nos empréstimos diretos do crédito privado "doubled to over 40% between 2020 and 2025" [71]. A segunda é quem compra o risco. O Conselho de Estabilidade Financeira (FSB) alertou em maio de 2026 que ratings privados, às vezes de agências pouco conhecidas, estão sendo usados para atrair seguradoras, que dependem de notas para cumprir as regras de capital: "The growing use of private ratings, sometimes from lesser-known providers, to facilitate investment by rating-reliant investors (including insurers) also warrants monitoring." [72]. E resumiu: "Private credit remains untested to a prolonged economic downturn and so warrants close attention." [72].

Onde a analogia falha

Há diferenças importantes, e omiti-las seria fazer com 2008 o que os otimistas fazem com a IA. Em 2008, o risco estava em bancos financiados com depósitos e dívida de curtíssimo prazo, e o ativo era a hipoteca de milhões de famílias. O crédito privado de hoje é financiado sobretudo com capital de longo prazo, e o próprio FMI escreveu, em 2024, que "Private credit loans are funded largely with long-term capital, mitigating maturity transformation risks. The use of leverage appears modest, as do liquidity and interconnectedness risks." [69]. Em abril de 2026, o FMI ainda avaliava que as hyperscalers "thus far have financed investments robustly without materially weakening their financial positions" [31]. A FCIC também teve votos divergentes: parte dos comissários rejeitou a narrativa da maioria, e o voto assinado por Keith Hennessey, Douglas Holtz-Eakin e Bill Thomas tem uma seção intitulada "The dangerous imprecision of the term 'shadow banking'" [60].

O que 2008 ensina, portanto, não é que a história se repete no mesmo tamanho. São cinco mecanismos para procurar. Compromissos fora do balanço voltam para dentro na crise, por contrato ou por reputação. Gatilhos ligados a preço de mercado e a rating antecipam as perdas. Ratings podem medir a garantia de outra empresa, e não o ativo. Garantias que caem junto com o risco falham quando são chamadas. E índices vigiados pelo mercado viram alvo de engenharia. As próximas seções procuram cada um deles no financiamento da IA.

O círculo: quem investe, quem vende e quem compra

A seção sobre o telecom mostrou o mecanismo: Lucent, Nortel e Cisco emprestavam dinheiro a operadoras para que elas comprassem seus equipamentos, e a receita dos fabricantes crescia com demanda que eles mesmos financiavam. Em 2025 e 2026, o mecanismo voltou maior e com mais formas. Fornecedores de chips e de nuvem compram participações em laboratórios de IA, que assinam compromissos de compra com esses mesmos fornecedores. Há bônus em ações para quem compra, garantias de aluguel, compromissos de comprar a capacidade que sobrar e financiamentos liberados conforme a entrega do produto.

O desenho genérico cabe num laço:

fornecedor de chips ou de nuvem
   |  1. investe no laboratório ou garante a dívida dele
   v
laboratório de IA
   |  2. assina compromisso de compra com o fornecedor
   v
carteira de contratos (RPO) do fornecedor
   |  3. receita futura e valor de mercado sobem
   v
fornecedor capta mais e investe de novo
   |
   +--> volta ao passo 1

Os acordos

A tabela reúne os principais acordos documentados. Os valores são os anunciados ou registrados pelas empresas, muitos deles "até", condicionados a entregas ou ainda não definitivos:

As fontes estão nos parágrafos a seguir, que detalham cada caso.

NVIDIA e OpenAI

Em setembro de 2025, NVIDIA e OpenAI anunciaram uma carta de intenção para pelo menos 10 GW de sistemas NVIDIA: "To support the partnership, NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed." [73]. O dinheiro do fornecedor seria liberado à medida que o cliente instalasse os produtos do fornecedor. No relatório trimestral seguinte, a NVIDIA avisou: "There is no assurance that we will enter into definitive agreements with respect to the OpenAI opportunity" [74]. O acordo de US$ 100 bilhões não saiu. Em fevereiro de 2026, a NVIDIA entrou com US$ 30 bilhões na rodada de US$ 110 bilhões da OpenAI, "at a $730B pre-money valuation", e a OpenAI garantiu "3GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems" [75]. Em março, Jensen Huang disse que a oportunidade de US$ 100 bilhões provavelmente estava "not in the cards", porque a OpenAI ia abrir o capital [76].

A NVIDIA não investe só na OpenAI. Em julho de 2026, a empresa informou "equity investments of $99 billion and equity investment commitments of $25 billion", e o valor contábil de suas participações em empresas fechadas tinha passado de US$ 3,8 bilhões para US$ 47,9 bilhões em um ano [77]. No relatório anual, ela própria descreve parte desses investimentos: "These investments include AI model makers that purchase our products directly or through CSPs." [78]. Os ganhos não realizados com essas participações entram no lucro. Quando a avaliação dos clientes sobe, sobe também o resultado de quem vende para eles.

No mesmo trimestre, a NVIDIA criou o que chamou de "a new business model": nuvens de IA compram sua infraestrutura e a NVIDIA se compromete a alugar capacidade delas, contratos que "the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates". Esses compromissos somavam US$ 36 bilhões [77]. Na prática, o fornecedor vende o equipamento e garante um piso de demanda para quem o comprou. A seção "Fora do balanço" mostra o passo seguinte: as garantias de até US$ 105 bilhões sobre os aluguéis da OpenAI.

AMD e OpenAI

A AMD escolheu outra forma. Em outubro de 2025, deu à OpenAI o direito de comprar até 160 milhões de ações da AMD a US$ 0,01 cada, liberadas em lotes conforme a OpenAI comprasse GPUs AMD, até 6 GW, e conforme a ação da AMD atingisse metas que "escalate to $600 per share for the final tranche" [79]. Pela minha conta, 160 milhões de ações eram cerca de 9,8% das ações da AMD em circulação no fim de outubro de 2025 [80]. A diretora financeira da AMD previu que a parceria traria "tens of billions of dollars in revenue for AMD" [79]. Na contabilidade, o bônus em ações vira desconto na receita: "The grant date fair value of the warrant will be recognized as a reduction to revenue when revenue is recognized." [80]. Em fevereiro de 2026, a AMD repetiu o modelo com a Meta, no mesmo tamanho e com as mesmas metas [81]. E passou a garantir o aluguel de um data center de um parceiro, com exposição máxima de US$ 4,1 bilhões; garantias desse tipo, diz a empresa, "may be issued in exchange for warrants" [82].

Microsoft e OpenAI

A Microsoft é sócia, fornecedora e beneficiária de receita da OpenAI ao mesmo tempo. Depois da reestruturação de outubro de 2025, sua participação foi avaliada em cerca de US$ 135 bilhões, "roughly 27 percent", e a OpenAI "has contracted to purchase an incremental $250B of Azure services" [83]. Em janeiro de 2026, a diretora financeira informou que "Approximately 45% of our commercial RPO balance is from OpenAI" [84]; pela minha conta, cerca de US$ 280 bilhões. No ano fiscal encerrado em junho de 2026, a Microsoft registrou US$ 24,1 bilhões de receita com a OpenAI e tinha US$ 6,0 bilhões a receber dela, contra US$ 11,9 bilhões investidos desde o início. O lucro do ano incluiu ainda US$ 6,5 bilhões de ganhos líquidos com a participação na OpenAI, que segundo a empresa "primarily relate to the dilution gain from the OpenAI Recapitalization" [3]. Ser diluída a uma avaliação maior gerou lucro contábil para a sócia.

Amazon, OpenAI e Anthropic

Com a Amazon, o dinheiro anda nos dois sentidos entre as mesmas duas empresas. Em fevereiro de 2026, a Amazon se comprometeu a investir US$ 50 bilhões na OpenAI, e a OpenAI ampliou seu contrato com a AWS, de US$ 38 bilhões, "by $100 billion over 8 years", com cerca de 2 GW de capacidade nos chips Trainium [85]. Com a Anthropic, a Amazon investiu US$ 8 bilhões em títulos conversíveis entre 2023 e 2025 e mais US$ 5 bilhões em ações no segundo trimestre de 2026, e abriu uma linha de financiamento de até US$ 20 bilhões [7]. Do outro lado, a Anthropic anunciou: "We are committing more than $100 billion over the next ten years to AWS technologies, securing up to 5GW of new capacity to train and run Claude." [86].

O balanço da Amazon mostra o efeito da avaliação. No segundo trimestre de 2026, a empresa registrou "upward adjustments of approximately $50.5 billion" no valor de suas ações da Anthropic, por causa das novas rodadas de investimento da empresa [7]. Pela minha conta, é cerca de três vezes o lucro operacional da AWS no mesmo trimestre, de US$ 16,6 bilhões. A Anthropic, por sua vez, levantou US$ 65 bilhões em maio de 2026 a uma avaliação de US$ 965 bilhões e disse que sua receita anualizada tinha passado de US$ 47 bilhões [87]. O crescimento é real; o ganho da investidora, por enquanto, é contábil.

Anthropic, Google e o aviso no relatório da Alphabet

A Anthropic repete o desenho com vários fornecedores. Em novembro de 2025, a empresa anunciou que "has committed to purchase $30 billion of Azure compute capacity", enquanto "NVIDIA and Microsoft are committing to invest up to $10 billion and up to $5 billion respectively in Anthropic" [88]. O Google é o outro grande fornecedor. Em outubro de 2025, a Anthropic anunciou que usaria até um milhão de TPUs do Google, uma expansão "worth tens of billions of dollars" [89]. O relatório anual da Alphabet traz o aviso mais claro sobre o risco desse tipo de acordo, sem nomear empresas: "Our financial instruments also include debt and equity investments in companies with which we also entered into commercial arrangements at or near the same time." E continua: "if our investment in such companies becomes impaired, we may need to re-evaluate the accounting for the commercial arrangement, including reducing any remaining performance obligations." [4]. Ou seja, se o investimento perder valor, a carteira de contratos também pode encolher. As duas coisas caem juntas.

CoreWeave e SB Energy

O mesmo desenho aparece entre empresas menores. Quando fechou seu primeiro contrato com a CoreWeave, em março de 2025, a OpenAI recebeu US$ 350 milhões em ações da própria CoreWeave [90]; os contratos entre as duas chegaram a cerca de US$ 22,4 bilhões em setembro [91]. A NVIDIA, que fornece todas as GPUs da CoreWeave, comprou US$ 2 bilhões em ações da empresa e se comprometeu a comprar a capacidade que sobrar sem cliente até 2032 [92], [93]. Em janeiro de 2026, OpenAI e SoftBank investiram US$ 1 bilhão na SB Energy, desenvolvedora de data centers do grupo SoftBank; a OpenAI assinou com ela um aluguel de 1,2 GW, e a SB Energy "will become a major customer of OpenAI" [94]. Numa única operação, a OpenAI investe no próprio locador, que vira seu cliente e pertence ao grupo de um de seus maiores acionistas.

Quanto vale a carteira

Esses acordos aparecem nas carteiras de contratos a cumprir dos fornecedores. Em meados de 2026, a carteira comercial da Microsoft era de US$ 678 bilhões [3], a da Oracle de US$ 664 bilhões [18], a do Google Cloud de US$ 514 bilhões [14], a da AWS de cerca de US$ 496 bilhões [7] e a da CoreWeave de US$ 103,7 bilhões [93]. Pela minha conta, a soma passa de US$ 2,4 trilhões. Do outro lado estão, em boa parte, os mesmos poucos laboratórios: só a OpenAI fala em compromissos de cerca de US$ 1,4 trilhão em oito anos, para uma receita anualizada de cerca de US$ 50 bilhões no fim de setembro de 2026 [95], [96].

Os economistas do BIS mediram a circularidade de forma sistemática. Entre 2021 e 2025, "55.2% of incoming investments in AI firms came from other AI firms", e 46,4% do valor dos investimentos entre empresas de IA envolvia também uma relação de fornecimento entre investidor e investida. A consequência, segundo eles: "circular relationships make reported demand partly endogenous to firms’ own financing decisions. For example, when a supplier finances a customer, part of the supplier’s revenue growth reflects its own capital investment, rather than organic final demand." E a comparação com o telecom é deles: "The parallel with the telecommunications boom of the late 1990s is instructive: upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors’ equipment" [51].

O FMI, no relatório de outubro de 2026, pôs números no círculo: "Fourteen publicly disclosed transactions since 2023 account for approximately $300 billion of aggregate equity investments led by hyperscalers in financially weaker AI firms, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments through services or chip orders". Cerca de US$ 320 bilhões dessa dívida está com as empresas investidas, "including debt backed by investor guarantees, contingent financing, and other off-balance-sheet debt". E nomeou o risco: "distress at key investee firms could present significant "wrong-way risk" to the AI ecosystem: Equity and debt investments in the firms depreciate in value precisely when future revenue streams tied to them are impaired." [97]. Jason Wu, do FMI, resumiu na entrevista coletiva: "If there are questions about the monetization in AI, the equity investments will deplete in value, the debt may be difficult to service, and the revenue commitments in the future may not materialize", o que chamou de "triple whammy" [98]. Em abril, o Fundo já tinha estimado que, de uma alta de cerca de 12 pontos percentuais nas ações das empresas do círculo entre setembro e dezembro de 2025, 7 pontos vinham do reforço mútuo entre elas [31].

Nem todo acordo é circular

Há limites para o argumento. O acordo da OpenAI com a Broadcom, de 10 GW em aceleradores próprios, não tem investimento do fornecedor no cliente; as empresas assinaram um "term sheet" e não divulgaram valor [99]. Muitos compromissos são condicionados a entregas, e alguns nunca viraram contrato, como os US$ 100 bilhões da NVIDIA. Isso reduz o risco efetivo, mas também mostra que os números de manchete superestimam a demanda firme. E a Microsoft afirma que, no trimestre de abril a junho de 2026, todo o crescimento de sua carteira veio de clientes que não são laboratórios de fronteira, como mostra a seção "O que pode desmentir a tese".

O círculo não é fraude, e cada acordo tem lógica comercial: os laboratórios precisam de capacidade garantida, e os fornecedores precisam de clientes para a capacidade que constroem. O problema é o que o conjunto faz. Ele infla a receita e a carteira dos fornecedores com demanda que eles mesmos financiaram. Transforma a avaliação privada dos laboratórios em lucro contábil dos investidores. E concentra o risco num punhado de clientes que ainda queimam caixa. Se um deles falhar, os outros perdem pelos três lados ao mesmo tempo: na participação, no crédito e na receita prometida.

Fora do balanço: Hyperion, xAI e a engenharia dos data centers

As seções sobre a Enron e sobre 2008 mostraram como veículos fora do balanço funcionam e como voltam para dentro na crise. Em 2025 e 2026, a mesma caixa de ferramentas voltou a ser usada para financiar a IA: sociedades de propósito específico, joint ventures em que a empresa fica com a minoria, garantias de valor residual e aluguéis de curto prazo renováveis. A diferença em relação à Enron é importante: tudo está divulgado, auditado e, até onde se sabe, dentro das regras. A semelhança também: a dívida fica longe do balanço da empresa que usa o ativo e que, na prática, sustenta o pagamento.

Hyperion: um data center de US$ 27 bilhões fora do balanço da Meta

Em 21 de outubro de 2025, a Meta anunciou uma joint venture com fundos geridos pela Blue Owl Capital para construir o campus Hyperion, na Louisiana. Os fundos da Blue Owl ficaram com 80%, a Meta com 20%, e o custo total de desenvolvimento foi estimado em cerca de US$ 27 bilhões. A Meta alugou todo o campus por contratos com "a four-year initial term with options to extend" e deu à joint venture "a residual value guarantee for the first 16 years of operations" [100]. Recebeu ainda uma distribuição única de US$ 2,55 bilhões, segundo o relatório anual; o comunicado falava em cerca de US$ 3 bilhões [8].

O dinheiro veio de um título. Segundo a Bloomberg, o Morgan Stanley montou "over $27 billion of debt and about $2.5 billion of equity into a special purpose vehicle", com a PIMCO como compradora principal, e os papéis receberam nota A+ da S&P, de acordo com um relatório de rating. A Bloomberg resumiu o desenho: "Under the SPV structure, Meta is not borrowing the capital itself, the financing entity is." [101]. O emissor aparece num documento de um dos fundos da Blue Owl enviado à SEC: a Beignet Investor LLC, dona dos 80% da joint venture, emitiu "$27,293,849 in senior secured notes with an interest rate of 6.581% and a maturity date of May 30, 2049" (valores em milhares de dólares) [102]. Na manhã seguinte à emissão, os títulos já eram oferecidos a 105% do valor de face [101].

O relatório anual da Meta explica por que nada disso entra no balanço dela. Os aluguéis começam em 2029, com compromisso inicial de cerca de US$ 12,31 bilhões. A garantia de valor residual tem um limite agregado de cerca de US$ 28 bilhões, que diminui com o tempo, e "As of December 31, 2025, RVG payments are not probable and therefore, no liability has been recorded." Sobre a joint venture: "We do not have the power to direct the activities that most significantly impact the Venture's economic performance. Therefore, we are not the primary beneficiary and do not consolidate the variable interest entity (VIE). Our maximum exposure to loss related to the Venture was $45.95 billion as of December 31, 2025" [8]. Em junho de 2026, a exposição máxima era de US$ 46,03 bilhões [9].

O argumento está no relatório do primeiro trimestre de 2026. Segundo a Meta, as decisões que mais pesam no resultado da joint venture são "decisions pertaining to remarketing the data center campus, including but not limited to, negotiations with future lease tenants and individual property sales", e essas decisões não são dela [103]. É uma leitura possível da regra contábil, que exige poder de dirigir as atividades mais relevantes e exposição significativa a perdas para consolidar [104]. Mas as decisões de recolocar o campus no mercado só importam se a Meta sair, e quem decide se sai, a cada quatro anos, é a Meta. A Meta também constrói e administra o campus. A auditoria percebeu a delicadeza: a Ernst & Young incluiu a consolidação da joint venture entre os principais assuntos de auditoria e escreveu que "Auditing the Company's determination of the primary beneficiary of the VIE was especially challenging" [8]. No Wall Street Journal, o colunista Jonathan Weil foi mais direto no título, "AI Meets Aggressive Accounting at Meta's Gigantic New Data Center", e no subtítulo: "Favorable treatment off the balance sheet hinges on some convenient assumptions." [105].

A regra sobre garantias ajuda a entender o desenho. Pela norma americana de arrendamentos, o passivo de aluguel inclui só os valores de garantia residual que a empresa provavelmente vai pagar. A Deloitte observa que, por isso, contratos em que boa parte da economia está numa garantia de valor residual "may continue to yield favorable accounting results (e.g., reduced leverage) under ASC 842" [106]. E aluguéis que ainda não começaram só entram no balanço do inquilino na data de início [107].

Pela minha conta, a exposição máxima declarada, de US$ 46 bilhões, equivalia a 78% de toda a dívida de longo prazo da Meta no fim de 2025, de US$ 58,7 bilhões [8]. E o modelo está sendo repetido. Em julho de 2026, a Meta assinou exclusividade para uma segunda joint venture, em El Paso, no Texas, de novo com 20%, e informou que dará "residual value guarantees with a maximum aggregate exposure of approximately $13 billion" [9].

Quem ficou com o risco

Num título privado, saber quem é o dono é difícil. Uma parte aparece nos relatórios que os fundos registrados nos Estados Unidos entregam à SEC, os formulários N-PORT. Pela minha análise desses dados, com posições quase todas de 31 de março de 2026, 317 fundos tinham US$ 9,44 bilhões dos títulos da Beignet, cerca de 35% da emissão. Fundos de registrantes com o nome PIMCO concentravam US$ 6,57 bilhões, e só o PIMCO Income Fund tinha US$ 4,11 bilhões, cerca de 15% da emissão. Fundos exclusivos de produtos de seguradoras, como anuidades variáveis, tinham cerca de US$ 473 milhões, e cinco fundos fechados tinham tomado empréstimos dando os títulos em garantia [108]. O preço médio informado pelos fundos caiu de 105,7% do valor de face em dezembro de 2025 para 102,9% em março de 2026 [109], [108]. Os outros dois terços da emissão estão com quem não aparece nesses dados, como carteiras próprias de seguradoras, fundos de pensão e fundos privados.

O capital próprio também chegou ao investidor individual. O fundo da Blue Owl que detém 63% da Beignet Investor, com compromisso de US$ 1,53 bilhão, é um fundo imobiliário sem negociação em bolsa, vendido em oferta privada a "accredited investors", pessoas físicas e instituições com patrimônio elevado [102]. O risco de um único campus de uma única empresa foi repartido entre fundos de renda fixa, ETFs, planos de previdência e investidores ricos, como as hipotecas de 2006 foram repartidas entre milhares de carteiras.

xAI: as GPUs num veículo próprio

A xAI, de Elon Musk, usou outra versão. Segundo a Bloomberg, em outubro de 2025 a empresa montou uma captação de US$ 20 bilhões que incluía uma sociedade de propósito específico "that will buy Nvidia processors and rent them to xAI", com cerca de US$ 7,5 bilhões de capital e até US$ 12,5 bilhões de dívida; a NVIDIA investiria até US$ 2 bilhões no capital, e a xAI alugaria os chips por cinco anos [110]. Num texto explicativo, a Bloomberg resumiu: "And xAI's exposure? Just a 5-year lease; nothing else." [111]. Em janeiro de 2026, a Apollo liderou US$ 3,5 bilhões em financiamento para um veículo da Valor Equity Partners comprar US$ 5,4 bilhões em infraestrutura, incluindo GPUs GB200, e alugá-la a uma subsidiária da xAI. A Apollo informou que "NVIDIA invested in VCI as an anchor Limited Partner" [112]. O fornecedor dos chips é sócio do veículo que compra os chips para alugar ao cliente.

O mesmo texto da Bloomberg trouxe a comparação com o passado na voz de um estrategista do UBS, Matthew Mish: "During the dot-com era, most of the growth was financed with equity not debt. So when it burst, the impact in the economy was manageable," e "Now, there's a rapid growth of capex in AI firms driven by debt and it's starting to be kept off balance sheet." [111].

Garantias de quem vende

A forma mais nova do desenho é a garantia dada pelo fornecedor. Em agosto de 2026, a NVIDIA informou à SEC, no item do formulário que trata de obrigações financeiras diretas e fora do balanço, garantias de valor residual de até US$ 105 bilhões sobre aluguéis da SB Energy a uma afiliada da OpenAI, para cerca de 4,25 GW de capacidade num campus em Ohio. Se a OpenAI ficar insolvente ou deixar de pagar, a NVIDIA cobre "any shortfall between the guaranteed minimum value of a lease (as set forth in the Agreements) and amounts recovered through a replacement lease or sale". As garantias acabam, entre outros casos, quando a OpenAI conseguir "a satisfactory credit rating" [113]. Segundo a CNBC, as conversas tinham começado em julho com uma garantia de até US$ 250 bilhões, que permitiria à OpenAI tomar dívida para um campus de 10 GW "on the strength of Nvidia's credit" [114]. Em troca, o campus vai abrigar só infraestrutura NVIDIA [77]. Há contrapesos: as garantias entram em vigor por fases, a primeira prevista para o ano fiscal de 2029, diminuem ao longo dos 20 anos de aluguel, e a OpenAI se comprometeu a reembolsar a NVIDIA [77]. Jensen Huang descreveu o objetivo: "We are securing long-lived infrastructure for NVIDIA compute so OpenAI can deploy the most productive AI factories that can be upgraded repeatedly with each new generation delivering more intelligence and better economics." [115]. O prédio dura 20 anos, os chips mudam a cada geração, e quem garante o aluguel é quem vende os chips.

O Google fez o mesmo com antigas mineradoras de bitcoin convertidas em data centers. Garantiu os aluguéis da neocloud Fluidstack para que elas conseguissem financiamento: cerca de US$ 3,2 bilhões no caso da TeraWulf, com warrants que levariam sua participação a cerca de 14%; US$ 1,4 bilhão no da Cipher Mining; e, no da Hut 8, uma garantia que "covers the lease payments and related pass-through obligations" de um contrato de US$ 7,0 bilhões por 15 anos, com bancos financiando até 85% do custo do projeto [116], [117], [118]. Segundo a Bloomberg, o Google registra essas garantias como derivativos de crédito [111].

Em agosto de 2026, a NVIDIA levou a ideia a outra escala. Assinou memorandos com Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs e KKR para criar plataformas de financiamento capazes de "mobilize over $500 billion of third-party capital" e apresentou seus chips como ativo de investimento: "NVIDIA compute is an investable asset" [119]. Segundo a CNBC, a NVIDIA terá a opção de garantir 25% de cada empréstimo. Larry Fink, presidente da BlackRock, fez a comparação que este artigo vinha fazendo: "This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s". E o executivo da KKR responsável por infraestrutura digital completou: "you can securitize it or effectively divide that risk and sell it to investors who want to participate anywhere in that stack" [120]. Os acordos eram memorandos de entendimento, sem contratos fechados.

O que não aparece no balanço

Somados, os aluguéis já assinados e ainda não iniciados das maiores empresas do ciclo passam de US$ 1 trilhão. Nenhum deles está no balanço, porque a regra contábil só reconhece o aluguel quando ele começa:

As fontes são os relatórios trimestrais e anuais das empresas [3], [18], [9], [7], [5], [93]. Pela minha conta, a soma, de cerca de US$ 1,15 trilhão, é mais que o dobro da dívida financeira de Microsoft, Alphabet, Amazon, Meta e Oracle juntas em meados de 2026, de cerca de US$ 480 bilhões. São valores nominais, a pagar ao longo de até 20 anos, e parte depende de condições contratuais; não são dívida. Mas são compromissos fixos, e a Moody's diz tratá-los como equivalentes a dívida [17].

A fronteira entre investimento e aluguel também se move. Em julho de 2026, a Microsoft aumentou de 15 para 25 anos a vida útil estimada de seus data centers e explicou o efeito: "more of our future datacenter leases will shift from finance leases to operating leases as a result of this update. Finance leases are included in capital expenditures while operating leases are not." Com isso, a previsão de capex do ano caiu de cerca de US$ 190 bilhões para cerca de US$ 175 bilhões, com os planos de investimento "unchanged" [13]. O investimento é o mesmo; o número que o mercado acompanha ficou menor.

Os reguladores deram nome ao fenômeno. Em março de 2026, o BIS escreveu: "These arrangements amount to "shadow borrowing": obligations that are economically akin to debt but largely reside outside corporate balance sheets. By channelling sizeable private credit into AI infrastructure, these structures strengthen links between hyperscalers and non-bank investors such as private credit vehicles and insurers." [121]. No relatório anual de junho, acrescentou que a construção de data centers é cada vez mais terceirizada a quem aluga de volta às hyperscalers "on long-dated contracts with embedded exit clauses. The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times." [122]. O FMI, em outubro de 2026, apontou como fonte futura de necessidade de financiamento "the eventual realization of sizable off-balance-sheet commitments" [123]. E lembrou que as securitizações ligadas a data centers, US$ 46 bilhões desde 2018, devem chegar a US$ 150 bilhões em 2028 [31].

Nada disso é ilegal, e a contabilidade de hoje é muito mais exigente que a de 2001 ou de 2007. A pergunta que 2008 deixou é outra: o que acontece com esses compromissos quando o ativo perde valor. A garantia de valor residual da Meta só é paga se a Meta não renovar os aluguéis e o campus valer menos que o limite, ou seja, se a demanda por data centers cair. A garantia da NVIDIA só é chamada se a OpenAI não pagar. Os dois gatilhos dependem do mesmo evento, uma correção da IA, que ao mesmo tempo derruba o valor das ações e do fluxo de caixa de quem garante. A dívida não sumiu. Mudou de endereço, e continua ligada ao mesmo risco.

A física da depreciação: GPU não é imóvel

Toda dívida tem um lastro, e o lastro decide quanto o credor recupera se algo der errado. Num data center de IA, o lastro tem três camadas com vidas muito diferentes. O prédio e o terreno duram décadas: a Alphabet deprecia seus data centers e escritórios "over periods of seven to 40 years" [4]. Os equipamentos de energia e refrigeração duram cerca de uma década: a CoreWeave deprecia "Data center equipment" em 8 a 12 anos, e a Amazon aumentou de 10 para 13 anos a vida útil do "heavy equipment" que sustenta seus centros de distribuição e data centers [124], [125], [6]. E os servidores, onde está a maior parte do dinheiro, valem por poucos anos. A pergunta desta seção é quantos.

O ritmo do fornecedor

Em junho de 2024, Jensen Huang, presidente da NVIDIA, anunciou o ritmo dos lançamentos: "Our company has a one-year rhythm. Our basic philosophy is very simple: build the entire data center scale, disaggregate and sell to you parts on a one-year rhythm, and push everything to technology limits" [126]. Até então, o salto entre gerações levava dois anos: o comunicado de lançamento da Blackwell, em março de 2024, descreve a arquitetura anterior, Hopper, como "launched two years ago" [127]. No ritmo anunciado, quem compra um servidor hoje e o deprecia em seis anos ainda vai tê-lo no balanço depois de cinco ou seis lançamentos anuais do fornecedor.

O próprio fornecedor diz quanto cada geração antiga fica para trás. No lançamento da Blackwell, a NVIDIA afirmou que a nova plataforma permitia rodar modelos de um trilhão de parâmetros "at up to 25x less cost and energy consumption than its predecessor" [127]. Na página do sistema GB200 NVL72, a comparação com a H100 é de "30x" na inferência de modelos de linguagem e de "25x" em eficiência energética [128]. Em março de 2026, a geração seguinte, Vera Rubin, prometia "one-tenth the token cost relative to NVIDIA Blackwell" [129]. E em agosto de 2026 a empresa citou testes da consultoria SemiAnalysis segundo os quais o Vera Rubin NVL72 entrega "30x higher throughput per megawatt and 45x lower token costs than NVIDIA GB300 NVL72" [130]. São números de quem vende, medidos em condições escolhidas por quem vende, e devem ser lidos assim. Mas o sinal é o que importa: o vendedor está dizendo ao mercado que o equipamento comprado no ano anterior produz tokens bem mais caros que o novo.

A página da H100 no site da NVIDIA chega a pôr preço na diferença. Segundo ela, com dados da SemiAnalysis de abril de 2026, a H100 produz inferência a cerca de US$ 0,09 por milhão de tokens, e a B200, a cerca de US$ 0,02: "roughly 4.5x cheaper than H100" [131]. A comparação usa programas diferentes nos dois chips e velocidades por usuário diferentes, então não é exata. O mesmo site registra o outro lado da moeda: o custo por token da própria Blackwell caiu cinco vezes depois do lançamento, só com software, "driven by continuous software optimization" [132]. O dono de uma GPU ganha com a otimização de software, mas perde quando a geração seguinte chega e o preço do token cai para todos.

O mecanismo econômico é simples. Quando a oferta acompanha a demanda, o preço de aluguel de uma hora de GPU tende a ser puxado pelo custo de produzir tokens com o equipamento mais eficiente disponível. Uma GPU antiga continua funcionando, mas só consegue cobrar o que sobra entre o preço de mercado do token e o próprio custo, que inclui a energia que ela gasta a mais. Por isso a vida útil econômica de uma GPU não é a mesma coisa que a vida física.

O desgaste físico

A vida física também não é longa quando a máquina trabalha no limite. A Meta publicou um dos raros relatos detalhados de falhas em treino de grande escala. Durante 54 dias de pré-treino do Llama 3 405B, num cluster de até 16 mil GPUs H100 "each running at 700W TDP with 80GB HBM3", houve 466 interrupções, 419 delas inesperadas: "Approximately 78% of the unexpected interruptions are attributed to confirmed hardware issues, such as GPU or host component failures, or suspected hardware-related issues like silent data corruption and unplanned individual host maintenance events." [133]. Na tabela de causas, 148 interrupções vieram de GPUs com defeito e 72 de falhas na memória HBM3 [133]. A automação deu conta: o tempo efetivo de treino ficou "higher than 90%". Mas uma frota que perde componentes nesse ritmo envelhece mais rápido que um servidor de banco de dados.

O prédio também envelhece

A ideia de que o prédio dura décadas e só o chip é trocado esbarra na potência. Uma H100 consome até 700 W; um servidor DGX B200, com oito GPUs Blackwell, "~14.3 kW max" [131], [132]. Os racks GB200 NVL72 já são refrigerados a líquido [128]. Engenheiros da NVIDIA escreveram em 2025 que a distribuição de energia usada hoje dentro dos racks "begins to hit physical limits" acima de 200 kW por rack, e anunciaram uma arquitetura de 800 volts em corrente contínua "to support 1 MW IT racks starting in 2027". No padrão atual, um único rack de 1 MW "requires up to 200 kg of copper busbar" [134]. A Agência Internacional de Energia (IEA) calcula que a densidade de potência dos servidores de IA "increased by 11 times" entre 2020 e 2025 e deve quadruplicar de novo até 2027 [22]. Um prédio projetado para uma geração de chips pode precisar de obras pesadas, em energia e refrigeração, para receber a seguinte.

A contabilidade foi na direção oposta

Enquanto o ritmo dos chips acelerava, a contabilidade das grandes empresas fez o contrário: alongou a vida útil dos servidores. Uma vida útil mais longa espalha o custo por mais anos e aumenta o lucro de cada ano. As mudanças estão nos relatórios anuais:

As justificativas variam. A Microsoft citou "investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology" [136]. A Meta, em 2022, "expected longer refresh cycles in our data centers" [137]. A Amazon é a exceção, e a mais reveladora. Em 2025, ela encurtou a vida de parte dos servidores e deu o motivo: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." [6]. No fim de 2024, a empresa já tinha decidido aposentar antes da hora parte dos servidores e lançado "approximately $920 million of accelerated depreciation and related charges" [125]. A Microsoft, no relatório anual de 2026, passou a descrever a vida útil de servidores e equipamentos de rede como "two to six years" [3].

Os valores envolvidos crescem com o investimento. Na Meta, a depreciação de servidores e equipamentos de rede foi de US$ 13,36 bilhões em 2025, contra US$ 7,32 bilhões em 2023, e o valor bruto desses ativos passou de US$ 68,4 bilhões para US$ 98,0 bilhões em um ano [8]. O FMI fez a conta para o setor inteiro. Pela depreciação divulgada por dez hyperscalers, a vida útil implícita de todo o imobilizado, prédios incluídos, é de cerca de sete anos, mas GPUs e chips avançados "could face obsolescence within two years". O Fundo também testou uma vida útil mais curta: "Under a scenario of three-year useful life of capital and assuming hyperscalers fulfill these additional financing needs by issuing debt, their debt levels could rise from the present $800 billion to more than $1 trillion." Pelo modelo do FMI, os spreads de CDS dessas empresas subiriam cerca de 60 pontos-base, "a significant increase considering that these spreads are currently in the range of 20 to 160 basis points" [31]. O FMI trata a obsolescência como risco de negócio, sem efeito de primeira ordem na estabilidade financeira. Para quem empresta contra esses ativos, porém, é a pergunta central.

Burry, a NVIDIA e o preço do aluguel

Em novembro de 2025, o investidor Michael Burry levou a discussão para o centro do mercado. Escreveu que "Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era" [138] e que "Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment. Yet this is exactly what all the hyperscalers have done" [139]. Segundo a Business Insider, ele estimou que Meta, Google, Oracle, Microsoft e Amazon subestimariam a depreciação em cerca de US$ 176 bilhões entre 2026 e 2028 e que, em 2028, o lucro da Oracle estaria inflado em cerca de 26% e o da Meta em cerca de 20%, "without elaborating on his calculations". Jim Chanos, investidor que apostou contra a Enron, foi na mesma linha: "the chips basically depreciate over two years" [139].

A palavra "fraude" é de Burry e precisa ser lida com cuidado. As mudanças de vida útil estão nos relatórios anuais, com o efeito em dólares, como mostra a tabela acima; a discussão é sobre a estimativa, não sobre ocultação. A NVIDIA respondeu num memorando a analistas. Segundo a CNBC, a empresa argumentou que os clientes depreciam GPUs em quatro a seis anos "based on real-world longevity and utilization patterns" e que as A100, lançadas em 2020, "continue to run at high utilization rates and retain meaningful economic value well beyond the two to three years claimed by critics". Burry rebateu: "I am not claiming Nvidia is Enron. It is clearly Cisco." [140].

Até agora, o mercado de aluguel deu razão em parte aos dois lados. Até 2025, os preços caíram. Em junho daquele ano, a AWS cortou em até 45% o preço das instâncias com GPUs NVIDIA [141], e o preço por hora de uma máquina com oito H100 na tabela pública da empresa passou de US$ 98,32 em agosto de 2023 para US$ 55,04 em outubro de 2026 [142], [143]. Em 2026, porém, o aluguel de capacidade reservada subiu. Nos blocos de capacidade da AWS, a H100 foi de US$ 3,93 por GPU-hora em junho de 2025 para US$ 5,97 em 13 de outubro de 2026, alta de 52% [144], [145]. Na Lambda, uma neocloud, o preço avulso da H100 passou de US$ 2,99 por hora em 2025 para US$ 3,99 a partir de abril de 2026, e o de clusters por algumas semanas, de US$ 2,40 em novembro de 2025 para US$ 5,54 a US$ 6,16 em outubro de 2026 [146], [147]. Segundo a CNBC, um lote de H100 de 2022 que ficou livre com o fim de um contrato na CoreWeave foi realugado "at 95% of their original price" [148].

A explicação mais citada é a escassez. A The Verge registrou que "The price to rent old chips has been rising, and Silicon Data projects that it will continue rising through 2028", e um analista ouvido pelo site atribuiu a alta a "a major shortage of inference chips" [149]. A falta de memória HBM apontada pela IEA vai na mesma direção [22]. Em agosto de 2026, a bolsa de derivativos CME anunciou contratos futuros sobre índices de aluguel de H100 e B200, citando "sharp swings in compute prices" [150]. Num texto de outubro de 2026 em defesa da durabilidade das GPUs, a NVIDIA citou estimativas de terceiros, baseadas em preços de revenda, de cinco a seis anos de vida útil para um sistema H100 e de nove a dez anos para o GB300 NVL72 [151].

É o melhor argumento contra a obsolescência rápida, e ele precisa ser levado a sério. Mas preço sustentado por escassez não é o mesmo que valor de longo prazo. O mesmo texto da NVIDIA afirma: "Power is the binding constraint on an AI factory. This makes tokens per second per megawatt the number that governs earning capacity." E acrescenta que o Vera Rubin entrega "over 30x higher throughput per megawatt" que o GB300 [151]. Com a energia limitada, quem troca um rack antigo por um novo multiplica a produção do mesmo megawatt. Quando a escassez de chips passar, a pressão para aposentar as máquinas antigas volta, e com ela a pergunta de Burry.

Por que isso importa para quem empresta

A tabela abaixo é só aritmética, sobre um investimento de US$ 100 em servidores depreciado em linha reta. Ela mostra quanto do custo ainda está no balanço depois de três anos, dependendo da vida útil escolhida:

Com seis anos de vida útil, metade do custo de uma GPU ainda está no balanço quando ela completa três anos, e pelo menos duas arquiteturas mais novas já estão à venda. Se o valor de mercado dessa GPU de três anos estiver bem abaixo de metade do preço original, a diferença é perda que ainda não apareceu no resultado. Para quem tem a GPU como garantia de um empréstimo, é uma garantia menor do que parece. A CoreWeave, que vive de alugar GPUs, descreve esse risco no próprio prospecto: "Any changes to the significant assumptions underlying our estimates or to the estimates of our components’ useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and future prospects." [124].

Há bons argumentos do outro lado. GPUs que deixam de ser úteis para treinar modelos de fronteira podem passar anos rodando inferência ou tarefas menos exigentes. A otimização de software aumenta a produção do mesmo chip, como a própria NVIDIA mostra. E a Alphabet lembra que a escolha da vida útil usa "historical asset performance, expected technology advancements, and our future infrastructure deployment plans" [4]: não é um número arbitrário. A conclusão não é que as empresas erraram a conta. É que o colateral que sustenta boa parte da dívida da IA perde valor econômico mais rápido que os prédios, mais rápido que as linhas de transmissão e, em várias estruturas, mais rápido que a própria dívida. É o quinto marcador da tabela de bolhas, e ele está confirmado por quem vende os chips e por quem os contabiliza.

Energia: o gargalo que pode virar ativo encalhado

Um data center de IA é, antes de tudo, uma ligação elétrica. Os chips mudam a cada ano, mas a subestação, a linha de transmissão e a usina que os alimentam são projetadas para durar 20, 30 ou 40 anos. É o mesmo descasamento de prazos da seção anterior, só que agora fora do balanço das empresas de tecnologia. Quem constrói a infraestrutura de energia de um data center assume o risco de a demanda desaparecer antes de a obra se pagar. Quando isso acontece, o que sobra é chamado de ativo encalhado: um investimento que continua de pé, mas não gera a receita que justificou sua construção.

Quanto a IA consome

A IEA estima que os data centers consumiram 415 TWh em 2024, "around 1.5% of the world’s electricity consumption". Os Estados Unidos responderam por 45% desse total, a China por 25% e a Europa por 15%. No cenário-base da agência, o consumo "is set to more than double to around 945 TWh by 2030", pouco mais que o consumo total de eletricidade do Japão hoje. Nos Estados Unidos, os data centers respondem por quase metade do crescimento da demanda até 2030, e no fim da década o país "is set to consume more electricity for data centres than for the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined." [152]. A atualização de abril de 2026 confirmou a trajetória: o consumo dos data centers cresceu 17% em 2025, e o dos data centers dedicados à IA, 50% [22].

Para os Estados Unidos, o Lawrence Berkeley National Laboratory (LBNL) calculou que os data centers consumiram 176 TWh em 2023, 4,4% da eletricidade do país. Para 2028, a projeção vai de 325 a 580 TWh, de 6,7% a 12,0% do consumo previsto [153]. O cenário alto é quase o dobro do baixo. Quem constrói uma usina ou uma linha de transmissão para esse mercado precisa escolher um número dentro dessa faixa e conviver com ele por décadas. O mesmo relatório supõe que os servidores de IA durem tanto quanto os convencionais, cerca de 5,7 anos em 2023: "For AI accelerated and nonaccelerated servers, the same lifetime assumptions are used." [153]. Até as projeções de energia carregam a hipótese de vida útil questionada na seção anterior.

A fila da rede

O limite não está só na geração, mas na ligação. A IEA calcula que, sem medidas para destravar a rede, "around 20% of planned data centre projects could be at risk of delays". Nas economias avançadas, uma linha de transmissão nova leva de quatro a oito anos, e "wait times for critical grid components such as transformers and cables have doubled in the past three years" [152]. O FMI fala em prazos de interconexão "now stretching from two to seven years" [31].

A fila americana de interconexão mostra o ritmo da rede. Segundo o LBNL, "As of the end of 2025, over 2,060 gigawatts (GW) of total generation and storage capacity were actively seeking connection to the grid." Para os projetos concluídos em 2025, a mediana entre o pedido e a entrada em operação ficou acima de cinco anos. E "Only 13% of capacity that submitted interconnection requests from 2000-2020 had reached commercial operations by the end of 2025; 75% of that capacity had been withdrawn" [154]. Essa fila é de usinas e baterias, não de data centers, mas mostra o que acontece com pedidos de conexão: a maioria nunca vira obra. A própria NVIDIA reconheceu o problema em março de 2026: "Energy is now the biggest bottleneck for AI infrastructure buildouts, with over $300 billion in equipment backlogs and more than 200 gigawatts of projects waiting in U.S. interconnection queues." [155].

Turbinas e reatores

Sem espaço na rede, muitos projetos buscam geração própria, quase sempre a gás. A GE Vernova, uma das maiores fabricantes de turbinas, informou que sua carteira de equipamentos a gás, somada a reservas de produção, cresceu de 62 para 83 GW no quarto trimestre de 2025 e de 100 para 116 GW no segundo trimestre de 2026, com a expectativa de chegar a pelo menos 125 GW no fim do ano. A produção da empresa chega ao ritmo de 20 GW por ano no terceiro trimestre de 2026, deve ir a 24 GW em 2028 e tem meta de 30 GW em 2030 [156], [157]. Pela minha conta, os 116 GW contratados ou reservados equivalem a quase seis anos da produção atual. A IEA registra "A 70% surge in gas turbine orders in 2025" e calcula que a geração a gás dentro do terreno do data center precisa ser superdimensionada de 30% a 70% em relação à demanda para garantir o fornecimento [22].

A outra aposta é nuclear, com contratos de prazo muito mais longo que qualquer chip:

A reativação de Crane ganhou ainda um empréstimo de US$ 1 bilhão do Departamento de Energia dos Estados Unidos [162], o que põe parte do risco do ciclo no balanço do governo. Para quem vende a energia, o contrato de 20 anos é a garantia que sustenta o investimento. Para quem compra, é um custo fixo de duas décadas, assumido para alimentar equipamentos depreciados em cinco ou seis anos.

Quem paga se a demanda sumir

Algumas distribuidoras americanas já cobram garantias pensando nesse risco. Em Ohio, a AEP fechou em 2024 um acordo com a equipe do regulador estadual (PUCO), a defensoria dos consumidores e outras partes. O acordo "requires large new data center customers to pay for a minimum of 85% of the energy they say they need each month – even if they use less – to cover the cost of infrastructure needed to bring electricity to those facilities." [163]. O regulador aprovou a tarifa em julho de 2025, com faixas de consumo mínimo limitadas a 85% da capacidade contratada. O contrato inicial dura o período de rampa, de até quatro anos, mais oito anos. O cliente que não tem rating alto precisa dar garantia de 50% de todos os pagamentos mínimos do contrato. Só depois do quinto ano após a rampa ele pode sair, pagando 36 meses de encargos mínimos. E a carta de compromisso "requires customers to reimburse Ohio Power 100% of the buildout costs if the customer cancels — or delays the project by more than 12 months — prior to the target energization date." [164]. No Texas, a lei estadual SB 6, sobre a conexão de grandes cargas, busca, segundo a AEP, "minimizing the potential for stranded infrastructure costs" [165].

É a lógica de qualquer credor prudente: a distribuidora não consegue verificar se a demanda prometida vai existir, então exige garantia de quem prometeu. O risco não desaparece. Passa para o dono do data center, que pode ser uma sociedade de propósito específico como as da seção "Fora do balanço". Se a distribuidora não conseguir cobrar, o custo volta para a conta de luz dos outros consumidores. A IEA descreve esse risco: data centers "often oversize their grid connections initially", e o descompasso entre data centers rápidos e investimentos lentos em energia "could raise electricity prices in some places if unmitigated." [22].

A agência também liga o ritmo da energia ao do crédito: "Data centre investments have grown too large to be funded from company balance sheets alone, and large amounts of funding from capital markets will be critical for their buildout. As a result, the pace of data centre growth, and the resulting increase in energy consumption, will be sensitive to market sentiment, including expectations for returns on investment in data centres and AI deployment, as well as to broader macroeconomic and financing conditions." [22]. Ou seja, a demanda de energia que justifica turbinas e linhas de transmissão depende do humor do mercado de crédito, que pode mudar em semanas.

A física do lado da rede

Há ainda um problema de engenharia que não aparece nas planilhas. Um cluster de treinamento não consome energia de forma estável. A Meta relatou que, no treino do Llama 3, dezenas de milhares de GPUs podiam aumentar ou reduzir o consumo ao mesmo tempo, por exemplo ao esperar um checkpoint, e isso "can result in instant fluctuations of power consumption across the data center on the order of tens of megawatts, stretching the limits of the power grid." [133]. Ao mesmo tempo, a IEA lembra que um data center de IA "is 10 times more capital-intensive than an aluminium smelter, which means curtailing its operations to provide flexibility to the grid is very costly." [152]. O data center precisa rodar quase sem parar para pagar o capital, e a rede precisa absorver oscilações que outras indústrias não produzem.

A eficiência corta nos dois sentidos. Segundo a IEA, "Software and hardware advances have resulted in the energy use per AI task dropping by at least an order of magnitude annually in recent years." [22]. Cada tarefa fica mais barata, o que pode multiplicar o uso. Mas a mesma eficiência reduz a energia necessária para um dado volume de trabalho e, como visto na seção anterior, derruba o valor das máquinas mais antigas.

A energia é a parte do ciclo da IA que mais se parece com as ferrovias e com a fibra: ativos de vida longa, construídos antes da demanda, que continuam existindo mesmo que os investidores percam dinheiro. Em 2002, a fibra apagada virou banda barata para a década seguinte. Subestações e turbinas ociosas também encontrariam outro uso. A dúvida é quem paga a conta enquanto isso: o dono do data center, se as tarifas com consumo mínimo funcionarem; os credores desse dono, se ele quebrar; ou os demais consumidores de eletricidade, se as regras falharem.

Os elos mais fracos

Crises de crédito raramente começam pelas instituições mais fortes. Em 2007, os primeiros a cair foram os veículos e fundos que dependiam de financiamento de curto prazo e de garantias que perdiam valor [60]. No ciclo da IA, as grandes hyperscalers têm balanços sólidos: segundo o Banco da Inglaterra, "the AI hyperscalers all have a credit rating of AA- or higher" [166]. O risco se concentra em três tipos de empresa: o laboratório de IA que assina compromissos muito maiores que a própria receita, a empresa tradicional que se endividou para virar fornecedora de capacidade e as neoclouds, que alugam GPUs compradas com dívida garantida pelas próprias GPUs. Os exemplos mais documentados são OpenAI, Oracle e CoreWeave.

OpenAI: compromissos de trilhão, receita de dezenas de bilhões

O crescimento da OpenAI é real. A diretora financeira, Sarah Friar, escreveu em janeiro de 2026 que a receita anualizada passou de "$2B ARR in 2023, $6B in 2024, and $20B+ in 2025", acompanhando a capacidade de computação, que foi de 0,2 GW para cerca de 1,9 GW no mesmo período [167]. Receita anualizada é o ritmo do fim do ano, não a receita do ano: segundo a CNBC, a receita de 2025 foi de US$ 13,1 bilhões. Em março de 2026, a empresa disse estar "generating $2B in revenue per month" [168], [169]. O uso também é real: em fevereiro de 2026, o ChatGPT tinha "more than 900M weekly active users" e mais de 50 milhões de assinantes [75]. Pela minha conta, cerca de 5,5% dos usuários semanais pagavam.

O problema é a escala dos compromissos. Em novembro de 2025, Sam Altman escreveu: "We are looking at commitments of about $1.4 trillion over the next 8 years" [95]. Pela minha conta, isso dá uma média de US$ 175 bilhões por ano. No fim de setembro de 2026, segundo a CNBC, a OpenAI informou aos investidores uma receita anualizada de cerca de US$ 50 bilhões, menos de um terço disso; os US$ 68 bilhões que circularam no mês anterior incluíam receita bruta de parceiros [96]. A OpenAI não publica demonstrações financeiras auditadas. O que se sabe sobre o caixa vem de documentos enviados a acionistas e relatados pelo site The Information: a empresa projetou queimar US$ 115 bilhões até 2029 [170], depois previu "$111 Billion More Cash Burn Through 2030" [171] e, em abril de 2026, havia a expectativa de que queimaria "more than $200 billion before it starts generating cash" [172]. No primeiro trimestre de 2026, a empresa queimou US$ 3,7 bilhões, "more than half its $5.7 billion in revenue" [173].

O dinheiro veio de rodadas cada vez maiores. A OpenAI captou US$ 40 bilhões a uma avaliação de US$ 300 bilhões em março de 2025 e, um ano depois, fechou US$ 122 bilhões a US$ 852 bilhões. Pela primeira vez, abriu a rodada a pessoas físicas por meio de bancos, "raising over $3 billion from individual investors", e anunciou sua inclusão em fundos negociados em bolsa da ARK Invest [169]. Em junho de 2026, apresentou à SEC um pedido confidencial de abertura de capital [174]. Parte do dinheiro dos sócios também veio de dívida. A SoftBank, que investiu ao todo US$ 64,6 bilhões por cerca de 13% da OpenAI, pagou o aporte de US$ 30 bilhões de 2026 primeiro com um empréstimo-ponte de bancos e, na última parcela, com títulos de cerca de US$ 11,1 bilhões, com rating BB+, abaixo do grau de investimento, e juros de até 9,75% ao ano [175], [176], [177]. Estrategistas ouvidos pela CNBC em maio já liam a fila de megaofertas, com SpaceX, OpenAI e Anthropic, como sinal de topo de mercado. Um deles comparou: "Back in 1999, we saw the same kind of thing where people were just rushing to get these IPOs out." [178].

O balanço da OpenAI é leve. Segundo o The Information, que leu as demonstrações da empresa, em 31 de março de 2026 ela tinha "zero debt" [179], e a linha de crédito bancária de cerca de US$ 4,7 bilhões estava intacta [169]. A dívida existe, mas está no balanço de outros: nos aluguéis de data centers assinados por Oracle e desenvolvedores, nos contratos das neoclouds e nas garantias da NVIDIA, que a seção "Fora do balanço" detalha. A própria NVIDIA explicou por que precisa garantir os clientes: "AI model makers have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity to secure the AI infrastructure necessary to grow." [77]. É a definição de vendor financing: o fornecedor dá o crédito que o mercado não dá ao cliente.

Em novembro de 2025, a diretora financeira chegou a falar num "backstop" federal para financiar chips e voltou atrás no mesmo dia: "I used the word 'backstop' and it muddied the point" [180]. Altman disse que a OpenAI não tem nem quer garantias do governo para seus data centers e escreveu: "But we of course could be wrong, and the market—not the government—will deal with it if we are." [181]. O analista Gil Luria, da D.A. Davidson, tirou a conclusão: "If sanity prevails and this federal guarantee doesn't materialize, it could become clear that OpenAI cannot live up to its extravagant commitments. In which case we would view the companies propped up by OpenAI promises as the most vulnerable," e citou Broadcom, AMD, Oracle, CoreWeave e Oklo [182]. Em abril de 2026, o Wall Street Journal noticiou, segundo a CNBC, que a OpenAI não tinha atingido as próprias metas de receita e de usuários e que a diretora financeira se preocupava com a capacidade de pagar contratos futuros de computação. A empresa chamou a reportagem de "ridiculous", mas as ações da Oracle e dos fabricantes de chips caíram no mesmo dia [183]. Em 8 de outubro de 2026, no dia em que a CNBC confirmou a receita anualizada de US$ 50 bilhões, as ações da NVIDIA caíram 3%, as da Oracle quase 6% e as da CoreWeave quase 8% [96]. O canal de transmissão já está à vista.

Oracle: uma empresa de software que virou credora da IA

Nenhuma grande empresa mudou tanto de perfil. No fim do ano fiscal encerrado em maio de 2025, a carteira de contratos a cumprir da Oracle (o RPO, na sigla em inglês) era de US$ 137,8 bilhões. Três meses depois, era de US$ 455 bilhões, "up 359%". A então presidente-executiva, Safra Catz, disse: "It was an astonishing quarter" [184]. A tabela mostra o que veio junto, em bilhões de dólares:

Os números vêm dos relatórios e comunicados da empresa à SEC; os do trimestre de março a maio de 2025 saem da diferença entre o ano fiscal e os nove meses anteriores [185], [12], [10], [186], [18]. No ano fiscal de 2026, o fluxo de caixa livre foi negativo em US$ 23,7 bilhões, nas palavras da própria Oracle [187]. E só cerca de 13% dos US$ 664 bilhões de RPO devem virar receita nos 12 meses seguintes [18].

A Oracle não diz quanto do RPO vem de cada cliente, mas avisa no relatório anual: "The economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations." [10]. Analistas fizeram a conta: segundo a D.A. Davidson, pelo menos US$ 300 bilhões dos US$ 523 bilhões de RPO de novembro de 2025 vinham da OpenAI [188]; segundo o Bank of America, mais da metade do RPO em junho de 2026 [189]. Perguntado se a OpenAI conseguiria pagar US$ 60 bilhões por ano em nuvem, o copresidente-executivo Clay Magouyrk respondeu "of course" [190].

Para pagar a obra, a Oracle foi ao mercado. Em setembro de 2025, emitiu US$ 18 bilhões em títulos em seis séries, a mais longa vencendo em 2065, com cupom de 6,10% [191]. Em fevereiro de 2026, mais US$ 25 bilhões em oito séries, a mais longa até 2066 com cupom de 6,85%, além de US$ 5 bilhões em ações preferenciais conversíveis [192]. De junho a agosto de 2026, vendeu US$ 20 bilhões em ações no mercado [186]. Os ratings ficaram em Baa2 e BBB, ambos com perspectiva negativa, "just two notches above junk status", nas palavras da CNBC sobre a avaliação da Moody's [17]. Pelos próprios term sheets da empresa, o prêmio sobre os títulos do Tesouro subiu entre as duas emissões: de 1,05 para 1,45 ponto percentual nos papéis de dez anos e de 1,37 para 1,95 ponto nos de 40 anos [191], [192].

Fora do balanço, o compromisso é ainda maior. Os contratos de aluguel já assinados e ainda não iniciados, "substantially all related to data center arrangements", com prazos de 15 a 19 anos, foram de US$ 43,4 bilhões em maio de 2025 para US$ 288 bilhões em agosto de 2026 [12], [18]. É mais que o dobro de toda a dívida financeira da empresa. A Oracle também garantiu até US$ 3,3 bilhões de um empréstimo tomado por um de seus locadores [10].

Parte do investimento passou a ser paga pelos clientes. No trimestre de junho a agosto de 2026, a Oracle recebeu US$ 11,4 bilhões em pagamentos antecipados "that included a significant financing component", ou seja, com juros embutidos [18]. A empresa criou então uma nova métrica, fora das regras contábeis, a "net cash outlay for capital expenditures", que desconta do capex esses pagamentos e parte dos fluxos com fabricantes de equipamento classificados como financiamento. Por essa medida, o investimento do ano fiscal de 2026 foi de US$ 47,7 bilhões, contra US$ 55,7 bilhões de capex pela contabilidade; no trimestre seguinte, US$ 18,0 bilhões contra US$ 28,5 bilhões [186]. A Oracle não identifica quem adiantou o dinheiro. Entre os clientes de nuvem que ela própria cita estão AMD, Meta, NVIDIA, OpenAI, TikTok e xAI [193], e a OpenAI depende de novas rodadas de investimento para pagar as próprias contas.

O mercado reagiu. Em junho de 2026, a ação acumulava queda de cerca de 55% desde o pico de setembro de 2025 e teve "the steepest weekly drop since a 20% plunge in August 2001, during the depths of the dot-com bust" [194]. Em julho, Torsten Slok, da gestora Apollo, disse à CNBC: "CDS for Oracle is at the same level it was at in 2008." [195]. Compradores dos títulos de setembro de 2025 processaram a empresa, alegando que ela não informou que precisaria de muito mais dívida, alegação ainda não julgada [196]. Em setembro de 2026, a Oracle enviou um aviso de força maior ao desenvolvedor do Project Jupiter, no Novo México, para se proteger de custos maiores; segundo o Financial Times, citado pela CNBC, os US$ 18 bilhões em dívida ligados a esse data center já eram negociados "at stressed levels" [197].

CoreWeave e as neoclouds: a GPU como garantia

A CoreWeave aluga capacidade de GPUs NVIDIA a laboratórios e hyperscalers. Em 30 de junho de 2026, tinha US$ 35,1 bilhões em dívida para US$ 5,0 bilhões de patrimônio, cerca de sete vezes, além de US$ 16,3 bilhões em passivos de aluguel e de US$ 35,5 bilhões em aluguéis assinados e ainda não iniciados, fora do balanço [93]. No segundo trimestre de 2026, a receita foi de US$ 2,6 bilhões, 112% a mais que um ano antes, mas os juros líquidos somaram US$ 640 milhões, cerca de um quarto da receita, e o prejuízo foi de US$ 626 milhões. No semestre, a empresa investiu US$ 14,1 bilhões e gerou US$ 3,7 bilhões de caixa nas operações [93].

O detalhe que liga a CoreWeave à seção sobre depreciação está nos contratos de empréstimo. O limite do crédito é calculado "based upon the depreciable cost of computing equipment", e numa das linhas, "based upon the depreciable cost of graphics processing unit servers" [93]. Ou seja, quanto a empresa pode tomar depende do valor contábil das GPUs, que depende da vida útil escolhida. As taxas efetivas vão de 7% a 15% ao ano, e parte da dívida está em subsidiárias de propósito específico, isoladas em caso de falência, que davam em garantia US$ 18,2 bilhões em ativos de longo prazo [93].

Em 2026, esse crédito virou classe de ativo. Em março, uma linha de US$ 8,5 bilhões recebeu rating A3 da Moody's, que a CoreWeave descreveu como "the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract" [198]. Em agosto, outra linha trouxe um descasamento explícito: "the DDTL 5.5 Facility carries an approximate five-year maturity while its underlying customer contracts average approximately three years in length. By financing these shorter-date commitments, lenders are signalling confidence in the long-term value of NVIDIA GPUs running on CoreWeave’s cloud platform and a willingness to underwrite renewal risk." [199]. O credor empresta por cinco anos contra contratos de três e aposta que a GPU ainda terá cliente quando eles acabarem.

A receita é concentrada. A Microsoft respondeu por cerca de 67% da receita de 2025; no segundo trimestre de 2026, os três maiores clientes somaram 36%, 26% e 10% [200], [93]. A empresa avisa que, se um cliente grande falhar, pode ficar com capacidade ociosa e com as despesas de equipamentos, aluguéis e financiamento, e acrescenta: "We do not currently maintain credit insurance to insure against customer credit risk." [93]. Do outro lado está a NVIDIA, que fornece todas as GPUs da CoreWeave, comprou US$ 2 bilhões em ações da empresa em janeiro de 2026 e se comprometeu, em setembro de 2025, a comprar a capacidade que sobrar sem cliente até 2032, num contrato inicial de US$ 6,3 bilhões [92], [93]. A seção "O círculo" já tratou desse contrato.

A Nebius, outra neocloud listada nos Estados Unidos, mostra o mesmo desenho. No primeiro semestre de 2026, investiu US$ 8,1 bilhões e faturou US$ 0,98 bilhão; o caixa das operações só foi positivo graças a US$ 4,4 bilhões em pagamentos antecipados de clientes. Em janeiro, aumentou de quatro para cinco anos a vida útil de servidores e equipamentos de rede [201]. O contrato de US$ 17,4 bilhões com a Microsoft só começa a valer quando a Nebius confirmar "that it has secured any additional financing required in respect of the capital expenditures under the Agreement" [202]. O contrato depende do financiamento, e o financiamento depende do contrato.

O custo do dinheiro subiu. No fim de setembro de 2026, o título de dez anos do Tesouro americano rendia perto de 5,17%, segundo a CNBC, e credores relataram mais seletividade: "Instead of a roster of 50 neoclouds, there's probably 20 that the market's truly interested in." [203]. A CoreWeave calcula que cada ponto percentual de alta nos juros acrescenta cerca de US$ 30 milhões por trimestre à sua despesa com a dívida de taxa flutuante [93]. Analistas do Mizuho já falavam em julho em "neoclouds generating negative free cash flow, widening credit spreads and capital raise concerns" [195].

Os três elos estão presos uns aos outros. Os compromissos da OpenAI são parte relevante da carteira da Oracle e da CoreWeave; a dívida da Oracle e da CoreWeave se apoia nesses contratos; e a NVIDIA garante pedaços da cadeia. Se a receita da OpenAI decepcionar, a pressão chega primeiro a Oracle e CoreWeave, depois aos credores delas e, por fim, às garantias da NVIDIA. É o wrong-way risk descrito na seção sobre 2008: a proteção perde valor justamente quando é chamada.

Macroeconomia e os alertas oficiais

O estouro de 2000 derrubou a bolsa e trouxe uma recessão de oito meses. O de 2008 quebrou bancos e trouxe a recessão mais longa desde a Segunda Guerra, de 18 meses [40]. A diferença esteve no canal: ações num caso, crédito no outro. Esta seção mede o peso da IA na economia americana e mostra o que bancos centrais e organismos internacionais disseram sobre o canal pelo qual uma correção passaria.

Quanto do crescimento vem da IA

Em outubro de 2025, a revista Fortune divulgou uma conta do economista Jason Furman, de Harvard: "Technically, as Furman notes, investment in information-processing equipment and software was only 4% of U.S. GDP for the first half of 2025, yet it also accounted for fully 92% of GDP growth over that period." O próprio Furman ponderou que, sem o boom da IA, os juros e a eletricidade provavelmente estariam mais baratos, o que daria algum crescimento a outros setores [204]. Refiz a conta com os dados da época e ela bate: os 92% saem de subtrair os valores em dólares encadeados do investimento em tecnologia dos do PIB [205]. O problema é o método. O BEA, que calcula o PIB americano, avisa que, para produtos com preços em queda rápida, como computadores, "the chained-dollar levels (as distinct from chain-weighted indexes and percent changes) will overstate their relative importance to GDP growth." [206].

Pela tabela de contribuições do próprio BEA, na versão revisada de setembro de 2026, equipamentos de processamento de informação e software responderam por cerca de 56% do crescimento do primeiro semestre de 2025, 30% do crescimento do ano inteiro (0,70 de 2,3 pontos percentuais) e cerca de 41% do primeiro semestre de 2026, pelas minhas contas [23]. São números brutos. Boa parte do equipamento é importada, e o PIB soma o investimento mas desconta as importações. As importações americanas de computadores, periféricos e peças passaram de US$ 217 bilhões em 2024 para US$ 362 bilhões em 2025 e chegaram a um ritmo anual de US$ 601 bilhões no segundo trimestre de 2026 [23]. As estimativas oficiais variam muito conforme o método:

As fontes da tabela são, na ordem, [204], [207], [166], [208], [23] e [31]. Em termos brutos, o investimento em tecnologia respondeu por algo entre metade e dois terços do crescimento americano no primeiro semestre de 2025. Descontadas as importações, o efeito foi bem menor. A distância entre as estimativas já diz algo: o peso da IA na economia americana depende do método, e quem afirma um número único está escolhendo um.

O efeito não para no investimento. A construção de data centers nos Estados Unidos chegou a um ritmo anual de US$ 84,95 bilhões em agosto de 2026 (dado preliminar), 73% acima de um ano antes, enquanto a construção total caía 1,7% [209]. E a alta das ações também mexe com o consumo: a equipe do Banco da Inglaterra estimou que a valorização das ações americanas respondeu por "around 10% of the increase in US real consumption since the start of 2024, with AI-related equities the main driver during this period." [166].

Economistas do BIS, o banco dos bancos centrais, colocaram o ciclo em perspectiva histórica. Pelos cálculos deles, os gastos com fábricas de equipamentos de TI e data centers equivaliam a cerca de 1% do PIB americano em meados de 2025, "similar in size to the US shale boom of the mid-2010s and half as large as the rise in IT investment during the dot-com boom of the 1990s." E acrescentaram: "The end of previous investment booms was associated with a slowdown in GDP growth of more than 1 pp on average (Graph 4.B). Notably, the largest contraction was after the US dot-com boom, even though the boom was small relative to GDP." [208]. O tamanho do ciclo, sozinho, não diz quanto ele pode machucar.

Concentração

O segundo canal é a bolsa. Em outubro de 2025, o comitê de política financeira do Banco da Inglaterra registrou que as cinco maiores empresas do S&P 500 somavam cerca de 30% do índice, fatia "higher than at any point in the past 50 years", e que, por isso, "any AI-led price adjustment would have a high level of pass-through into the returns for investors exposed to the aggregate index." [210]. O FMI calculou no mesmo mês que as sete maiores de tecnologia eram 33% do índice e que uma medida de concentração baseada no índice Herfindahl-Hirschman estava "substantially higher than during the dot-com bubble" [207]. Em julho de 2026, o Banco da Inglaterra estimou que as empresas ligadas à IA já eram cerca de metade do S&P 500, contra um quarto em 2022, e que a fatia podia crescer com as ofertas de ações de OpenAI e Anthropic [1].

O BIS aplicou um teste estatístico usado para detectar bolhas, que mede se um preço entrou em trajetória explosiva. Em dezembro de 2025, tanto o S&P 500 quanto o ouro tinham "entered explosive territory", e foi "the only time in at least the last 50 years in which gold and equities have entered this territory simultaneously". O próprio BIS lembra o limite do teste: "it provides no information on when bubbles may burst." [211].

Da ação para a dívida

O que mais preocupa os bancos centrais não é o preço das ações, e sim a mudança no modo de pagar. O Banco Central Europeu escreveu em maio de 2026 que empresas e infraestrutura de IA "have started to rely increasingly on credit financing" e lembrou um padrão histórico: "Strong growth in either equity prices or business debt in isolation has not been associated with financial crises in the past. That said, 15% of those years that showed particularly strong growth in both equity prices and business debt have been followed by a financial crisis within two years". O gráfico que acompanha o texto tem um título direto: "Greater debt financing for AI is concerning for financial stability" [212]. O mesmo relatório ressalva que o crescimento da dívida das empresas, no agregado, ainda é baixo nos Estados Unidos e na zona do euro.

Os números do Banco da Inglaterra, de julho de 2026, mostram a velocidade. As cinco maiores hyperscalers tinham só 3% da dívida corporativa americana com grau de investimento no fim de 2025, mas responderam por mais de 15% das emissões de 2026 até o início de maio. Emissores ligados à IA foram 41% das emissões americanas de alto rendimento (exceto refinanciamentos) no ano, e, segundo a OCDE, a fatia do crédito privado no financiamento da IA subiu de 9% em 2024 para 34% em 2025 [1]. Em setembro, o comitê do banco citou uma estimativa do Morgan Stanley: cerca de US$ 450 bilhões em dívida ligada à IA emitidos em 2026 até o início do mês, mais que o dobro de todo o ano de 2025 [213]. Na pesquisa do Fed com participantes do mercado, a IA passou de 30% das menções entre os riscos em 2025 para 50% em 2026, e os entrevistados citaram que os investimentos "are increasingly funded by debt, creating leverage in the system" [214], [70].

A tabela resume a evolução dos alertas oficiais:

O relatório anual do BIS, de junho de 2026, faz a ligação histórica que este artigo propõe. As manias dos canais nos anos 1830, das ferrovias nos anos 1840, da eletrificação nos anos 1920 e da internet nos anos 1990 "all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify. These episodes ended with an eventual reversal in investment, inducing economy-wide recessions." E sobre o ciclo atual: "The scale and pace of the current AI investment boom accompanied by expectations of large productivity payoffs bear resemblance to these precedents, highlighting potential downside risks in the near term." [122].

Julho de 2026: um ensaio

O mercado já teve um ensaio. Segundo o Banco da Inglaterra, "Equity valuations for AI companies fell sharply in July. The scale of the adjustment was amplified by an unwinding of stretched positions and associated deleveraging activity." Houve perdas grandes para investidores alavancados, mas "there was no spillover to core markets." [213]. O BIS registrou a quebra de "a highly leveraged AI-focused hedge fund", uma queda de quase 15% na relação preço/lucro das hyperscalers e de 40% na dos fabricantes de semicondutores, e a alta dos spreads de CDS dessas empresas, enquanto os do resto do S&P 500 ficavam estáveis [217]. Em 27 de julho, a Apple passou a NVIDIA como a empresa mais valiosa do mundo; segundo a CNBC, os investidores "rewarded its reluctance to spend heavily on capital expenditures for AI, preferring to rent capacity instead of building its own." [218].

O episódio mostrou as duas coisas ao mesmo tempo. O mercado já cobra prêmio de quem investe e se endivida mais. E, enquanto a maior parte do ciclo for paga com caixa e ações, uma queda forte da bolsa ainda não vira crise de crédito.

O que os reguladores não dizem

Nenhum desses documentos prevê uma crise. O Banco da Inglaterra escreveu em dezembro de 2025 que "The predominantly cash and equity-financed nature of AI development to date limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term." [166]. O FMI, em abril de 2026, considerou que "financial stability risks remain contained" [31]. E o BCE notou que, ao contrário de 2000, as hyperscalers "combine high profit margins, strong earnings growth, little debt and diversified underlying businesses beyond AI" [215].

O que todos apontam é o mecanismo: a passagem do financiamento por caixa e ações para o financiamento por dívida, dentro e fora do balanço. É ela que pode transformar uma correção de bolsa, como a de 2000, num problema de crédito, como o de 2008. O Banco da Inglaterra simulou um cenário hipotético de correção ligada à IA, com queda de 45% nas ações americanas em seis trimestres e alta de 350 pontos-base nos spreads de crédito, que tiraria 2,2 pontos percentuais do PIB britânico [1]. E lembrou, em setembro, que as projeções de crescimento e as contas públicas também dependem da IA: "A reassessment of those expectations could therefore affect not only AI-related asset valuations but also sovereign debt markets." [213].

E o Brasil?

O Brasil não está no centro do círculo descrito neste artigo. Nenhuma empresa brasileira vende GPUs, treina modelos de fronteira ou emite dezenas de bilhões de dólares em títulos para erguer data centers. Mas o país está ligado ao ciclo por três canais: a política pública, que aposta em atrair data centers; a energia, que é o principal argumento do país e a parte mais durável do investimento; e os investidores, expostos às ações americanas e às condições financeiras globais.

Redata: o incentivo que chegou atrasado

A principal aposta federal é o Regime Especial de Tributação para Serviços de Datacenter, o Redata. A Medida Provisória 1.318, de 17 de setembro de 2025, suspendeu PIS/Cofins, IPI e Imposto de Importação na compra de equipamentos de tecnologia por data centers habilitados. Em troca, a empresa precisava oferecer ao mercado interno pelo menos 10% da capacidade instalada com o benefício, atender toda a sua demanda de energia com fontes limpas ou renováveis, ter eficiência hídrica (WUE) de no máximo 0,05 litro por kWh e investir 2% do valor dos produtos em pesquisa e desenvolvimento [219]. A exposição de motivos afirma que "cerca de 60% das cargas digitais nacionais" são processadas no exterior e que operar no Brasil custa, em média, 30% mais, "devido principalmente à tributação sobre equipamentos de Tecnologia da Informação e Comunicação (TIC)". A Receita estimou a renúncia fiscal em R$ 5,20 bilhões em 2026, R$ 1,00 bilhão em 2027 e R$ 1,05 bilhão em 2028 [220]. O número mais repetido, porém, não está em nenhum documento técnico. Na cerimônia de assinatura, o vice-presidente Geraldo Alckmin disse: "Há uma expectativa de que o Brasil possa atrair R$ 2 trilhões de investimentos, ao longo de 10 anos" [221]. Nenhuma metodologia foi publicada para esse valor.

A MP perdeu a validade em 25 de fevereiro de 2026 sem ser votada [222]. Segundo a Agência Brasil, o presidente do Senado, Davi Alcolumbre, decidiu não votar o texto no prazo [223]. O conteúdo voltou como projeto de lei, aprovado pela Câmara em 24 de fevereiro e pelo Senado só em 1º de setembro [224], [225], e foi sancionado como Lei 15.504, de 15 de setembro de 2026. A lei trocou "fontes limpas ou renováveis" por "fontes renováveis ou de baixa emissão" e manteve um detalhe decisivo: os benefícios de PIS/Cofins e IPI "produzirão efeitos até 31 de dezembro de 2026", por causa da transição para a CBS e o IBS da reforma tributária [226]. No mesmo dia, uma lei complementar abriu uma exceção às regras fiscais para que o benefício pudesse ser concedido [227]. Na sanção, a Agência Brasil registrou que parte das regras "ainda dependerá de regulamentação para definir procedimentos de habilitação das empresas, critérios técnicos e outras condições necessárias para a utilização efetiva dos benefícios tributários" [228]. Até 13 de outubro de 2026, uma busca no Diário Oficial não encontrou o decreto de regulamentação [229]. Se ele demorar, a janela dos principais benefícios tributários em 2026 será curta.

Anúncios e atos oficiais

Os valores anunciados para data centers no Brasil são enormes, mas mudam de natureza quando se lê o documento oficial:

O caso da ByteDance é o mais instrutivo. Na aprovação pelo conselho das Zonas de Processamento de Exportação, o projeto "prevê investimentos totais de R$ 108 bilhões até 2035 em equipamentos de alta tecnologia" e, "Para o decênio posterior a 2035 (36-46), a empresa espera investir outros R$ 135 bi em atualizações tecnológicas" [232]. Quase todo o número da manchete é equipamento e sua troca ao longo de duas décadas, o mesmo ativo de vida curta discutido na seção sobre depreciação. O primeiro marco concreto é bem menor: "A expectativa é de que, até o terceiro trimestre de 2027, pelo menos 40MW estejam em plena operação." [234].

No Rio Grande do Sul, o governo estadual informou que o Ministério de Minas e Energia publicou "a resolução que garante a conexão de 5 gigawatts (GW) de energia para cidade de data centers em Eldorado do Sul" [236]. A portaria publicada não cita potência nenhuma. Ela reconhece que a alternativa de acesso à rede "atende aos critérios de mínimo custo global", condiciona a ligação à "disponibilidade sistêmica verificada pelo ONS" e é revogada se não houver contrato de uso do sistema de transmissão até 31 de dezembro de 2033 [237]. A Elea falou em 1,5 GW de energia renovável "a ser entregue imediatamente" na Rio AI City [238], mas as fases concretas descritas pela própria empresa somam cerca de 200 MW [239]. Em abril de 2026, o comunicado da venda da Elea para a I Squared Capital mencionou "mais de 1 GW de capacidade planejada e mais de 300 MW em terrenos com energia garantida" [240]. Para comparar escalas: o SP6 da Equinix, um data center real e em operação, custou US$ 114 milhões, e a empresa diz investir no país "em média, cerca de R$ 1 bilhão por ano ao longo da última década" [241].

Nada disso indica fraude. Anunciar potencial é comum em infraestrutura. O problema, num ciclo de euforia, é que o número da manchete circula como se fosse investimento feito e acaba servindo de base para política pública. O planejamento elétrico mostra o tamanho da diferença. Numa apresentação de setembro de 2025, a Empresa de Pesquisa Energética (EPE) mostrou que os pedidos de data centers acompanhados pelo ministério somavam 10.721 MW até 2029 e 20.986 MW até 2038, dos quais 13.399 MW no Sudeste. A previsão da própria EPE para a carga nova de data centers na rede básica em 2029 fica em torno de 1.000 MW médios, pela leitura do gráfico [242]. As unidades são diferentes, potência pedida e energia média, então a comparação vale como ordem de grandeza. Uma busca no Diário Oficial encontrou pelo menos 15 portarias de 2024 e 2025 que reconhecem o acesso de data centers à rede básica, e nenhuma informa potência [243]. Para filtrar pedidos especulativos, a ANEEL passou a exigir garantias financeiras de quem pede acesso à rede de transmissão [244], a mesma lógica da tarifa de Ohio descrita na seção sobre energia.

Energia sobrando, rede faltando

O argumento brasileiro é a energia limpa. As renováveis foram 86,8% da matriz elétrica em 2025 [245]. E há energia sobrando. Pelos meus cálculos com os dados abertos do ONS, as usinas eólicas e solares acompanhadas pelo operador deixaram de gerar 37,2 TWh em 2025 por ordem de corte, 20,7% do que poderiam ter produzido. O pico foi em outubro de 2025, com 30,4%, e 84% do volume cortado no ano estava no Nordeste. De janeiro a 12 de outubro de 2026, os dados preliminares somam 31,3 TWh, ou 22,1% [246], [247]. Com a Lei 15.269/2025, os custos de cortes por "sobreoferta de energia elétrica, referente a eventos motivados pela impossibilidade de alocação de geração de energia elétrica na carga" não podem mais ser cobertos pelo encargo previsto na lei do setor [248]. Na prática, o gerador não é compensado por esse tipo de corte.

Os data centers são apresentados como compradores desse excedente, mas a conta não fecha sozinha. A energia cortada em 2025 é cerca de quatro vezes a carga nova de data centers que a EPE projeta para 2029 (1.000 MW médios equivalem a cerca de 8,8 TWh por ano). E a sobra está concentrada no Nordeste, enquanto a maior parte dos pedidos de data center está no Sudeste. O gargalo brasileiro, como o americano, é rede e localização, não geração.

A dívida que liga o Ceará à IA

O elo mais direto entre o ciclo da IA e a dívida de longo prazo no Brasil está na geração de energia. Em junho de 2026, a Casa dos Ventos concluiu nos Estados Unidos uma captação privada de US$ 1,1 bilhão, em duas emissões: US$ 825 milhões com prazo de 24 anos e US$ 252 milhões com prazo de 17 anos. As usinas financiadas "fazem parte de contratos de longo prazo de venda de energia, os PPAs, firmados com empresas como Ascenty, Omnia e ByteDance, dona do TikTok." A dívida é em dólar porque "os contratos de venda de energia também são em moeda americana", o que, segundo a empresa, "elimina o risco cambial da operação" [249]. O risco de câmbio sai, mas fica o de contraparte: dívida de 17 a 24 anos, sustentada por contratos com compradores cujo negócio depende de equipamentos depreciados em cinco ou seis anos. É o mesmo descasamento dos contratos nucleares americanos. Do outro lado estão, segundo a empresa, "23 investidores institucionais de segmentos variados, como seguradoras, gestores de ativos e fundos de pensão" [249].

Investidores e Banco Central

O investidor brasileiro também está exposto. Segundo a B3, o estoque de BDRs, os recibos de ações estrangeiras negociados na bolsa brasileira, caiu de R$ 34 bilhões em 2021 para R$ 18 bilhões em 2022 e chegou a R$ 50 bilhões em 2025, com cerca de 1 milhão de investidores. A NVIDIA (NVDC34) foi o terceiro BDR mais negociado no período coberto pelo relatório, com R$ 116 milhões por dia [250]. A queda de 2022 coincidiu com a das ações de tecnologia nos Estados Unidos [46], mas câmbio e resgates também podem ter pesado, e não encontrei estudo que separe cada fator. Os fundos destinados ao público em geral podem aplicar até 20% do patrimônio em ativos no exterior, e os de investidores qualificados, até 40% [251]. Os fundos de pensão fechados podem ter até 10% no exterior e mais até 10% em BDRs e ETFs internacionais negociados na B3 [252]. São tetos, não exposição efetiva, e não encontrei dado consolidado sobre a alocação real.

O Banco Central acompanha o tema. No Relatório de Estabilidade Financeira de maio de 2026, escreveu que "avaliações elevadas e maior concentração de investimentos em determinados segmentos aumentam a sensibilidade dos mercados a frustrações de expectativa, com risco de correções de preços mais intensas". Sobre o crédito privado americano que financia data centers, apontou "riscos relevantes associados à concentração e à opacidade". E lembrou que as economias emergentes "permanecem expostas a pressões cambiais e a saídas de capitais". Mas concluiu que, "Diante da reduzida exposição cambial e da pequena dependência de funding externo, a exposição do SFN a flutuações financeiras originadas no exterior é baixa." [253].

Numa correção, o Brasil sentiria o ciclo da IA de forma indireta: pela saída de capital e pelo câmbio, pelos projetos anunciados que encolheriam ou não sairiam do papel, pelos geradores cuja dívida depende de contratos com data centers e pelos investidores de varejo em BDRs, que já viram o estoque cair quase pela metade em 2022. O custo fiscal estimado do Redata, de R$ 7,25 bilhões em três anos, é pequeno perto dos valores anunciados. O país não carrega a dívida do ciclo. Carrega parte da narrativa e uma parte da energia.

O que pode desmentir a tese

Um argumento que não pode ser refutado não é análise. Esta seção reúne os melhores argumentos contra a ideia de que o ciclo da IA tem a forma de uma bolha de crédito, com os dados que os sustentam e o que eles não provam.

A demanda existe e está crescendo

O argumento mais forte é a receita. A Microsoft informou que o Azure passou de US$ 100 bilhões no ano fiscal de 2026, com alta de 41%, e que "Customer demand continues to exceed available capacity" [13]. A receita do Google Cloud cresceu 82% no segundo trimestre de 2026, para US$ 24,8 bilhões, e a carteira de contratos chegou a US$ 514 bilhões [254], [14]. A AWS cresceu 37% no mesmo trimestre, para um ritmo anual de US$ 169 bilhões, e informou que seu negócio de IA passou de US$ 25 bilhões por ano [25]. Na Meta, a receita subiu 28% e Mark Zuckerberg atribuiu o resultado à IA: "AI is accelerating our core business today" [16]. Do lado do uso, o aplicativo Gemini chegou a 950 milhões de usuários mensais [254], e a OpenAI fala em "our collective 1.2B weekly users" [255].

O que isso não prova é que a receita cubra o investimento, nem que a demanda seja independente do financiamento. Em janeiro de 2026, a Microsoft informou que "Approximately 45% of our commercial RPO balance is from OpenAI" [84]. Na Amazon, o prazo médio dos contratos de longo prazo ainda não reconhecidos subiu de 4,0 para 6,4 anos em um ano, com as ampliações dos contratos com OpenAI e Anthropic [7]. A própria Microsoft oferece o contraponto: em julho de 2026, disse que todo o crescimento da carteira no trimestre anterior veio de clientes fora dos laboratórios de fronteira e que "nearly 90%" da receita de nuvem do ano veio desses outros clientes [13]. É o dado mais importante a acompanhar. Se a demanda de quem não é laboratório de IA continuar crescendo nesse ritmo, o círculo descrito neste artigo pesa menos do que parece.

Os balanços são fortes

As quatro maiores tinham, juntas, cerca de US$ 533 bilhões em caixa e aplicações de curto prazo em junho de 2026: US$ 76,8 bilhões a Microsoft, US$ 242,5 bilhões a Alphabet, US$ 123,0 bilhões a Amazon e US$ 90,3 bilhões a Meta, pelas minhas contas sobre os balanços enviados à SEC [3], [5], [7], [9]. O BCE e o BIS lembraram que, ao contrário de 2000, a alta das ações vem acompanhada de lucro: segundo o BIS, isso é "dissimilar to the dotcom bubble of the late 1990s, which was largely fuelled by over-optimistic expectations that were not underpinned by realised earnings growth." [211], [215].

O que isso não prova é que o caixa venha da operação. A Alphabet levantou US$ 49,6 bilhões líquidos em ações em junho de 2026, "including capital expenditures to scale AI infrastructure", emitiu US$ 56,2 bilhões em dívida no semestre e registrou um ganho de US$ 98,0 bilhões, "primarily the result of net unrealized gains on our equity securities" [254]. A diretora financeira resumiu a dívida da empresa: "a year ago, we were at about $16 billion; and we went to about $100 billion across multiple currencies and geographies." [14]. Na Meta, o fluxo de caixa livre do segundo trimestre de 2026 foi de US$ 784 milhões, e o caixa de US$ 90,3 bilhões já era quase igual à dívida de longo prazo, de US$ 83,7 bilhões [16]. Os balanços continuam fortes, mas a folga encolheu rápido.

A dívida ainda é pequena perto do sistema

O Banco da Inglaterra escreveu que o financiamento predominantemente por caixa e ações "limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term" [166]. O BCE notou que o crescimento da dívida das empresas, no agregado, ainda é baixo [212], e em novembro de 2025 calculou que o setor de tecnologia era de 1,8% a 8,3% dos principais índices de títulos corporativos [215]. A seção sobre 2008 mostrou outro ponto a favor: o crédito privado é financiado sobretudo com capital de longo prazo, sem a dependência de dinheiro de curtíssimo prazo que derrubou os veículos de 2007.

O que isso não prova é que o risco esteja medido. As estatísticas de dívida não incluem os contratos de aluguel ainda não iniciados, que passam de US$ 1 trilhão nas maiores empresas, nem as garantias e os compromissos de compra descritos nas seções "O círculo" e "Fora do balanço". E a velocidade importa tanto quanto o estoque: as emissões ligadas à IA em 2026 já tinham mais que dobrado as de 2025 em setembro [213].

A produtividade vai aparecer

Os estudos de tarefa mostram ganhos reais. Num estudo com 5.172 atendentes de suporte, publicado no Quarterly Journal of Economics, o acesso a um assistente de IA "increases worker productivity, as measured by issues resolved per hour, by 15% on average" [256]. Pesquisadores do Fed de St. Louis estimaram que os usuários de IA generativa economizaram em média 5,4% das horas de trabalho, o que corresponderia a um aumento de cerca de 1,1% na produtividade agregada [257]. A adoção pelas empresas também cresce: na pesquisa do Census americano encerrada em 4 de outubro de 2026, 23,8% das empresas disseram ter usado IA nas duas semanas anteriores, contra 17,3% em novembro de 2025, já com a mesma pergunta [258]. O BIS resume a distância entre o micro e o macro: os estudos de tarefa relatam ganhos "between 20 and 50% in time savings", mas as estimativas de produtividade agregada "tend to be more conservative at less than 1% over a long horizon" [122].

O que isso não prova é que quem financia a infraestrutura vá ganhar. É a lição das ferrovias e da fibra: a tecnologia pode transformar a economia e, mesmo assim, o primeiro investidor perder dinheiro. Satya Nadella, presidente-executivo da Microsoft, parece concordar com o paralelo. Perguntado em julho de 2026 sobre o risco de excesso de capacidade, citou justamente o ano do pânico ferroviário americano: "All of us are reading this 1873 as the book to be read." A diretora financeira, Amy Hood, deu a resposta prática: o capex migrou para ativos de vida curta, e "if the demand environment changes, you just slow down what is, in fact, the largest component and the driver of COGS." [13]. É um bom argumento para a Microsoft. Mas a flexibilidade de quem compra GPUs é o risco de quem as financiou e de quem construiu o prédio para elas.

O preço do token pode criar a própria demanda

O custo de usar IA cai muito depressa. A NVIDIA diz que o custo por token da Blackwell caiu cinco vezes depois do lançamento, só com software [132], e a IEA calcula que a energia por tarefa de IA vem caindo "by at least an order of magnitude annually" [22]. Se a demanda for elástica o bastante, cada queda de preço multiplica o uso, e a capacidade instalada é absorvida. É a aposta de quem constrói. O problema é o prazo: se a demanda demorar mais que a dívida, quem financiou quebra antes de a capacidade encher, como aconteceu no telecom.

O que eu consideraria prova contrária

Com base nesses argumentos, estes são os sinais que me fariam abandonar a tese:

• receita de produtos de IA, e não de nuvem em geral, cobrindo uma parte crescente do capex por vários trimestres seguidos;
• a OpenAI chegando a fluxo de caixa positivo antes do previsto nos documentos relatados à imprensa;
• preço de aluguel de GPUs de gerações anteriores estável por mais de três anos, compatível com vidas úteis de seis anos;
• spreads de crédito de Oracle e das neoclouds caindo enquanto elas refinanciam as dívidas que vencem em 2027 e 2028;
• crescimento da carteira das hyperscalers sustentado por clientes que não são laboratórios de IA.

Se esses sinais aparecerem juntos, o ciclo da IA terá sido caro, mas não uma bolha de crédito. Até outubro de 2026, o quadro era misto. Na Microsoft, a receita anualizada de IA crescia 123% ao ano [20], mais que o capex, que subiu cerca de 80% no ano fiscal de 2026, e no último trimestre a carteira cresceu só com clientes de fora dos laboratórios. O aluguel de H100, chips de 2022, voltou a subir em 2026, outro ponto a favor da durabilidade, como mostra a seção sobre depreciação. Já os spreads de Oracle e das neoclouds subiram, e a OpenAI continuava queimando caixa.

Cenários e sinais de alerta

Bolhas não marcam data para estourar. Greenspan falou em exuberância irracional em 1996, e a bolsa subiu por mais três anos; o teste estatístico do BIS, como o próprio BIS avisa, não diz quando a bolha acaba. Por isso esta seção não faz previsão de data. Descreve como seria uma correção, a partir de dois ensaios que já aconteceram, e lista os indicadores que mostram em que direção o ciclo está indo.

Dois ensaios

O primeiro veio da tecnologia. No fim de dezembro de 2024, a chinesa DeepSeek publicou o relatório técnico de seu modelo V3 e calculou o custo do treino final em US$ 5,576 milhões, supondo um aluguel de US$ 2 por hora de GPU H800. O próprio relatório avisava que o número excluía "the costs associated with prior research and ablation experiments on architectures, algorithms, or data" [259]. Em 27 de janeiro de 2025, o mercado reagiu: "Nvidia lost close to $600 billion in market cap on Monday, the biggest drop for any company on a single day in U.S. history." A Broadcom caiu 17% e a Oracle, 14%. Analistas da Cantor falaram em "great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs." [260]. A recuperação foi rápida: em outubro de 2025, a NVIDIA se tornou a primeira empresa a passar de US$ 5 trilhões em valor de mercado [261]. O episódio mostrou que uma mudança de eficiência pode reprecificar da noite para o dia a capacidade instalada e quem a vende.

O segundo veio das finanças. Em julho de 2026, como mostra a seção sobre macroeconomia, as ações ligadas à IA caíram forte, um fundo alavancado quebrou e as ações das hyperscalers passaram a ser negociadas a múltiplos menores, mas não houve contágio para o resto do mercado. Os dois ensaios foram eventos de bolsa. Nenhum chegou ao crédito. A pergunta é o que faria chegar.

Três cenários

No primeiro cenário, a demanda alcança a capacidade. A receita de produtos de IA cresce a ponto de cobrir uma parte relevante do investimento, o aluguel de GPUs se mantém firme mesmo depois do fim da escassez, e as empresas refinanciam sem dificuldade as dívidas que vencem em 2027 e 2028. As perdas ficam com algumas neoclouds e com investidores que entraram tarde. É o cenário que as empresas descrevem, e os dados de nuvem da seção anterior dão algum apoio a ele.

No segundo, a correção é de bolsa, como em 2000. A monetização da IA decepciona, as ações caem, e as hyperscalers fazem o que a diretora financeira da Microsoft descreveu: desaceleram as compras de ativos de vida curta [13]. Os fornecedores de chips e de memória perdem pedidos, as neoclouds mais endividadas são reestruturadas, e o crescimento desacelera. O Banco da Inglaterra simulou um caso assim, com queda de 45% nas ações americanas e alta de 350 pontos-base nos spreads [1]. O FMI calculou que o desmonte do investimento em IA poderia reduzir o PIB, o investimento e o consumo globais "by as much as 0.4, 1.3, and 0.6 percent, respectively, by 2028" [97]. E o BIS lembra que o fim dos booms de investimento anteriores trouxe, em média, uma desaceleração de mais de um ponto percentual no PIB [208].

No terceiro, a correção chega ao crédito. Um laboratório grande não consegue levantar a rodada seguinte e renegocia seus compromissos. Oracle e neoclouds enfrentam o refinanciamento num momento ruim: a CoreWeave, por exemplo, tinha US$ 4,4 bilhões em dívida vencendo no resto de 2026 e US$ 6,2 bilhões em 2027 [93]. Garantias de aluguel e de valor residual começam a ser cobradas. Fundos que marcam títulos de data centers a mercado reduzem os preços, e fundos semilíquidos de crédito privado, onde os pedidos de resgate já superaram as entradas no primeiro trimestre de 2026, enfrentam mais saídas [70]. É o cenário em que o wrong-way risk aparece: quem garante a dívida perde valor ao mesmo tempo que a garantia é chamada.

Não atribuo probabilidades a esses cenários. O argumento deste artigo é mais modesto e mais verificável: a estrutura de financiamento montada em 2025 e 2026 torna o terceiro cenário possível, de um jeito que não era possível quando o ciclo era pago com caixa e ações.

Quem perde primeiro

Pela estrutura de capital descrita nas seções anteriores, as perdas seguiriam uma ordem. Primeiro, os acionistas dos laboratórios e das neoclouds, incluindo os investidores individuais que entraram pelas rodadas abertas ao varejo. Depois, os credores das neoclouds, cujos empréstimos têm GPUs como garantia. Em seguida, os fornecedores que deram garantias ou compraram participações em clientes, como NVIDIA, AMD e Google, que perdem ao mesmo tempo receita e valor de investimento. Depois, os donos dos títulos dos veículos de data centers, se os aluguéis e as garantias de valor residual não cobrirem a dívida, entre eles seguradoras e fundos de renda fixa. Por último, as próprias hyperscalers, que continuam obrigadas a pagar os aluguéis que assinaram. É uma ordem provável, não garantida: em 2007, compromissos que pareciam distantes voltaram para o balanço antes da hora, por contrato ou por reputação.

O que acompanhar

Os indicadores abaixo são públicos e mostram, trimestre a trimestre, em que cenário o ciclo está:

Nenhum desses sinais, sozinho, decide a questão. Juntos, mostram se a conta está sendo paga pela receita ou empurrada para a frente.

Conclusão

A pergunta deste artigo nunca foi se a inteligência artificial é real. As ferrovias eram reais, a fibra óptica também, e as duas mudaram a economia depois de arruinar boa parte de quem as financiou. A pergunta é quem paga a conta se a receita demorar mais que a dívida.

Os seis marcadores da seção "Como se reconhece uma bolha" dão uma resposta parcial. O choque real está presente: o investimento das quatro maiores empresas de nuvem foi de US$ 228 bilhões em 2024 para US$ 376 bilhões em 2025 e pode passar de US$ 700 bilhões em 2026, e a receita de nuvem e de IA cresce. O crédito financiando a expansão também: as hyperscalers responderam por mais de 15% das emissões americanas de grau de investimento em 2026 até maio, e as emissões ligadas à IA já tinham passado, em setembro, do dobro de todo o ano de 2025. A inovação nos passivos é evidente, das joint ventures não consolidadas às garantias de valor residual, dos empréstimos com GPUs como garantia às plataformas para transformar chips em ativo de investimento, que o presidente da BlackRock comparou ao começo do mercado de títulos hipotecários. A receita que depende do próprio financiamento está medida: US$ 1,4 trilhão em compromissos nos acordos que o FMI chamou de circulares, e 55,2% do investimento recebido por empresas de IA vindo de outras empresas de IA, segundo o BIS. Os dois últimos marcadores aparecem só em parte. O colateral envelhece rápido no papel, com uma geração de chips por ano e depreciação de seis, mas o aluguel das GPUs antigas subiu em 2026 por causa da escassez. E os preços estão altos, com o CAPE perto de 41 e a concentração da bolsa em níveis recordes, mas apoiados em lucros que existem.

O que 2008 ensina se aplica quase ponto por ponto. Compromissos fora do balanço voltam para dentro na crise, e hoje eles somam mais de US$ 1 trilhão em aluguéis não iniciados, além das garantias da Meta, da NVIDIA, da AMD e do Google. Gatilhos ligados a preço de mercado antecipam perdas, e os empréstimos das neoclouds são dimensionados pelo valor contábil das GPUs. Notas de crédito podem medir quem garante, e não o ativo: os títulos A+ da Beignet financiam um campus cujo único inquilino e garantidor é a Meta. Garantias que caem junto com o risco falham quando são chamadas, e o FMI já usa a expressão wrong-way risk para o ecossistema de IA. E métricas vigiadas pelo mercado viram alvo de engenharia: vidas úteis mais longas, uma reclassificação de aluguéis que reduz o capex divulgado sem mudar o investimento, uma nova medida de capex líquido de pagamentos antecipados de clientes.

As diferenças também precisam ser ditas. As hyperscalers têm caixa e pouca alavancagem, ao contrário dos bancos de 2007, o crédito privado não depende de dinheiro de curtíssimo prazo como os veículos de papel comercial, a divulgação é melhor e as regras de consolidação são mais duras. Não há milhões de famílias endividadas na ponta. Se o ciclo quebrar, o mais provável é algo entre 2000 e 2008: uma correção de bolsa, com perdas de crédito concentradas em laboratórios, neoclouds, fornecedores que deram garantias e os fundos que compraram a dívida deles.

O que preocupa é a direção. O ciclo começou pago com caixa e ações e está migrando, depressa, para dívida, boa parte dela fora dos balanços de quem usa os ativos. Até outubro de 2026, a conta da IA estava espalhada em notas explicativas, garantias contingentes e veículos que ninguém precisa consolidar. É exatamente o tipo de conta que, em 2008, só apareceu inteira quando já era tarde para dividi-la.

Referências

1. [1]Bank of England, “Financial Stability Report – July 2026,” Bank of England, 7 jul. 2026. [Online]. Disponível em: https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026 ↩
2. [2]U.S. Securities and Exchange Commission, “Company facts: Microsoft Corporation (CIK 0000789019),” EDGAR XBRL API, 13 out. 2026. [Online]. Disponível em: https://data.sec.gov/api/xbrl/companyfacts/CIK0000789019.json ↩
3. [3]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2026,” U.S. Securities and Exchange Commission, 29 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm ↩
4. [4]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, 5 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm ↩
5. [5]Alphabet Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, 23 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm ↩
6. [6]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, 6 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm ↩
7. [7]Amazon.com, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, 31 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm ↩
8. [8]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, 29 jan. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm ↩
9. [9]Meta Platforms, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, 30 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm ↩
10. [10]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2026,” U.S. Securities and Exchange Commission, 22 jun. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm ↩
11. [11]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2023,” U.S. Securities and Exchange Commission, 27 jul. 2023. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/789019/000095017023035122/msft-20230630.htm ↩
12. [12]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2025,” U.S. Securities and Exchange Commission, 18 jun. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000095017025087926/orcl-20250531.htm ↩
13. [13]Microsoft Corporation, “Microsoft Fiscal Year 2026 Fourth Quarter Earnings Conference Call,” Microsoft Investor Relations, 29 jul. 2026. [Online]. Disponível em: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 ↩
14. [14]Alphabet Inc., “2026 Q2 Earnings Call (transcrição),” Alphabet Investor Relations, 22 jul. 2026. [Online]. Disponível em: https://abc.xyz/investor/events/event-details/2026/2026-Q2-Earnings-Call-2026-GgTAq7Is0z/default.aspx ↩
15. [15]A. Palmer, “Amazon hikes 2026 capex to $220 billion due to higher memory costs,” CNBC, 30 jul. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html ↩
16. [16]Meta Platforms, Inc., “Meta Reports Second Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 29 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm ↩
17. [17]H. Son, “Moody's says 'unprecedented' AI spending threatens credit quality of Amazon, Meta, Alphabet and others,” CNBC, 24 jul. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/07/24/moodys-ai-spending-credit-quality-amazon-meta-alphabet.html ↩
18. [18]Oracle Corporation, “Form 10-Q for the quarterly period ended August 31, 2026,” U.S. Securities and Exchange Commission, 11 set. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm ↩
19. [19]Alphabet Inc., “2025 Q4 Earnings Call (transcrição),” Alphabet Investor Relations, 4 fev. 2026. [Online]. Disponível em: https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx ↩
20. [20]Microsoft Corporation, “Microsoft Fiscal Year 2026 Third Quarter Earnings Conference Call,” Microsoft Investor Relations, 29 abr. 2026. [Online]. Disponível em: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3 ↩
21. [21]Meta Platforms, Inc., “Meta Reports First Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 29 abr. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028364/meta-03312026xexhibit991.htm ↩
22. [22]International Energy Agency, “Key Questions on Energy and AI,” resumo executivo, 16 abr. 2026. [Online]. Disponível em: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ↩
23. [23]U.S. Bureau of Economic Analysis, “National Income and Product Accounts: arquivos de dados anuais e trimestrais (NipaDataA.txt e NipaDataQ.txt),” versão de 30 set. 2026. [Online]. Disponível em: https://apps.bea.gov/national/Release/TXT/NipaDataA.txt ↩
24. [24]A. Odlyzko, “Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s,” University of Minnesota, versão preliminar, 15 jan. 2010. [Online]. Disponível em: https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf ↩
25. [25]Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 30 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm ↩
26. [26]C. P. Kindleberger e R. Z. Aliber, “Financial Crisis: A Hardy Perennial,” cap. 1 de Manias, Panics, and Crashes: A History of Financial Crises, 5. ed. Hoboken: John Wiley & Sons, 2005 (excerto do editor). [Online]. Disponível em: https://media.wiley.com/product_data/excerpt/46/04714671/0471467146.pdf ↩
27. [27]C. P. Kindleberger e R. Z. Aliber, “Anatomy of a Typical Crisis,” cap. 2 de Manias, Panics and Crashes: A History of Financial Crises, 5. ed. Basingstoke: Palgrave Macmillan, 2005 (prévia do editor). [Online]. Disponível em: https://page-one.springer.com/pdf/preview/10.1057/9780230628045_2 ↩
28. [28]H. P. Minsky, “The Financial Instability Hypothesis,” Working Paper No. 74, The Jerome Levy Economics Institute of Bard College, maio 1992. [Online]. Disponível em: https://www.levyinstitute.org/pubs/wp74.pdf ↩
29. [29]C. Perez, “Great Surges of development and alternative forms of globalization,” Working Papers in Technology Governance and Economic Dynamics No. 15, The Other Canon Foundation e Tallinn University of Technology, jul. 2007. [Online]. Disponível em: https://hum.ttu.ee/wp/paper15.pdf ↩
30. [30]C. Perez, “The double bubble at the turn of the Century: Technological roots and structural implications,” CFAP/CERF Working Paper No. 31, University of Cambridge, 10 maio 2009; publicado no Cambridge Journal of Economics, v. 33, n. 4, pp. 779–805, 2009. [Online]. Disponível em: https://cp.informaticaprojects.com/wp-content/downloads/publications/theoretical-framework/C.PEREZ_CJE_DoubleBubble_WP%20CERF%20No%2031.pdf ↩
31. [31]International Monetary Fund, “Chapter 1,” em Global Financial Stability Report: Global Financial Markets Confront the War in the Middle East and Amplification Risks, 14 abr. 2026. [Online]. Disponível em: https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf ↩
32. [32]R. J. Shiller, “Speculative Asset Prices,” Prize Lecture, 8 dez. 2013. [Online]. Disponível em: https://www.nobelprize.org/uploads/2018/06/shiller-lecture.pdf ↩
33. [33]R. J. Shiller, “U.S. Stock Markets 1871-Present and CAPE Ratio,” planilha ie_data.xls, atualizada em 5 out. 2026. [Online]. Disponível em: https://shillerdata.com/ ↩
34. [34]A. Greenspan, “The Challenge of Central Banking in a Democratic Society,” discurso no American Enterprise Institute, Board of Governors of the Federal Reserve System, 5 dez. 1996. [Online]. Disponível em: https://www.federalreserve.gov/boarddocs/speeches/1996/19961205.htm ↩
35. [35]A. Odlyzko, “This time is different: An example of a giant, wildly speculative, and successful investment mania,” The B.E. Journal of Economic Analysis & Policy, v. 10, n. 1, art. 60, 2010. [Online]. Disponível em: https://www-users.cse.umn.edu/~odlyzko/doc/mania01.pdf ↩
36. [36]A. Odlyzko, “The Railway Mania: Fraud, disappointed expectations, and the modern economy,” Journal of the Railway & Canal Historical Society, n. 215, pp. 2–12, nov. 2012 (versão revisada de 14 ago. 2012). [Online]. Disponível em: https://www-users.cse.umn.edu/~odlyzko/doc/mania06.pdf ↩
37. [37]C. P. Kindleberger, “Bubbles in History,” em The New Palgrave Dictionary of Economics, 2. ed. Basingstoke: Palgrave Macmillan, 2008 (prévia do editor). [Online]. Disponível em: https://page-one.springer.com/pdf/preview/10.1057/978-1-349-95121-5_3073-1 ↩
38. [38]G. Richardson e T. Sablik, “Banking Panics of the Gilded Age,” Federal Reserve History, Federal Reserve System, 4 dez. 2015. [Online]. Disponível em: https://www.federalreservehistory.org/essays/banking-panics-of-the-gilded-age ↩
39. [39]K. Giesecke, F. A. Longstaff, S. Schaefer e I. Strebulaev, “Corporate Bond Default Risk: A 150-Year Perspective,” NBER Working Paper n. 15848, mar. 2010 (© 2010 by Kay Giesecke, Francis A. Longstaff, Stephen Schaefer, and Ilya Strebulaev). [Online]. Disponível em: https://www.nber.org/system/files/working_papers/w15848/w15848.pdf ↩
40. [40]National Bureau of Economic Research, “US Business Cycle Expansions and Contractions,” dados atualizados em 14 mar. 2023. [Online]. Disponível em: https://www.nber.org/research/data/us-business-cycle-expansions-and-contractions ↩
41. [41]S. Mixon, “The Crisis of 1873: Perspectives from Multiple Asset Classes,” The Journal of Economic History, v. 68, n. 3, pp. 722–757, set. 2008. [Online]. Disponível em: https://doi.org/10.1017/S0022050708000624 ↩
42. [42]E. A. Couper, J. P. Hejkal e A. L. Wolman, “Boom and Bust in Telecommunications,” Economic Quarterly, Federal Reserve Bank of Richmond, v. 89, n. 4, outono 2003. [Online]. Disponível em: https://www.richmondfed.org/-/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf ↩
43. [43]A. M. Odlyzko, “Internet traffic growth: Sources and implications,” em Optical Transmission Systems and Equipment for WDM Networking II, Proc. SPIE, v. 5247, pp. 1–15, 2003. [Online]. Disponível em: https://www-users.cse.umn.edu/~odlyzko/doc/itcom.internet.growth.pdf ↩
44. [44]J. M. Kraushaar, “Fiber Deployment Update — End of Year 1998,” Industry Analysis Division, Common Carrier Bureau, Federal Communications Commission, 9 set. 1999, cópia do Internet Archive. [Online]. Disponível em: https://web.archive.org/web/20000817010609/http://www.fcc.gov:80/Bureaus/Common_Carrier/Reports/FCC-State_Link/Fiber/fiber98.pdf ↩
45. [45]M. Doms, “The Boom and Bust in Information Technology Investment,” FRBSF Economic Review, Federal Reserve Bank of San Francisco, pp. 19–34, 2004. [Online]. Disponível em: https://www.frbsf.org/wp-content/uploads/sites/4/er19-34bk.pdf ↩
46. [46]NASDAQ OMX Group, “NASDAQ Composite Index [NASDAQCOM],” FRED, Federal Reserve Bank of St. Louis, série diária consultada em out. 2026. [Online]. Disponível em: https://fred.stlouisfed.org/series/NASDAQCOM ↩
47. [47]Bank for International Settlements, 72nd Annual Report, 1 April 2001–31 March 2002, Basileia, 8 jul. 2002. [Online]. Disponível em: https://www.bis.org/publications/aer-2002.pdf ↩
48. [48]Lucent Technologies Inc., “Form 10-K405 for the fiscal year ended September 30, 2001,” U.S. Securities and Exchange Commission, 28 dez. 2001. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1006240/000095011701501896/a31752.txt ↩
49. [49]Nortel Networks Corporation, “Form 10-K for the fiscal year ended December 31, 2001,” U.S. Securities and Exchange Commission, 11 mar. 2002. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/72911/000113031902000168/t06646e10-k.htm ↩
50. [50]Cisco Systems, Inc., “Form 10-K for the fiscal year ended July 28, 2001,” U.S. Securities and Exchange Commission, 24 set. 2001. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/858877/000109581101505065/f75710e10-k.txt ↩
51. [51]J. Frost, R. Kansal, K. Rishabh, V. Shreeti e L. S. Y. Zhang, “Circular relationships among AI firms,” BIS Bulletin, n. 137, Bank for International Settlements, 1 out. 2026. [Online]. Disponível em: https://www.bis.org/publications/bulletin-137-circular-relationships-among-ai-firms.pdf ↩
52. [52]U.S. Securities and Exchange Commission, “SEC Charges Qwest Communications International Inc. with Multi-Faceted Accounting and Financial Reporting Fraud,” Litigation Release n. 18936, 21 out. 2004. [Online]. Disponível em: https://www.sec.gov/litigation/litreleases/lr18936.htm ↩
53. [53]U.S. Securities and Exchange Commission, “In the Matter of Global Crossing Ltd., Thomas J. Casey, Dan J. Cohrs, and Joseph P. Perrone,” Exchange Act Release n. 51517, 11 abr. 2005. [Online]. Disponível em: https://www.sec.gov/files/litigation/admin/34-51517.pdf ↩
54. [54]U.S. Securities and Exchange Commission, “SEC Charges WorldCom with $3.8 Billion Fraud,” Litigation Release n. 17588, 27 jun. 2002. [Online]. Disponível em: https://www.sec.gov/litigation/litreleases/lr17588.htm ↩
55. [55]WorldCom, Inc., “Form 10-K for the fiscal year ended December 31, 2002,” U.S. Securities and Exchange Commission, 12 mar. 2004. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/723527/000119312504039709/d10k.htm ↩
56. [56]W. C. Powers Jr., R. S. Troubh e H. S. Winokur Jr., “Report of Investigation by the Special Investigative Committee of the Board of Directors of Enron Corp.,” 1 fev. 2002, anexo 99.2 ao Form 8-K da Enron Corp. de 7 fev. 2002. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1024401/000090951802000089/big.txt ↩
57. [57]U.S. Securities and Exchange Commission, Office of the Chief Accountant, Office of Economic Analysis e Division of Corporation Finance, “Report and Recommendations Pursuant to Section 401(c) of the Sarbanes-Oxley Act of 2002 On Arrangements with Off-Balance Sheet Implications, Special Purpose Entities, and Transparency of Filings by Issuers,” jun. 2005. [Online]. Disponível em: https://www.sec.gov/news/studies/soxoffbalancerpt.pdf ↩
58. [58]Financial Accounting Standards Board, “Summary of Interpretation No. 46: Consolidation of Variable Interest Entities—an interpretation of ARB No. 51,” jan. 2003, cópia do Internet Archive. [Online]. Disponível em: https://web.archive.org/web/20040611063520/http://www.fasb.org:80/st/summary/finsum46.shtml ↩
59. [59]Financial Accounting Standards Board, “Summary of Interpretation No. 46 (revised December 2003),” dez. 2003, cópia do Internet Archive. [Online]. Disponível em: https://web.archive.org/web/20040611052814/http://www.fasb.org:80/st/summary/finsum46r.shtml ↩
60. [60]Financial Crisis Inquiry Commission, The Financial Crisis Inquiry Report: Final Report of the National Commission on the Causes of the Financial and Economic Crisis in the United States. Washington: U.S. Government Publishing Office, jan. 2011. [Online]. Disponível em: https://www.govinfo.gov/content/pkg/GPO-FCIC/pdf/GPO-FCIC.pdf ↩
61. [61]Board of Governors of the Federal Reserve System, “Asset-Backed Commercial Paper Outstanding [ABCOMP],” FRED, Federal Reserve Bank of St. Louis, série semanal consultada em 13 out. 2026. [Online]. Disponível em: https://fred.stlouisfed.org/series/ABCOMP ↩
62. [62]D. M. Covitz, N. Liang e G. A. Suarez, “The Evolution of a Financial Crisis: Panic in the Asset-Backed Commercial Paper Market,” Finance and Economics Discussion Series 2009-36, Federal Reserve Board, 18 ago. 2009. [Online]. Disponível em: https://www.federalreserve.gov/pubs/feds/2009/200936/200936pap.pdf ↩
63. [63]Citigroup Inc., “Citi Commits Support Facility for Citi-Advised SIVs,” comunicado de imprensa (anexo 99.1 ao Form 8-K), 13 dez. 2007. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/831001/000114420407067635/v097312_ex99-1.htm ↩
64. [64]A. R. Valukas, “Report of Anton R. Valukas, Examiner,” v. 1, In re Lehman Brothers Holdings Inc., Chapter 11 Case No. 08-13555, U.S. Bankruptcy Court, S.D.N.Y., 11 mar. 2010, cópia do Internet Archive. [Online]. Disponível em: https://web.archive.org/web/20100331000019/http://lehmanreport.jenner.com:80/VOLUME%201.pdf ↩
65. [65]Basel Committee on Banking Supervision, “CRE50 – Counterparty credit risk definitions and terminology,” Basel Framework, versão publicada em 5 jul. 2024. [Online]. Disponível em: https://www.bis.org/basel_framework/chapter/CRE/50.htm ↩
66. [66]Basel Committee on Banking Supervision, “Strengthening the resilience of the banking sector,” documento consultivo, dez. 2009. [Online]. Disponível em: https://www.bis.org/publ/bcbs164.pdf ↩
67. [67]Financial Accounting Standards Board, “Accounting Standards Update No. 2009-17, Consolidations (Topic 810): Improvements to Financial Reporting by Enterprises Involved with Variable Interest Entities,” dez. 2009. [Online]. Disponível em: https://storage.fasb.org/ASU2009-17.pdf ↩
68. [68]Board of Governors of the Federal Reserve System, “H.8 Assets and Liabilities of Commercial Banks in the United States,” 9 abr. 2010. [Online]. Disponível em: https://www.federalreserve.gov/releases/h8/20100409/ ↩
69. [69]International Monetary Fund, “Chapter 2: The Rise and Risks of Private Credit,” em Global Financial Stability Report: The Last Mile: Financial Vulnerabilities and Risks, 16 abr. 2024. [Online]. Disponível em: https://www.imf.org/-/media/Files/Publications/GFSR/2024/April/English/ch2.ashx ↩
70. [70]Board of Governors of the Federal Reserve System, Financial Stability Report, maio 2026. [Online]. Disponível em: https://www.federalreserve.gov/publications/files/financial-stability-report-20260508.pdf ↩
71. [71]P. Abbassi, I. Aldasoro e S. Doerr, “Financing the digital economy: the role of private credit,” BIS Quarterly Review, Bank for International Settlements, 14 set. 2026. [Online]. Disponível em: https://www.bis.org/publications/financing-digital-economy-role-private-credit.pdf ↩
72. [72]Financial Stability Board, “Report on Vulnerabilities in Private Credit,” 6 maio 2026. [Online]. Disponível em: https://www.fsb.org/uploads/P060526.pdf ↩
73. [73]NVIDIA, “OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems,” NVIDIA Newsroom, 22 set. 2025. [Online]. Disponível em: https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems ↩
74. [74]NVIDIA Corporation, “Form 10-Q — quarterly report for the period ended October 26, 2025,” U.S. Securities and Exchange Commission (EDGAR), 19 nov. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1045810/000104581025000230/nvda-20251026.htm ↩
75. [75]OpenAI, “Scaling AI for everyone,” OpenAI, 27 fev. 2026. [Online]. Disponível em: https://openai.com/index/scaling-ai-for-everyone/ ↩
76. [76]A. Capoot, “Nvidia CEO Huang says $30 billion OpenAI investment 'might be the last',” CNBC, 4 mar. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/03/04/nvidia-huang-openai-investment.html ↩
77. [77]NVIDIA Corporation, “Form 10-Q for the quarterly period ended July 26, 2026,” U.S. Securities and Exchange Commission, 26 ago. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm ↩
78. [78]NVIDIA Corporation, “Form 10-K — annual report for the fiscal year ended January 25, 2026,” U.S. Securities and Exchange Commission (EDGAR), 25 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm ↩
79. [79]Advanced Micro Devices, Inc., “Form 8-K” e anexo 99.1, “AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs,” U.S. Securities and Exchange Commission, 6 out. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/2488/000119312525230895/d28189d8k.htm ↩
80. [80]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended September 27, 2025,” U.S. Securities and Exchange Commission, 5 nov. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/2488/000000248825000166/amd-20250927.htm ↩
81. [81]Advanced Micro Devices, Inc., “Form 8-K” (warrant emitido à Meta Platforms), U.S. Securities and Exchange Commission, 24 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/2488/000000248826000045/amd-20260223.htm ↩
82. [82]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended June 27, 2026,” U.S. Securities and Exchange Commission, 5 ago. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/2488/000000248826000123/amd-20260627.htm ↩
83. [83]Microsoft, “The next chapter of the Microsoft–OpenAI partnership,” Official Microsoft Blog, 28 out. 2025. [Online]. Disponível em: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/ ↩
84. [84]Microsoft, “Microsoft Fiscal Year 2026 Second Quarter Earnings Conference Call,” Microsoft Investor Relations, 28 jan. 2026. [Online]. Disponível em: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q2 ↩
85. [85]OpenAI, “OpenAI and Amazon announce strategic partnership,” OpenAI, 27 fev. 2026. [Online]. Disponível em: https://openai.com/index/amazon-partnership/ ↩
86. [86]Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute,” Anthropic, 20 abr. 2026. [Online]. Disponível em: https://www.anthropic.com/news/anthropic-amazon-compute ↩
87. [87]Anthropic, “Anthropic raises $65B in Series H funding at $965B post-money valuation,” Anthropic, 28 maio 2026. [Online]. Disponível em: https://www.anthropic.com/news/series-h ↩
88. [88]Anthropic, “Microsoft, NVIDIA, and Anthropic announce strategic partnerships,” Anthropic, 18 nov. 2025. [Online]. Disponível em: https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships ↩
89. [89]Anthropic, “Expanding our use of Google Cloud TPUs and Services,” Anthropic, 23 out. 2025. [Online]. Disponível em: https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services ↩
90. [90]CoreWeave, Inc., “CoreWeave Announces Agreement with OpenAI to Deliver AI Infrastructure,” CoreWeave Investor Relations, 10 mar. 2025. [Online]. Disponível em: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Announces-Agreement-with-OpenAI-to-Deliver-AI-Infrastructure/default.aspx ↩
91. [91]CoreWeave, Inc., “CoreWeave Expands Agreement with OpenAI by up to $6.5B,” CoreWeave Investor Relations, 25 set. 2025. [Online]. Disponível em: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Expands-Agreement-with-OpenAI-by-up-to-6-5B/default.aspx ↩
92. [92]CoreWeave, Inc., “Current Report (Form 8-K) — Order Form with NVIDIA Corporation,” U.S. Securities and Exchange Commission (EDGAR), 15 set. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm ↩
93. [93]CoreWeave, Inc., “Quarterly Report (Form 10-Q) for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission (EDGAR), 12 ago. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm ↩
94. [94]OpenAI, “OpenAI and SoftBank Group partner with SB Energy,” OpenAI, 9 jan. 2026. [Online]. Disponível em: https://openai.com/index/stargate-sb-energy-partnership/ ↩
95. [95]J. Bort, “Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments,” TechCrunch, 6 nov. 2025. [Online]. Disponível em: https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/ ↩
96. [96]A. Capoot e K. Rooney, “Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report,” CNBC, 8 out. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html ↩
97. [97]International Monetary Fund, “Global Financial Stability Report, October 2026: Resilience, Risk, and Transformation,” IMF, 12 out. 2026. [Online]. Disponível em: https://doi.org/10.5089/9798229055673.082 ↩
98. [98]A. Bao, “IMF warns bond yields could see a 'disorderly' spike as debt, inflation risks worsen,” CNBC, 13 out. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/10/13/treasury-yields-imf-debt-inflation-.html ↩
99. [99]OpenAI e Broadcom, “OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators,” OpenAI, 13 out. 2025. [Online]. Disponível em: https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/ ↩
100. [100]Meta Platforms, “Meta Announces Joint Venture With Funds Managed by Blue Owl Capital to Develop Hyperion Data Center,” Meta Newsroom, 21 out. 2025. [Online]. Disponível em: https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/ ↩
101. [101]C. Arroyo e L. Benitez, “Meta, Blue Owl Seal $30 Billion Private Capital Deal for AI,” Bloomberg, 16 out. 2025. [Online]. Disponível em: https://www.bloomberg.com/news/articles/2025-10-16/blue-owl-seals-largest-private-capital-deal-for-meta-s-ai-growth ↩
102. [102]Blue Owl Real Estate Net Lease Trust, “Form 10-Q: quarterly report for the period ended September 30, 2025,” U.S. Securities and Exchange Commission (EDGAR), 7 nov. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1944366/000194436625000126/osnl-20250930.htm ↩
103. [103]Meta Platforms, Inc., “Form 10-Q: quarterly report for the period ended March 31, 2026,” U.S. Securities and Exchange Commission (EDGAR), 30 abr. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm ↩
104. [104]Deloitte, “Roadmap: Consolidation — 7.1 Introduction (Determining the Primary Beneficiary),” DART – Deloitte Accounting Research Tool, set. 2026. [Online]. Disponível em: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc810-10/roadmap-consolidation/chapter-7-determining-primary-beneficiary/7-1-introduction ↩
105. [105]J. Weil, “AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center,” The Wall Street Journal, 24 nov. 2025. [Online]. Disponível em: https://www.wsj.com/tech/meta-ai-data-center-finances-d3a6b464 ↩
106. [106]Deloitte, “Roadmap: Leases — 6.7 Amounts That It Is Probable That the Lessee Will Owe Under a Residual Value Guarantee,” DART – Deloitte Accounting Research Tool, ago. 2026. [Online]. Disponível em: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-6-lease-payments/6-7-amounts-that-it-is ↩
107. [107]Deloitte, “Roadmap: Leases — 8.4 Recognition and Measurement,” DART – Deloitte Accounting Research Tool, ago. 2026. [Online]. Disponível em: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-8-lessee-accounting/8-4-recognition-measurement ↩
108. [108]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q2_nport.zip,” SEC Division of Economic and Risk Analysis, 8 jul. 2026. [Online]. Disponível em: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q2_nport.zip ↩
109. [109]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q1_nport.zip,” SEC Division of Economic and Risk Analysis, 3 abr. 2026. [Online]. Disponível em: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q1_nport.zip ↩
110. [110]C. Arroyo e E. Ludlow, “xAI to Raise $20 Billion After Nvidia and Others Boost Round,” Bloomberg, 7 out. 2025. [Online]. Disponível em: https://www.bloomberg.com/news/articles/2025-10-07/musk-s-xai-nears-20-billion-capital-raise-tied-to-nvidia-chips ↩
111. [111]C. Arroyo, “How AI Companies Are Keeping Debt Off Their Balance Sheets,” Bloomberg, 31 out. 2025. [Online]. Disponível em: https://www.bloomberg.com/news/articles/2025-10-31/meta-xai-starting-trend-for-billions-in-off-balance-sheet-debt ↩
112. [112]Apollo Global Management, “Apollo Backs $5.4 Billion Valor and xAI Data Center Compute Infrastructure Transaction with $3.5 Billion Capital Solution,” Apollo (GlobeNewswire), 7 jan. 2026. [Online]. Disponível em: https://www.apollo.com/insights-news/pressreleases/2026/01/apollo-backs-5-4-billion-valor-and-xai-data-center-compute-infrastructure-transaction-with-3-5-billion-capital-solution-3214463 ↩
113. [113]NVIDIA Corporation, “Form 8-K (itens 1.01, 2.03 e 7.01),” U.S. Securities and Exchange Commission, 17 ago. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm ↩
114. [114]A. Capoot e K. Rooney, “Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans,” CNBC, 27 jul. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html ↩
115. [115]NVIDIA, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” NVIDIA Newsroom, 17 ago. 2026. [Online]. Disponível em: https://nvidianews.nvidia.com/news/nvidia-guarantees-sb-energy-s-ports-pike-technology-campus-in-ohio-to-exclusively-host-nvidia-ai-compute ↩
116. [116]TeraWulf Inc., “TeraWulf Announces Fluidstack Expansion with 160 MW CB-5 Lease at Lake Mariner,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 18 ago. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1083301/000110465925079463/tm2523651d3_ex99-1.htm ↩
117. [117]Cipher Mining Inc., “Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 25 set. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1819989/000095010325012168/dp234624_ex9901.htm ↩
118. [118]Hut 8 Corp., “Hut 8 Signs 15-Year, 245 MW AI Data Center Lease at River Bend Campus with Total Contract Value of $7.0 Billion,” Form 8-K, Exhibit 99.1 (SEC EDGAR), 17 dez. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1964789/000110465925122052/hut-20251217xex99d1.htm ↩
119. [119]NVIDIA, “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital,” NVIDIA Newsroom, 10 ago. 2026. [Online]. Disponível em: https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital ↩
120. [120]K. Leswing, “Wall Street just endorsed Jensen Huang's ‘big concept’ for AI. What now?,” CNBC, 11 ago. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/08/11/wall-street-endorsed-jensen-huangs-big-concept-for-ai-what-now.html ↩
121. [121]Bank for International Settlements, “Markets recalibrate amid shifting currents,” BIS Quarterly Review, 16 mar. 2026. [Online]. Disponível em: https://www.bis.org/publ/qtrpdf/r_qt2603a.htm ↩
122. [122]Bank for International Settlements, “I. Progress and peril,” BIS Annual Economic Report 2026, 28 jun. 2026. [Online]. Disponível em: https://www.bis.org/publications/aer-2026/progress-peril ↩
123. [123]D. Gautam, J. Kramer, H. Kraus e A. Yokoyama, “Background Analysis 1.1. Rising Leverage in the AI Ecosystem,” Global Financial Stability Report, October 2026 (IMF), 12 out. 2026. [Online]. Disponível em: https://www.elibrary.imf.org/display/book/9798229055673/9798229055673.xml ↩
124. [124]CoreWeave, Inc., “Form S-1 Registration Statement,” U.S. Securities and Exchange Commission, 3 mar. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm ↩
125. [125]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2024,” U.S. Securities and Exchange Commission, 7 fev. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm ↩
126. [126]B. Caulfield, “‘Accelerate Everything,’ NVIDIA CEO Says Ahead of COMPUTEX,” NVIDIA Blog, 2 jun. 2024. [Online]. Disponível em: https://blogs.nvidia.com/blog/computex-2024-jensen-huang/ ↩
127. [127]NVIDIA, “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing,” comunicado de imprensa, 18 mar. 2024. [Online]. Disponível em: https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing ↩
128. [128]NVIDIA, “NVIDIA GB200 NVL72,” página do produto, versão de 27 ago. 2026, consultada em 13 out. 2026. [Online]. Disponível em: https://www.nvidia.com/en-us/data-center/gb200-nvl72/ ↩
129. [129]NVIDIA, “NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer,” NVIDIA Technical Blog, 16 mar. 2026. [Online]. Disponível em: https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/ ↩
130. [130]S. Koparkar, “Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents,” NVIDIA Blog, 24 ago. 2026. [Online]. Disponível em: https://blogs.nvidia.com/blog/vera-rubin-nvl72-efficiency-ai-agents/ ↩
131. [131]NVIDIA, “NVIDIA H100 GPU,” página do produto, versão de 1 jul. 2026, consultada em 13 out. 2026. [Online]. Disponível em: https://www.nvidia.com/en-us/data-center/h100/ ↩
132. [132]NVIDIA, “NVIDIA DGX B200,” página do produto, versão de 29 abr. 2026, consultada em 13 out. 2026. [Online]. Disponível em: https://www.nvidia.com/en-us/data-center/dgx-b200/ ↩
133. [133]Llama Team, AI @ Meta, “The Llama 3 Herd of Models,” arXiv:2407.21783, 23 jul. 2024. [Online]. Disponível em: https://arxiv.org/abs/2407.21783 ↩
134. [134]M. Blake, M. Hsu, H. Petty e J. Huntington, “NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories,” NVIDIA Technical Blog, 20 maio 2025. [Online]. Disponível em: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/ ↩
135. [135]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2023,” U.S. Securities and Exchange Commission, 31 jan. 2024. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm ↩
136. [136]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2022,” U.S. Securities and Exchange Commission, 28 jul. 2022. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/789019/000156459022026876/msft-10k_20220630.htm ↩
137. [137]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2022,” U.S. Securities and Exchange Commission, 2 fev. 2023. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1326801/000132680123000013/meta-20221231.htm ↩
138. [138]M. Burry (@michaeljburry), “Understating depreciation by extending useful life of assets artificially boosts earnings…,” X, 10 nov. 2025. [Online]. Disponível em: https://x.com/michaeljburry/status/1987918650104283372 ↩
139. [139]J. Sor, “Michael Burry's latest argument against hyperscalers like Meta and Oracle echoes one made by an investor who shorted Enron,” Business Insider, 11 nov. 2025. [Online]. Disponível em: https://www.businessinsider.com/michael-burry-ai-stocks-nvidia-meta-oracle-hyperscalers-jim-chanos-2025-11 ↩
140. [140]Y. Li, “Nvidia name-checks Michael Burry in secret memo pushing back on AI bubble allegations,” CNBC, 25 nov. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/11/25/nvidia-pushes-back-on-charges-that-ai-investment-is-a-bubble.html ↩
141. [141]C. Yun, “Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances,” AWS News Blog, 5 jun. 2025. [Online]. Disponível em: https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/ ↩
142. [142]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2023-08-23),” AWS Pricing (captura Wayback), 25 ago. 2023. [Online]. Disponível em: https://web.archive.org/web/20230825130912/https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
143. [143]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2026-10-12),” AWS Pricing, 12 out. 2026. [Online]. Disponível em: https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
144. [144]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS (captura Wayback), 9 jun. 2025. [Online]. Disponível em: https://web.archive.org/web/20250609034829/https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
145. [145]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS, 13 out. 2026. [Online]. Disponível em: https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
146. [146]Lambda, “GPU cloud pricing,” Lambda (captura Wayback), 5 nov. 2025. [Online]. Disponível em: https://web.archive.org/web/20251105190524/https://lambda.ai/pricing ↩
147. [147]Lambda, “GPU cloud pricing: rent NVIDIA H100, H200, and B200,” Lambda (captura Wayback), 1 out. 2026. [Online]. Disponível em: https://web.archive.org/web/20261001114850/https://lambda.ai/pricing ↩
148. [148]K. Leswing, “The question everyone in AI is asking: How long before a GPU depreciates?,” CNBC, 14 nov. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html ↩
149. [149]E. Lopatto, “Nvidia’s new financial strategy does not compute,” The Verge, 19 ago. 2026. [Online]. Disponível em: https://www.theverge.com/ai-artificial-intelligence/981668/nvidias-goldman-blackrock-gpu-compute-asset ↩
150. [150]CME Group, “CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks,” CME Group Press Release, 11 ago. 2026. [Online]. Disponível em: https://www.cmegroup.com/media-room/press-releases/2026/8/11/cme_group_and_silicondatatolaunchcomputefuturesonoctober5tounloc.html ↩
151. [151]S. Koparkar, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” NVIDIA Blog, 1 out. 2026. [Online]. Disponível em: https://blogs.nvidia.com/blog/productive-durable-fungible-ai-factories/ ↩
152. [152]International Energy Agency, “Energy and AI,” resumo executivo, 10 abr. 2025 (licença CC BY 4.0). [Online]. Disponível em: https://www.iea.org/reports/energy-and-ai/executive-summary ↩
153. [153]A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet e D. Sartor, “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, LBNL-2001637, dez. 2024. [Online]. Disponível em: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf ↩
154. [154]J. Rand, A. Cheyette, C. Talley, S. Zhang, W. Gorman, R. H. Wiser, J. Seel, S. Jeong e F. Kahrl, “Queued Up: 2026 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2025,” Lawrence Berkeley National Laboratory, jun. 2026. [Online]. Disponível em: https://emp.lbl.gov/queues ↩
155. [155]NVIDIA, “NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint With Broad Industry Support,” comunicado de imprensa, 16 mar. 2026. [Online]. Disponível em: https://nvidianews.nvidia.com/news/nvidia-releases-vera-rubin-dsx-ai-factory-reference-design-and-omniverse-dsx-digital-twin-blueprint-with-broad-industry-support ↩
156. [156]GE Vernova Inc., “GE Vernova reports fourth quarter and full year 2025 financial results; increasing outlook with Prolec GE acquisition,” anexo 99 ao Form 8-K, U.S. Securities and Exchange Commission, 28 jan. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000012/gevpressrelease4q25.htm ↩
157. [157]GE Vernova Inc., “GE Vernova reports second quarter 2026 financial results and raises 2026 financial guidance,” anexo 99 ao Form 8-K, U.S. Securities and Exchange Commission, 22 jul. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000147/gevpressrelease2q26.htm ↩
158. [158]Constellation Energy, “Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to The Grid,” comunicado de imprensa, 20 set. 2024. [Online]. Disponível em: https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html ↩
159. [159]M. Terrell, “New nuclear clean energy agreement with Kairos Power,” The Keyword (Google), 14 out. 2024. [Online]. Disponível em: https://blog.google/company-news/outreach-and-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/ ↩
160. [160]Constellation Energy, “Constellation, Meta Sign 20-Year Deal for Clean, Reliable Nuclear Energy in Illinois,” comunicado de imprensa, 3 jun. 2025. [Online]. Disponível em: https://www.constellationenergy.com/news/2025/constellation-meta-sign-20-year-deal-for-clean-reliable-nuclear-energy-in-illinois.html ↩
161. [161]Talen Energy, “Talen Energy Expands Nuclear Energy Relationship with Amazon,” comunicado de imprensa, 11 jun. 2025. [Online]. Disponível em: https://ir.talenenergy.com/news-releases/news-release-details/talen-energy-expands-nuclear-energy-relationship-amazon ↩
162. [162]Constellation Energy, “U.S. Government Backs Constellation's Plan to Launch Crane Clean Energy Center, Adding 835 MWs of New Baseload Power to the Grid,” comunicado de imprensa, 18 nov. 2025. [Online]. Disponível em: https://www.constellationenergy.com/news/2025/11/us-government-backs-constellations-plan-to-launch-crane-clean-energy-center-adding-835-mws-of-new-baseload-power-to-the-grid.html ↩
163. [163]American Electric Power, “AEP Ohio, PUCO Staff, Ohio Consumers’ Counsel, Ohio Energy Group and Others Agree on How to Address Growing Data Center Power Needs,” 23 out. 2024. [Online]. Disponível em: https://www.aep.com/news/stories/view/9829/aep-ohio-puco-staff-ohio-consumers-counsel-ohio-energy-group-and-others-agree-on-how-to-address-growing-data-center-power-needs/ ↩
164. [164]Ohio Power Company, “Data Center Tariff,” página institucional, consultada em 13 out. 2026. [Online]. Disponível em: https://www.aepohio.com/company/about/rates/data-center-tariff ↩
165. [165]American Electric Power Company, Inc., “Form 10-Q for the quarterly period ended June 30, 2025,” U.S. Securities and Exchange Commission, 30 jul. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/4904/000000490425000141/aep-20250630.htm ↩
166. [166]Bank of England, “Financial Stability Report – December 2025,” Bank of England, 2 dez. 2025. [Online]. Disponível em: https://www.bankofengland.co.uk/financial-stability-report/2025/december-2025 ↩
167. [167]S. Friar, “A business that scales with the value of intelligence,” OpenAI, 18 jan. 2026. [Online]. Disponível em: https://openai.com/index/a-business-that-scales-with-the-value-of-intelligence/ ↩
168. [168]A. Capoot et al., “OpenAI closes record-breaking $122 billion funding round as anticipation builds for IPO,” CNBC, 31 mar. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/03/31/openai-funding-round-ipo.html ↩
169. [169]OpenAI, “OpenAI raises $122 billion to accelerate the next phase of AI,” OpenAI, 31 mar. 2026. [Online]. Disponível em: https://openai.com/index/accelerating-the-next-phase-ai/ ↩
170. [170]S. Muppidi, “OpenAI Says Its Business Will Burn $115 Billion Through 2029,” The Information, 6 set. 2025. [Online]. Disponível em: https://www.theinformation.com/articles/openai-says-business-will-burn-115-billion-2029 ↩
171. [171]S. Muppidi e S. Palazzolo, “OpenAI Boosts Revenue Forecasts, Predicts $111 Billion More Cash Burn Through 2030,” The Information, 20 fev. 2026. [Online]. Disponível em: https://www.theinformation.com/articles/openai-boost-revenue-forecasts-predicts-112-billion-cash-burn-2030 ↩
172. [172]A. Gardizy e A. Efrati, “OpenAI CEO and CFO Diverge on IPO Timing,” The Information, 5 abr. 2026. [Online]. Disponível em: https://www.theinformation.com/articles/openai-ceo-cfo-diverge-ipo-timing ↩
173. [173]E. Woo, “OpenAI Burned $3.7 Billion in First Three Months of 2026,” The Information, 16 jun. 2026. [Online]. Disponível em: https://www.theinformation.com/articles/openai-burned-3-7-billion-first-three-months-2026 ↩
174. [174]OpenAI, “Confidential submission of draft S-1 to the SEC,” OpenAI, 8 jun. 2026. [Online]. Disponível em: https://openai.com/index/openai-submits-confidential-s-1/ ↩
175. [175]SoftBank Group Corp., “Follow-on Investments in OpenAI,” comunicado de imprensa, 27 fev. 2026. [Online]. Disponível em: https://group.softbank/en/news/press/20260227 ↩
176. [176]SoftBank Group Corp., “Execution of Bridge Facility Agreement Primarily for the Follow-on Investments in OpenAI,” comunicado de imprensa, 27 mar. 2026. [Online]. Disponível em: https://group.softbank/en/news/press/20260327 ↩
177. [177]SoftBank Group Corp., “Issuance of Foreign Currency-Denominated Senior Notes,” comunicado de imprensa, 24 set. 2026. [Online]. Disponível em: https://group.softbank/en/news/press/20260924 ↩
178. [178]CNBC, “Mega-IPOs could signal market top, say analysts as SpaceX and OpenAI prep record floats,” CNBC, 22 maio 2026. [Online]. Disponível em: https://www.cnbc.com/2026/05/22/ipo-flurry-top-market-analysts-ai-spacex-musk-altman.html ↩
179. [179]The Information, “How OpenAI’s Web of Business Relationships Could Complicate Its IPO,” The Information, 22 jun. 2026. [Online]. Disponível em: https://www.theinformation.com/articles/openais-light-balance-sheet-face-hard-look-ipo ↩
180. [180]CNBC, “OpenAI CFO Sarah Friar says company isn't seeking government backstop, clarifying prior comment,” CNBC, 6 nov. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/11/06/openai-cfo-sarah-friar-says-company-is-not-seeking-government-backstop.html ↩
181. [181]A. Capoot e M. Sigalos, “Sam Altman says OpenAI will top $20 billion in annualized revenue this year, hundreds of billions by 2030,” CNBC, 6 nov. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/11/06/sam-altman-says-openai-will-top-20-billion-annual-revenue-this-year.html ↩
182. [182]P. Singh et al., “An analyst worries OpenAI will miss its 'extravagant' spending commitments. These stocks would be the most vulnerable,” CNBC, 7 nov. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/11/07/these-stocks-are-at-risk-if-openai-misses-its-extravagant-spending-plans.html ↩
183. [183]S. Subin, “OpenAI's revenue, growth estimates fall short as company races toward IPO: Report,” CNBC, 28 abr. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/04/28/openais-revenue-growth-estimates-fall-short-report.html ↩
184. [184]Oracle Corporation, “Oracle Announces Fiscal Year 2026 First Quarter Financial Results,” anexo 99.1 ao Form 8-K, U.S. Securities and Exchange Commission, 9 set. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312525199175/orcl-ex99_1.htm ↩
185. [185]U.S. Securities and Exchange Commission, “Company facts: Oracle Corporation (CIK 0001341439),” EDGAR XBRL API, 13 out. 2026. [Online]. Disponível em: https://data.sec.gov/api/xbrl/companyfacts/CIK0001341439.json ↩
186. [186]Oracle Corporation, “Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues,” anexo 99.1 ao Form 8-K, U.S. Securities and Exchange Commission, 10 set. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-ex99_1.htm ↩
187. [187]Oracle Corporation, “Oracle Announces Record Q4 and FY 2026 Results Driven by Cloud Infrastructure & Cloud Applications,” anexo 99.1 ao Form 8-K, U.S. Securities and Exchange Commission, 10 jun. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526265848/orcl-ex99_1.htm ↩
188. [188]S. Mody, “Oracle's credit default swaps are plummeting as financing plan boosts investor confidence,” CNBC, 2 fev. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/02/02/oracles-credit-default-swaps-are-plummeting-after-financing-plans.html ↩
189. [189]J. Novet, “Oracle beats on earnings, but stock drops on plans to raise another $20 billion,” CNBC, 10 jun. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/06/10/oracle-orcl-q4-earnings-report-2026.html ↩
190. [190]J. Novet, “Oracle CEO Magouyrk: 'Of course' OpenAI can pay $60 billion per year for cloud infrastructure,” CNBC, 13 out. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/10/13/oracle-ceo-magouyrk-of-course-openai-can-pay-60-billion-per-year.html ↩
191. [191]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, 24 set. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312525215807/d49252dfwp.htm ↩
192. [192]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, 2 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526033882/d44245dfwp.htm ↩
193. [193]Oracle Corporation, comunicado de 1 fev. 2026 sobre o plano de captação de 2026, registrado como Form FWP, U.S. Securities and Exchange Commission, 2 fev. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1341439/000119312526032650/d36096dfwp.htm ↩
194. [194]J. Novet, “Oracle stock has worst week since 2001 dot-com bust as AI financing concerns escalate,” CNBC, 26 jun. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/06/26/oracle-stock-ends-worst-week-since-2001-as-investors-dwell-on-finances.html ↩
195. [195]T. Burns, “What more expensive corporate debt could mean for the AI buildout,” CNBC, 26 jul. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/07/26/what-more-expensive-corporate-debt-could-mean-for-the-ai-buildout.html ↩
196. [196]Reuters, “Oracle sued by bondholders over losses tied to AI buildout,” publicado pela CNBC, 14 jan. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/01/14/oracle-sued-by-bondholders-over-losses-tied-to-ai-buildout.html ↩
197. [197]S. Subin, “Oracle sends 'force majeure' notice about data center project — stock drops 3%,” CNBC, 24 set. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/09/24/oracle-data-center-force-majeure.html ↩
198. [198]CoreWeave, Inc., “CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing,” CoreWeave Investor Relations, 31 mar. 2026. [Online]. Disponível em: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx ↩
199. [199]CoreWeave, Inc., “CoreWeave Closes $2.6 Billion Loan Facility, Expanding Financing Flexibility for AI Infrastructure,” CoreWeave Investor Relations, 10 ago. 2026. [Online]. Disponível em: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-2-6-Billion-Loan-Facility-Expanding-Financing-Flexibility-for-AI-Infrastructure/default.aspx ↩
200. [200]CoreWeave, Inc., “Annual Report (Form 10-K) for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission (EDGAR), 2 mar. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm ↩
201. [201]Nebius Group N.V., “Unaudited Condensed Consolidated Financial Statements for the Three and Six Months Ended June 30, 2026,” Form 6-K, Exhibit 99.2 (SEC EDGAR), 12 ago. 2026. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d2.htm ↩
202. [202]Nebius Group N.V., “Report of Foreign Private Issuer (Form 6-K) — Commercial Agreement with Microsoft,” U.S. Securities and Exchange Commission (EDGAR), 8 set. 2025. [Online]. Disponível em: https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm ↩
203. [203]K. Leswing e I. O'Brien, “Debt-hungry AI companies face increased risk as bond yields spike,” CNBC, 27 set. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/09/27/debt-hungry-data-center-companies-increased-risk-bond-yields-spike.html ↩
204. [204]N. Lichtenberg, “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says,” Fortune, 7 out. 2025. [Online]. Disponível em: https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist/ ↩
205. [205]Federal Reserve Bank of St. Louis, “ALFRED — Real GDP (GDPC1) e Real Private Fixed Investment: Information Processing Equipment and Software (A679RX1Q020SBEA), safra de dados de 25 set. 2025,” ALFRED, 25 set. 2025. [Online]. Disponível em: https://alfred.stlouisfed.org/series?seid=A679RX1Q020SBEA ↩
206. [206]U.S. Bureau of Economic Analysis, “NIPA Handbook: Concepts and Methods of the U.S. National Income and Product Accounts — Chapter 4: Estimating Methods,” BEA, dez. 2023. [Online]. Disponível em: https://www.bea.gov/resources/methodologies/nipa-handbook/pdf/chapter-04.pdf ↩
207. [207]International Monetary Fund, “Global Financial Stability Report, October 2025: Shifting Ground beneath the Calm,” IMF, 14 out. 2025. [Online]. Disponível em: https://www.imf.org/en/Publications/GFSR/Issues/2025/10/14/global-financial-stability-report-october-2025 ↩
208. [208]I. Aldasoro, S. Doerr e D. Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin, n. 120, 7 jan. 2026. [Online]. Disponível em: https://www.bis.org/publications/bulletin-120-financing-ai-boom-cash-flows-debt ↩
209. [209]U.S. Census Bureau, “Monthly Construction Spending, August 2026 (Release CB26-158),” U.S. Census Bureau, 1 out. 2026. [Online]. Disponível em: https://www.census.gov/construction/c30/pdf/release.pdf ↩
210. [210]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – October 2025,” Bank of England, 8 out. 2025. [Online]. Disponível em: https://www.bankofengland.co.uk/financial-policy-committee-record/2025/october-2025 ↩
211. [211]Bank for International Settlements, “Volatility challenges risk-taking,” BIS Quarterly Review, 8 dez. 2025. [Online]. Disponível em: https://www.bis.org/publ/qtrpdf/r_qt2512a.htm ↩
212. [212]European Central Bank, “Financial Stability Review, May 2026,” ECB, 27 maio 2026. [Online]. Disponível em: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202605~50566915a7.en.html ↩
213. [213]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – September 2026,” Bank of England, 30 set. 2026. [Online]. Disponível em: https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026 ↩
214. [214]Board of Governors of the Federal Reserve System, “Financial Stability Report – November 2025,” Federal Reserve Board, 7 nov. 2025. [Online]. Disponível em: https://www.federalreserve.gov/publications/files/financial-stability-report-20251107.pdf ↩
215. [215]European Central Bank, “Financial Stability Review, November 2025,” ECB, 26 nov. 2025. [Online]. Disponível em: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202511~263b5810d4.en.html ↩
216. [216]A. Bailey (Financial Stability Board), “FSB Chair’s letter to G20 Finance Ministers and Central Bank Governors: August 2026,” Financial Stability Board, 28 ago. 2026. [Online]. Disponível em: https://www.fsb.org/uploads/P310826.pdf ↩
217. [217]Bank for International Settlements, “Yields climb, yet risk appetite holds firm,” BIS Quarterly Review, 14 set. 2026. [Online]. Disponível em: https://www.bis.org/publications/qr-202609/yields-climb-yet-risk-appetite-holds-firm ↩
218. [218]K. Leswing, “Apple ends day as world's most valuable company, passing Nvidia,” CNBC, 27 jul. 2026. [Online]. Disponível em: https://www.cnbc.com/2026/07/27/apple-most-valuable-company-nvidia.html ↩
219. [219]Presidência da República, “Medida Provisória nº 1.318, de 17 de setembro de 2025,” Diário Oficial da União, 18 set. 2025. [Online]. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/mpv/mpv1318.htm ↩
220. [220]G. Alckmin, F. Haddad e A. Silveira, “EMI nº 00010/2025 MDIC MF MME (Exposição de Motivos da MP 1.318/2025),” Presidência da República, 16 set. 2025. [Online]. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/Exm/Exm-1318-25.pdf ↩
221. [221]MDIC, “MP cria o Redata, que estimula datacenters e impulsiona economia digital no Brasil,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, 17 set. 2025. [Online]. Disponível em: https://web.archive.org/web/20250929123902/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/setembro/mp-cria-o-redata-que-estimula-datacenters-e-impulsiona-economia-digital-no-brasil ↩
222. [222]Congresso Nacional, “Ato Declaratório do Presidente da Mesa do Congresso Nacional nº 11, de 2026,” Diário Oficial da União, 27 fev. 2026. [Online]. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/congresso/adc-11-mpv1.318.htm ↩
223. [223]W. Máximo, “Governo vai buscar diálogo para reativar incentivo para data centers,” Agência Brasil, 25 fev. 2026. [Online]. Disponível em: https://web.archive.org/web/20260226002349/https://agenciabrasil.ebc.com.br/economia/noticia/2026-02/governo-vai-buscar-dialogo-para-reativar-incentivo-para-data-centers ↩
224. [224]Câmara dos Deputados, “PL 278/2026: ficha de tramitação,” Câmara dos Deputados, 28 set. 2026. [Online]. Disponível em: https://www.camara.leg.br/proposicoesWeb/fichadetramitacao?idProposicao=2600838 ↩
225. [225]Senado Federal, “PL 278/2026 (Redata),” Senado Federal, 28 set. 2026. [Online]. Disponível em: https://www25.senado.leg.br/web/atividade/materias/-/materia/172786 ↩
226. [226]Presidência da República, “Lei nº 15.504, de 15 de setembro de 2026,” Diário Oficial da União (Edição Extra), 15 set. 2026. [Online]. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/lei/l15504.htm ↩
227. [227]Presidência da República, “Lei Complementar nº 237, de 15 de setembro de 2026,” Diário Oficial da União (Edição Extra), 15 set. 2026. [Online]. Disponível em: https://www.in.gov.br/web/dou/-/lei-complementar-n-237-de-15-de-setembro-de-2026-731891987 ↩
228. [228]P. Peduzzi, “Sancionado regime especial para incentivar instalação de datacenters,” Agência Brasil, 15 set. 2026. [Online]. Disponível em: https://agenciabrasil.ebc.com.br/economia/noticia/2026-09/sancionado-regime-especial-para-incentivar-instalacao-de-datacenters ↩
229. [229]Imprensa Nacional, “Busca no DOU: ‘Redata’ (15 set. a 13 out. 2026),” Diário Oficial da União, consultada em 13 out. 2026. [Online]. Disponível em: https://www.in.gov.br/consulta/-/buscar/dou?q=Redata&s=todos&exactDate=personalizado&publishFrom=15-09-2026&publishTo=13-10-2026 ↩
230. [230]Microsoft News Center Brasil, “Microsoft anuncia investimento de R$ 14,7 bilhões ao longo de três anos em infraestrutura de nuvem e IA e oferece treinamento em IA em larga escala para capacitar 5 milhões de pessoas no Brasil,” Microsoft, 26 set. 2024. [Online]. Disponível em: https://news.microsoft.com/pt-br/microsoft-anuncia-investimento-de-r-147-bilhoes-ao-longo-de-tres-anos-em-infraestrutura-de-nuvem-e-ia-e-oferece-treinamento-em-ia-em-larga-escala-para-capacitar-5-milhoes-de-pessoas-no-brasil/ ↩
231. [231]About Amazon Brasil, “AWS investirá mais de R$10 bilhões para expandir infraestrutura no Brasil,” Amazon, 11 set. 2024. [Online]. Disponível em: https://www.aboutamazon.com.br/noticias/aws/aws-investira-mais-de-r-10-bilhoes-para-expandir-infraestrutura-no-brasil ↩
232. [232]MDIC, “CZPE aprova 12 projetos para R$ 585 bilhões em novos investimentos, incluindo datacenter, metais e amônia verde,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, 3 nov. 2025. [Online]. Disponível em: https://web.archive.org/web/20251104021618/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/outubro/czpe-aprova-12-projetos-para-r-585-bilhoes-em-novos-investimentos-incluindo-datacenter-metais-e-amonia-verde ↩
233. [233]Governo do Estado do Ceará, “Com R$ 200 bilhões de investimento, Data Center do TikTok no Ceará será o maior do Brasil,” Portal do Governo do Ceará, 3 dez. 2025. [Online]. Disponível em: https://web.archive.org/web/20251203170929/https://www.ceara.gov.br/2025/12/03/com-r-200-bilhoes-de-investimento-data-center-do-tiktok-no-ceara-sera-o-maior-do-brasil/ ↩
234. [234]Governo do Estado do Ceará (Ascom ZPE), “Com novo data center, ZPE Ceará consolida o Estado como hub tecnológico do País,” Portal do Governo do Ceará, 5 fev. 2026. [Online]. Disponível em: https://web.archive.org/web/20260205195108/https://www.ceara.gov.br/2026/02/05/com-novo-data-center-zpe-ceara-consolida-o-estado-como-hub-tecnologico-do-pais/ ↩
235. [235]Scala Data Centers, “Com investimento inicial de R$ 3 bilhões, governo do RS e Scala Data Centers assinam acordo para o maior projeto de infraestrutura digital do estado,” Scala Data Centers, 11 set. 2024. [Online]. Disponível em: https://scaladatacenters.com/com-investimento-inicial-de-r-3-bilhoes-governo-do-rs-e-scala-data-centers-assinam-acordo-para-o-maior-projeto-de-infraestrutura-digital-do-estado/ ↩
236. [236]Governo do Estado do Rio Grande do Sul, “Ministério de Minas e Energia autoriza reserva energética para cidade de data centers no Rio Grande do Sul,” Portal do Estado do RS, 14 maio 2025. [Online]. Disponível em: https://estado.rs.gov.br/ministerio-de-minas-e-energia-autoriza-reserva-energetica-para-cidade-de-data-centers-no-rio-grande-do-sul ↩
237. [237]MME/SNTEP, “Portaria SNTEP/MME nº 2.942, de 9 de maio de 2025,” Diário Oficial da União, 13 maio 2025. [Online]. Disponível em: https://www.in.gov.br/web/dou/-/portaria-sntep/mme-n-2.942-de-9-de-maio-de-2025-628902215 ↩
238. [238]Elea Data Centers, “Elea assina MoU com Tapestry para assegurar capacidade energética de 1,5 GW da Rio AI City,” Elea Data Centers, 6 nov. 2025. [Online]. Disponível em: https://eleadatacenters.com/2025/11/06/elea-assina-mou-com-tapestry-para-assegurar-capacidade-energetica-de-15-gw-da-rio-ai-city/ ↩
239. [239]Elea Data Centers, “Elea anuncia a Rio AI City, projeto com capacidade de até 3,2GW em energia verde para impulsionar a Inteligência Artificial no mundo,” Elea Data Centers, 8 maio 2025. [Online]. Disponível em: https://eleadatacenters.com/2025/05/08/elea-anuncia-a-rio-ai-city/ ↩
240. [240]Elea Data Centers, “Elea Data Centers é adquirida pela I Squared Capital e inicia nova fase de expansão na América Latina,” Elea Data Centers, 28 abr. 2026. [Online]. Disponível em: https://eleadatacenters.com/2026/04/28/elea-data-centers-e-adquirida-pela-i-squared-capital-e-inicia-nova-fase-de-expansao-na-america-latina/ ↩
241. [241]Equinix, “Equinix fortalece liderança na América Latina com inauguração de novo data center SP6 em São Paulo,” Equinix Newsroom, 16 abr. 2026. [Online]. Disponível em: https://newsroom.equinix.com/2026-04-16-Equinix-fortalece-lideranca-na-America-Latina-com-inauguracao-de-novo-data-center-SP6-em-Sao-Paulo ↩
242. [242]EPE (Superintendência de Estudos Econômicos e Energéticos), “A Expansão da Carga de Data Centers no mercado elétrico brasileiro (III COPAM),” Empresa de Pesquisa Energética, 10 set. 2025. [Online]. Disponível em: https://web.archive.org/web/20251004104216/https://www.epe.gov.br/sites-pt/areas-de-atuacao/energia-eletrica/WorkshopCOPAM/%5bEPE%5d%20Datacenter_AllexYukizaki%20III%20COPAM_09-2025.pdf ↩
243. [243]Imprensa Nacional, “Busca no DOU: ‘alternativa de acesso à Rede Básica’ (1º jan. 2024 a 13 out. 2026),” Diário Oficial da União, 13 out. 2026. [Online]. Disponível em: https://www.in.gov.br/consulta/-/buscar/dou?q=%22alternativa%20de%20acesso%20%C3%A0%20Rede%20B%C3%A1sica%22&s=todos&exactDate=personalizado&publishFrom=01-01-2024&publishTo=13-10-2026 ↩
244. [244]ANEEL, “Resolução Normativa ANEEL nº 1.122, de 20 de maio de 2025,” Diário Oficial da União, 26 maio 2025. [Online]. Disponível em: https://www.in.gov.br/web/dou/-/resolucao-normativa-aneel-n-1.122-de-20-de-maio-de-2025-631621948 ↩
245. [245]EPE, “Balanço Energético Nacional 2026: Relatório Síntese, ano base 2025,” Empresa de Pesquisa Energética, 3 ago. 2026. [Online]. Disponível em: https://www.epe.gov.br/pt/publicacoes-dados-abertos/publicacoes/balanco-energetico-nacional-2026 ↩
246. [246]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Eólicas (dados abertos),” Operador Nacional do Sistema Elétrico, 12 out. 2026. [Online]. Disponível em: https://dados.ons.org.br/dataset/restricao_coff_eolica_usi ↩
247. [247]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Fotovoltaicas (dados abertos),” Operador Nacional do Sistema Elétrico, 12 out. 2026. [Online]. Disponível em: https://dados.ons.org.br/dataset/restricao_coff_fotovoltaica ↩
248. [248]Presidência da República, “Lei nº 15.269, de 24 de novembro de 2025 (conversão da MP 1.304/2025),” Diário Oficial da União, 24 nov. 2025. [Online]. Disponível em: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/lei/l15269.htm ↩
249. [249]Casa dos Ventos, “Casa dos Ventos bate recorde com captação de US$ 1,1 bi nos EUA,” Casa dos Ventos, 18 jun. 2026. [Online]. Disponível em: https://casadosventos.com.br/noticias/captacao-nos-estados-unidos ↩
250. [250]B3, “Relatório anual de BDRs 2025,” B3, 8 dez. 2025. [Online]. Disponível em: https://www.b3.com.br/data/files/5D/D3/83/92/6430B9105B12E5A9AC094EA8/Relatorio%20Anual%20BDRs%202025.pdf ↩
251. [251]CVM, “Resolução CVM nº 175, de 23 de dezembro de 2022: Anexo Normativo I (texto consolidado),” Comissão de Valores Mobiliários, 5 mar. 2025. [Online]. Disponível em: https://conteudo.cvm.gov.br/export/sites/cvm/legislacao/resolucoes/anexos/100/resol175consolid_Anexo01.pdf ↩
252. [252]Conselho Monetário Nacional, “Resolução CMN nº 4.994, de 24 de março de 2022 (com alterações da Resolução CMN nº 5.202/2025),” Banco Central do Brasil, 24 mar. 2022. [Online]. Disponível em: https://www.bcb.gov.br/estabilidadefinanceira/exibenormativo?tipo=Resolu%C3%A7%C3%A3o%20CMN&numero=4994 ↩
253. [253]Banco Central do Brasil, “Relatório de Estabilidade Financeira, v. 25, n. 1,” Banco Central do Brasil, 22 maio 2026. [Online]. Disponível em: https://www.bcb.gov.br/content/publicacoes/ref/202605/RELESTAB202605-refPub.pdf ↩
254. [254]Alphabet Inc., “Alphabet Announces Second Quarter 2026 Results,” Alphabet Investor Relations, 22 jul. 2026. [Online]. Disponível em: https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf ↩
255. [255]OpenAI, “DevDay 2026 Recap,” OpenAI, 29 set. 2026. [Online]. Disponível em: https://openai.com/index/devday-2026-recap/ ↩
256. [256]E. Brynjolfsson, D. Li e L. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, vol. 140, n. 2, pp. 889–942, 4 fev. 2025. [Online]. Disponível em: https://doi.org/10.1093/qje/qjae044 ↩
257. [257]A. Bick, A. Blandin e D. Deming, “The Impact of Generative AI on Work Productivity,” Federal Reserve Bank of St. Louis — On the Economy, 27 fev. 2025. [Online]. Disponível em: https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity ↩
258. [258]U.S. Census Bureau, “Business Trends and Outlook Survey (BTOS): National estimates (AI use, collection ending Oct. 4, 2026),” U.S. Census Bureau, 8 out. 2026. [Online]. Disponível em: https://www.census.gov/hfp/btos/downloads/National.xlsx ↩
259. [259]DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437, 27 dez. 2024. [Online]. Disponível em: https://arxiv.org/abs/2412.19437 ↩
260. [260]S. Subin, “Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history,” CNBC, 27 jan. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html ↩
261. [261]S. Meredith, “Nvidia becomes first company to reach $5 trillion valuation, fueled by AI boom,” CNBC, 29 out. 2025. [Online]. Disponível em: https://www.cnbc.com/2025/10/29/nvidia-on-track-to-hit-historic-5-trillion-valuation-amid-ai-rally.html ↩

Summary

Between 2024 and 2026, the largest technology companies in the United States began investing in AI data centers at a pace that the Bank of England described as "unprecedented historically" [1]. Capex at Microsoft, Alphabet, Amazon and Meta went from $228 billion in 2024 to $376 billion in 2025, and the forecasts for 2026 add up to as much as $745 billion. This article examines how this investment is being paid for and who is left with the risk if the revenue is slow to arrive.

The thesis is that the AI cycle brings together the classic signs of a credit bubble, even with a real technology behind it. Suppliers invest in the customers that buy from them: the IMF counted $1.4 trillion in future revenue commitments across 14 deals of this kind since 2023. Part of the debt stays off the balance sheets. Meta built a $27 billion campus financed by $27.3 billion in bonds issued by the majority partner in a joint venture that Meta does not consolidate, and the largest companies in the sector have more than $1 trillion in leases that have been signed but have not yet commenced. The collateral ages fast: NVIDIA launches new products every year, while the companies depreciate servers over five or six years. And the risk is concentrated in a few names, such as OpenAI, Oracle and CoreWeave, and spreads to funds, insurers and individual investors.

The article uses history as evidence of mechanisms, not as a forecast: the British railway mania and the American panic of 1873, the fiber optics of 2000, Enron and the 2008 crisis. It also shows what could disprove the thesis, such as cloud demand growing outside the AI labs, balance sheets that are still strong and rental prices for older GPUs rising in 2026, and it ends with scenarios and a list of indicators to watch. One section deals with Brazil.

The size of the bet

In 2024, Microsoft, Alphabet, Amazon and Meta together spent about $228 billion on capex, the purchases of buildings, machinery and equipment recorded in the cash flow statement. In 2025, the figure was $376 billion, 65% more. In the first half of 2026 alone, it was $295 billion, almost 80% of the entire previous year. The calculations are mine, based on the figures the companies filed with the SEC, with Microsoft converted to the calendar year from its quarterly reports [2], [3], [4], [5], [6], [7], [8], [9]. Adding Oracle, whose capex more than doubled in the fiscal year that ended in May 2026, the annual pace of the five tops $430 billion [10].

The table shows each company's capex, in billions of dollars, and the most recent forecast for 2026. Microsoft and Oracle use fiscal years that do not match the calendar year, and the forecasts use different definitions: Microsoft's includes finance leases, and Meta's, the principal paid on those leases [11], [12], [13], [14], [15], [16].

Added together, the 2026 forecasts of Microsoft, Alphabet, Amazon and Meta come to between $720 billion and $745 billion, almost double 2025. The rating agency Moody's, in a July 2026 note cited by CNBC, projected capex of $785 billion in 2026 for six companies (the four, plus Oracle and CoreWeave) and of about $1 trillion in 2027 [17]. Oracle alone spent $28.5 billion in the quarter from June to August 2026, more than in the whole of fiscal year 2025 [18].

What the money buys

Most of the spending goes to short-lived equipment. Alphabet reported that, in 2025, it put "approximately 60% of that investment in servers and 40% in data centers and networking equipment" [19]. Microsoft said, about the quarter from April to June 2026: "Roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs" [13]. It is this equipment that loses value at the pace described in the section "The physics of depreciation".

Part of the increase is price, not quantity. In April 2026, Microsoft forecast about $190 billion of capex for the year, including "approximately $25 billion from the impact of higher component pricing" [20]. Meta raised its forecast in the same month and explained that the revision "reflects our expectations for higher component pricing this year" [21]. Amazon raised its own to $220 billion in July, and its chief executive, Andy Jassy, attributed the increase to the price of memory [15]. The IEA records a shortage of high-bandwidth memory, the HBM used in AI chips, which is expected to "persist through at least the end of 2027" [22]. In an investment cycle, whoever controls the scarce component tends to capture the margin, and in 2026 the scarce component is memory.

The companies defend the bill by pointing to demand. Microsoft said it was "confident in the return on these investments given higher demand signals and increasing product usage" [20]. Jassy went further: "But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027 too," and "In fact, the demand we already have for 2028 is striking." The AWS backlog, the contracts signed but not yet delivered, reached $496 billion [15]. These numbers are the best argument against the thesis of this article, and they come back in the section "What could disprove the thesis". But on the other side of the largest cloud contracts signed in 2025 and 2026 are AI labs that are still burning cash and depend on new funding rounds, as the section "The circle" shows.

The size relative to the economy

By my calculation, the capex of the four largest companies in 2025 was equivalent to about 1.2% of U.S. nominal GDP; the forecasts for 2026 take it to somewhere between 2.2% and 2.3% [23]. Not all of this money is spent in the country, because there are data centers abroad, and much of the equipment is imported, as the section on macroeconomics shows. Even so, the order of magnitude is striking. According to the IEA, "Capital expenditure of just five technology companies is now larger than global investment in oil and natural gas production." [22]. As a share of the economy, the cycle is still far from the British railway mania, when investment in railways reached 7.3% of GDP in 1847 [24]. What sets the current cycle apart is concentration: four or five companies account for most of the spending, and the largest technology companies, among them the four that invest the most, make up more than a third of the S&P 500, as the section on macroeconomics shows.

Cash no longer covers everything

Until recently, the big tech companies paid for their investment with the cash they generated, and had money left over. The cushion has shrunk. The table compares the cash generated by operations with capex, in billions of dollars, by my calculation from the figures filed with the SEC:

The companies themselves acknowledge the shift. Amazon reported in July 2026: "Free cash flow decreased to an outflow of $7.6 billion for the trailing twelve months, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment, net of proceeds from sales and incentives. This increase primarily reflects investments in artificial intelligence." [25]. Alphabet, in the same month: "We had negative free cash flow of $5.9 billion in the second quarter, driven by our investments in CapEx." [14].

When capex exceeds the cash generated, the difference comes from debt, from issuing shares or from long-term lease contracts. Moody's summed up the change, according to CNBC: "Previously, these companies relied on asset-light structures centered on software, intellectual property, and scalable cloud services that required modest capital investment," and "The transition from asset-light to asset-heavy models requires unprecedented levels of investment and capital raising." The agency estimated the direct debt of the six companies at about $460 billion and their lease commitments at $1.2 trillion, of which more than $820 billion is in contracts that have not yet commenced, and said it treats these commitments as equivalent to debt [17]. The section "Off the balance sheet" shows where these numbers come from.

None of this, on its own, proves a bubble. Profitable companies investing heavily in a new technology is exactly what one would expect from a technological revolution. The literature on bubbles says to look at something else: how the investment is paid for and who is left with the risk if the revenue does not come. That is the subject of the next section.

How to recognize a bubble

"Bubble" has become a fighting word. Those who think AI is overpriced use the term as an insult; those who disagree reply that the technology is real. Both things can be true at the same time. Charles Kindleberger and Robert Aliber, authors of the standard reference book on financial crises, separate the terms more carefully: "The term mania describes the frenzied pattern of purchases, often an increase in prices accompanied by an increase in trading volumes; individuals are eager to buy before the prices increase further. The term bubble suggests that when the prices stop increasing, they are likely—indeed almost certain—to decline." [26]. The same book contains a caveat worth keeping in view throughout the entire article: "Virtually every mania is associated with a robust economic expansion, but only a few economic expansions are associated with a mania." [26].

The script they describe begins with a real shock. Kindleberger and Aliber summarize Hyman Minsky's model this way: "Minsky argued that the events that lead to a crisis start with a ‘displacement’, some exogenous, outside shock to the macroeconomic system. If the shock was sufficiently large and pervasive, the economic outlook and the anticipated profit opportunities would improve in at least one important sector of the economy." [27]. Then comes the euphoria, in which "investment spending surges because credit is plentiful" and money seems to be free ("money always seems free in manias"). Next comes a pause in the rise, forced selling by those who bought with borrowed money and, sometimes, panic [26]. The book's central thesis is not psychological. It is a thesis about credit: "the cycle of manias and panics results from the pro-cyclical changes in the supply of credit" [26].

Minsky gave this story a vocabulary that is useful for reading balance sheets. He divides borrowers into three types. Hedge units "can fulfill all of their contractual payment obligations by their cash flows". Speculative units pay the interest out of revenue but need to roll over the principal: "Such units need to "roll over" their liabilities". In Ponzi units, operating cash flow does not cover even the interest, and they "can sell assets or borrow" to stay on their feet [28]. The second theorem of the financial instability hypothesis is the sentence that matters most here: "over periods of prolonged prosperity, the economy transits from financial relations that make for a stable system to financial relations that make for an unstable system" [28].

Applied to the AI cycle, the taxonomy separates companies that the debate tends to lump together. The large platforms that pay for their investment with the cash from their own operations are hedge units by Minsky's criterion, even when they issue debt. The cloud companies that specialize in GPUs, which buy chips with loans to be repaid out of future contracts and refinancing, are closer to the speculative profile. And a company whose purchase commitments add up to several times its own revenue, and which depends on new rounds of capital to honor them, enters the territory Minsky called Ponzi. The term is technical and accuses no one of fraud. It describes a structure that only stays standing as long as the financing keeps coming. The sections "The circle" and "The weakest links" show where each company stands.

Minsky also identified who pushes the system from one regime to the other: "bankers (using the term generically for all intermediaries in finance), whether they be brokers or dealers, are merchants of debt who strive to innovate in the assets they acquire and the liabilities they market." [28]. In 2007, the innovation was in securitization vehicles. In the AI cycle, it is in the special purpose vehicles that build data centers, in GPU-backed loans and in the private credit funds that buy this debt. The form changes; the role is the same.

The economist Carlota Perez adds the missing piece for understanding technology bubbles. In her reading, each technological revolution has an installation period, of 20 to 30 years, led by financial capital and brought to a close by a major bubble. After a turning point comes the deployment period, led by production capital [29]. The bubble is no accident: "A major technology bubble (MTB) is not an accidental event. It regularly occurs midway along the process of assimilation of each technological revolution." [30].

The most uncomfortable point in Perez's theory is that the bubble serves a purpose. "It is the switch to short-term gains during the bubble that attracts the necessary capital to be poured into the infrastructural networks of each revolution." [30]. The railways and fiber optics were overbuilt with money that expected a quick return, and they ended up being used by others, later. She describes the spirit of the peak in two sentences: "Soon there is more capital wanting a piece of the action than projects looking for funds." and "The illusion is that there are high profits to be had with very low risk." [30]. And she gives two criteria for recognizing a major technology bubble: "The two defining characteristics of these major technology bubbles are: (a) their concentration on the new technologies –especially the new infrastructural networks– and (b) their decoupling from the real economy." [30].

Perez distinguishes this type of bubble from the credit bubbles described by Minsky: "They are the result of opportunity pull rather than of easy credit push. But they are indeed bubbles." [30]. This distinction is the backbone of this article. The AI cycle began as an opportunity bubble, financed mainly with cash and equity. From 2025 on, it took on the marks of a credit bubble. The IMF recorded that the hyperscalers had raised more than $100 billion in bonds since January 2025, "supplemented by leveraged loans and intercorporate arrangements that are, at times, part of circular financing structures" [31]. It is this mix that makes 2008 relevant.

Robert Shiller, who has studied bubbles all his life, prefers a behavioral definition: "A situation in which news of price increases spurs investor enthusiasm which spreads by psychological contagion from person to person, in the process amplifying stories that might justify the price increase and bringing in a larger and larger class of investors, who, despite doubts about the real value of the investment, are drawn to it partly through envy of others’ successes and partly through a gambler’s excitement." [32]. The indicator that bears his name, the CAPE, divides the price of the S&P index by the ten-year average of real earnings. In Shiller's series, updated on October 5, 2026, the CAPE reached 44.20 in December 1999, the record, and 27.55 in May 2007. In August 2026 it stood at 40.99, the highest reading since 2000, and the preliminary reading for October 2026 is 40.71. The average since 1881 is 17.80 [33].

Recognizing a bubble early does not mean correctly calling when it will burst. On December 5, 1996, the chairman of the Federal Reserve, Alan Greenspan, asked in public: "But how do we know when irrational exuberance has unduly escalated asset values, which then become subject to unexpected and prolonged contractions as they have in Japan over the past decade?" [34]. The CAPE stood at 27.72 that month and kept rising for three more years, to 44.20 in December 1999 [33].

Shiller himself warns that the indicator does not pinpoint a date and that bubbles do not always burst all at once: "Though the abrupt ends of stock market booms in 1929, 2000 and 2007 might seem consonant with such a metaphor, these booms were reflated again before long" [32]. Andrew Odlyzko, who has studied the railway manias, documents a case that contradicts the rule: the British railway mania of the 1830s, which was speculative and profitable for those who invested. His conclusion serves as a brake on this text: "The example of the railway mania of the 1830s serves as a useful antidote to claims that bubbles are easy to detect or that all large and quick jumps in asset valuations are irrational." [35].

From here on, the article uses six markers, all taken from this literature and all verifiable with public data:

No marker, on its own, proves a bubble. The thesis of this article is that, in 2026, all six appear at the same time, and that three of them (credit, innovation in liabilities and circular revenue) are precisely the ones that separated a market correction from a financial crisis in the episodes that follow.

Railways: the infrastructure stays, the investor pays

The comparison most often cited for AI is the internet of 2000, but the most useful one may be older. In the Britain of the 1840s, railways were the technology that was going to change everything, and they did. Each line needed authorization from Parliament, which left an annual record of the euphoria. There were 91 miles approved in 1843, 805 in 1844, 2,700 in 1845 and a record 4,538 in 1846. In 1847, 1,354; in 1850, just 8 [24].

The size of the investment is the figure that speaks most to 2026. According to Odlyzko, railway spending "peaked in 1847, at a level of about £44 million (in contrast to about £50 million for the national government budget)". By his calculation, "railway investments of £44 million in 1847 came to 7.3% of a GDP of £600 million, which is comparable to $1.1 trillion for the U.S. today", and the "today" of the text is 2010. In the table in the same work, railways account for £43.9 million of the £88 million invested in the country in 1847, almost half of all British investment [24]. The author's strongest comparison is with the military budget: "At its height, individual capitalists, in pursuit of private profit, were plowing more than twice as much into the construction of a public infrastructure as their nation was spending on the military." [24].

Investment did not stop when the shares fell. Odlyzko explains why: "Even in 1850, when investors felt crushed by the stock market decline, investments were high, higher than they had ever been before 1845. The reason was not that any shareholders were enthusiastic about the spending. Rather, so much money had been spent that it was felt better to go on and complete the lines and earn small returns than abandon them completely and get nothing." [24]. This is the logic of sunk costs, and it matters for AI. A half-built data center campus is worth nothing, which is why capex tends to stay high for some time after confidence is gone.

For those who put money in, the outcome was bad. "The best known index of railway shares prices declined from 167.9 in the summer of 1845 to 60.5 at the end of 1849, a drop of 64%." [36]. The Economist summed up the paradox in 1855, in a sentence that Odlyzko reproduces: "[m]echanically or scientifically, the railways, with all their multiplied conveniences and contrivances, are an honour to our age and country: commercially, they are great failures." [24]. Charlotte Brontë, who owned shares in one of the lines, wrote in 1849, also according to Odlyzko: "The original price of Shares in this Railway was £50. At one time they rose to 120; and for some years gave a dividend of 10 per cent; they are now down at 20, and it is doubtful whether any dividend will be declared this half-year." [24].

The popular explanation puts the blame on fraudsters, such as George Hudson, who doctored balance sheets. Odlyzko disagrees: "accounting fraud was not the fundamental cause of the investment disaster of the Railway Mania. Distorted financials helped mask for a while the disappointing returns that new railways produced for investors, but were not the reason for the lack of profits. The main problem was the headlong expansion during the Mania. It tripled British railway mileage, and the new lines cost more than expected and brought revenues lower than expected." [36]. Annual revenue per mile fell from £3,470 in 1845 to £2,080 in 1850 [36]. The lesson holds for today: the problem with an infrastructure bubble is usually arithmetic, not moral. Too much capacity, built at too high a cost, for a demand that grows, but more slowly than promised.

Kindleberger describes how the profile of investors changes over the cycle: "In early railway investment, for example, initial investors were persons doing business along the rights of way who sought benefits from the railroad for their other concerns. They were followed by a second group of investors interested in the profits the railroad would earn, and by a third group, made up of speculators who, seeing the rise in the railroad’s shares, borrowed money or paid for the initial instalments with no intention of completing the purchase, to make a profit on resale." [37]. In AI, the first group is the companies that build capacity for their own products. The second is those that build to sell or rent capacity to third parties. The third appears when money starts coming in because the price is rising, and not because the cash flow justifies it.

Perez notes a detail of the railway mania that recalls Minsky's Ponzi: "During the railway mania in the 1840s, though, companies without profits did pay dividends out of the new capital they attracted in order to keep previous investors in and new investors coming." [30]. When new money pays those who came earlier, the structure depends on new money continuing to come in.

Odlyzko explicitly links the Victorian mania to that of 2000. For him, the two biggest disasters of real investment in the internet bubble were long-distance fiber in the U.S. and the European 3G spectrum auctions, "each of which cost on the order of $100 billion to €100 billion" [24]. In both cases, "there was hard quantitative evidence that once those went into service, there would not be enough demand to provide the expected profits." [24]. And the sentence he uses for the two episodes fits the third: "society decided to adopt the new offerings at its own speed, slower than that expected by inventors, promoters, and investors, and financial ruin followed in both cases." [24].

There is a counterpoint, and it comes from the same author. The railway mania of the 1830s, smaller and earlier, paid off for those who invested. Odlyzko also distinguishes infrastructure financed with equity from debt crises: "One could argue that the infrastructure construction funded by equity is radically different from the much more frequent, far more dramatic and attention-grabbing financial crises that are associated with large accumulations of debt, typically for consumption." [35]. This distinction decides the size of the damage. When the bill falls on shareholders, the loss is painful, but it does not spread as it does in a banking crisis. When it falls on leveraged creditors, it spreads.

1873: when railroads are paid for with debt

On the other side of the Atlantic, the story had a different ending, because the money was different. The American crisis of 1873, in the words of Federal Reserve historians, "arose from investments in railroads. Railroads had expanded rapidly in the nineteenth century, and investors in many early projects had earned high returns. As the Gilded Age progressed, investment in railroads continued, but new projects outpaced demand for new capacity, and returns on railroad investments declined." [38]. When collapses on the Vienna stock exchange led European investors to sell American railroad bonds, the companies ran out of money to operate and to refinance their debts. The bank Jay Cooke & Co., heavily exposed to the Northern Pacific, failed on September 18, 1873. Two days later, the New York Stock Exchange closed for the first time in its history and went ten days without trading, and at least a hundred banks failed across the country [38].

The difference from the British mania lay in the liabilities. "Railroads were built mainly with borrowed money", write Kay Giesecke, Francis Longstaff, Stephen Schaefer and Ilya Strebulaev in a study of 150 years of corporate bond defaults. Between 1873 and 1875, defaults amounted to "36 percent of the par value of the entire corporate bond market"; the worst three-year period of the Great Depression came to 12.88% [39]. The contraction that followed lasted 65 months, from October 1873 to March 1879, the longest in the NBER chronology [40]. The authors themselves draw the connection to recent cycles: the debt-financed technology booms of the 19th century were followed by crises "closely resembling patterns in the telecommunications and structured-credit-products cycles of the past decade." [39].

Here, too, there is a counterpoint. The economic historian Scott Mixon shows that, before 1873, stocks, options and bonds were already pricing in the risk. Investors "were not caught up in an irrational, speculative mania", and the crisis began with the most leveraged railroads [41]. There was no lack of warning. The problem was debt concentrated in the weakest links, which is precisely the question of the section "The weakest links".

The comparison of durations sums up the argument of this article. According to the NBER, the recession that followed the bursting of the internet bubble, which had been financed mainly with equity, lasted 8 months, from March to November 2001. The recession of 2007–2009, a credit crisis, lasted 18 months. That of 1873, a railroad bond crisis, lasted 65 [40]. The size of the damage depends less on the technology than on who is holding the debt.

The railways leave two lessons for AI. The infrastructure outlives the bubble: the lines of the mania were completed and remained in use after the shares sank, and the buildings, substations and transmission lines of today's data centers will also serve someone. But a railway serves for decades, and the chip that justifies the investment lasts only a few years, as the section on depreciation shows. And until 2024 the AI cycle was paid for mainly with cash and equity, like the British manias; since 2025 it has been shifting to debt, like the American railroad expansion that ended in the panic of 1873.

Telecom, 1996–2002: dark fiber and round-trip revenue

If the railways show the arithmetic of an infrastructure bubble, telecom in the late 1990s shows the accounting. It was the last cycle in which a network technology attracted capital on this scale, and almost all the mechanisms discussed in this article appeared there: demand measured by capacity, suppliers financing customers, cross-sales between competitors and a technical improvement that multiplied supply faster than the market could use it.

The investment was large and fast. Economists at the Federal Reserve Bank of Richmond calculated that, between the first quarter of 1996 and the fourth quarter of 2000, "investment in communications equipment grew from approximately $62 billion per year to over $135 billion per year in constant 1996 dollars", an average growth of almost 18% a year. The fall was equally fast: in the fourth quarter of 2001, the figure was below $93 billion, "only 69 percent of the same figure one year earlier", and the share of communications equipment in private investment fell from almost 7% in 2000 to 4.8% at the end of 2002 [42].

The myth that justified the investment

The story that sustained this capex was a statistic. Andrew Odlyzko describes it this way: "The most popular and extremely misleading myths of the dot-com and telecom bubbles was that “Internet traffic doubles every 100 days” (or 3 months, or 4 months)." [43]. The real figure, according to him, was much smaller and still enormous: traffic was "close to doubling each year since 1997". The main source of the myth had a name: "At its base, most of the support for the myth of astronomical growth rates in Internet traffic was coming from WorldCom." [43]. In a presentation in March 2000, Bernie Ebbers, WorldCom's chief executive, said that, because of demand, the company's capex in 2003 "would have to exceed $100 billion" [43].

The detail that matters most for 2026 is the confusion behind the myth: "most of the WorldCom claims about astronomical growth rates of their network were about network capacity, not traffic." [43]. Installed capacity is a decision made by investors. Traffic is a decision made by users. When the former starts being presented as if it were the latter, the market loses its thermometer. The section "The circle" shows that many of the AI demand figures, such as contracted gigawatts and contract backlogs, measure capacity commitments, not usage.

Too much fiber, and each strand carrying more

The FCC's last annual survey of fiber, with data through the end of 1998, already showed the picture. The older carriers used much of the fiber they had: 50% at AT&T, 85% at Sprint. The new ones, which were building entire networks from scratch, barely lit theirs. Qwest had 2% of its fiber lit in 1998, Williams 1%, GST 2% and NEON 5%. In that year alone, the fiber deployed by long-distance carriers grew by more than 30%, to more than 3.6 million fiber miles [44]. The FCC report explained the logic of the excess in one sentence: "Although up-front costs for fiber deployment are high, a significant portion of the total investment can be deferred until actual demand materializes." [44]. The Richmond economists add that the big four of 1996 had 72% of U.S. long-distance fiber and, in 1999, only 30%, while annual fiber deployment multiplied more than fourfold [42].

Mark Doms, of the Federal Reserve Bank of San Francisco, measured the race another way: "in 1996 there were 143,000 route miles of fiber in the country, but by 2000 this had increased to 320,000." In 1990, four companies had national fiber networks; in 2000, ten. "Chicago and Minneapolis are each served by twenty long-haul firms", compared with about five in 1990 [45].

And each strand came to carry much more. "In 1996, a strand of fiber transmitted data at approximately 2.5 gigabits per second (Gbps). By 2000, the capacity of the same fiber could reach 100 Gbps.", thanks to wavelength-division multiplexing; the capacity of a single fiber doubled "approximately every nine months between 1996 and 2000", faster than Moore's law [42]. The authors list three causes of the excess: the overenthusiasm typical of a new industry, a capacity per strand that grew more than expected and, "perhaps most importantly", a demand that "had not grown as fast as many had forecast: the pessimists turned out to be right." [42].

This is the most direct technical parallel with AI. When efficiency rises faster than demand, the physical capacity built for the forecast demand is left over, even if demand keeps growing. In AI, the efficiency comes from new chips every year and from models that deliver the same result with less computation. The section on depreciation deals with this in numbers.

What the bill came to

The Nasdaq telecom stock index went from 198 in April 1997 to 1,230.06 on March 10, 2000, and fell 62% by the end of that year. On May 16, 2003, it stood at 136. The telecom companies in the S&P alone lost about $700 billion in market value between 2000 and 2002, "more than 3.5 percent of the entire value of U.S. corporate equities at the stock market peak in 2000" [42]. The Nasdaq Composite index closed at 5,048.62 on that same March 10, 2000, and at 1,114.11 on October 9, 2002, a drop of 77.9%. It did not close above its 2000 peak again until April 23, 2015, more than fifteen years later [46].

The Bank for International Settlements (BIS) explained in 2002 why the banks came out relatively unscathed: "a considerable amount of the financing for the most risky elements of the late 1990s boom was obtained from outside the banking system. This is perhaps best illustrated by the fact that firms in the technology and telecommunications sectors relied heavily on vendor financing, venture capital and the equity and bond markets for their financial needs." [47]. It is the same financing mix that appears today, under other names: vendor credit, venture capital, stocks and bonds.

The supplier that lends to customers so they can buy from it

The 2001 financial statements of the equipment makers record the mechanism that would return in 2025. Lucent explained in its annual report: "Some of our customers worldwide are requiring their suppliers to arrange or provide long-term financing for them as a condition of obtaining or bidding on infrastructure projects. These projects may require financing in amounts ranging from modest sums to more than a billion dollars." [48]. Lucent's total customer financing commitments were $8.1 billion on September 30, 2000. In fiscal year 2001, the company recorded $2.2 billion in provisions for doubtful accounts and customer financing, about $1.3 billion of it on three projects, among them Winstar, which filed for bankruptcy protection in April 2001 [48]. Among the risk factors, a sentence that would apply to any supplier that invests in its own customers today: "We are vulnerable to downturns in our customers' businesses as a result of our customer financing arrangements." [48].

Nortel told the same story. "As a result of the competitive environment during 1999 and 2000, we provided significant amounts of medium-term and long-term customer financing" [49]. Total customer financing fell from $5.168 billion at the end of 2000 to $2.075 billion at the end of 2001, because "our ability to place customer financing with third-party lenders was significantly reduced during 2001". Nortel's net loss in 2001 was $27.3 billion [49]. Cisco treated credit as a sales weapon: "We believe customer financing is a competitive factor in obtaining business, particularly in supplying customers involved in significant infrastructure projects." Its loans could include "additional funds for soft costs associated with network installation and integration of our products and for working capital purposes" [50].

In October 2026, BIS economists used exactly this episode to explain the risk of circular relationships in AI: "upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors' equipment [...]. This meant that part of the equipment vendors' reported sales was being funded by the vendors themselves." [51].

Round-trip revenue

In some cases, the circularity went beyond financing and turned into manufactured revenue. In 2004, the SEC charged Qwest with fraudulently recognizing more than $3.8 billion in revenue between 1999 and 2002. Part of it came from the sale of rights to use fiber (IRUs) and, when demand fell, from swaps: "When the demand for IRUs declined, Qwest engaged in IRU "swaps" whereby Qwest bought IRUs from other companies in exchange for agreements from those companies to buy IRUs from Qwest." [52]. According to the SEC, the practice became so routine that employees compared it to an addiction and called the non-recurring sales "heroin". Qwest settled without admitting or denying the charges and paid a $250 million penalty [52].

Global Crossing is the most instructive case, because the SEC measured the circularity. In the first half of 2001, the company's capacity sales came to depend on simultaneous purchases from the same carriers: "GX’s ability to sell capacity often became linked to, and in some cases dependent on, GX’s simultaneous purchase of a nearly equal dollar amount of capacity." In the second quarter of 2001, these "reciprocal transactions" made up $516 million of the $567 million in IRU sales, or 91%, and without them the "Recurring Adjusted EBITDA" of the company "would have gone from the reported $472 million to a negative $43 million." [53]. Global Crossing filed for bankruptcy protection on January 28, 2002. It was a disclosure case, closed with a settlement, and the SEC did not charge fraud [53].

WorldCom, the source of the traffic myth, ended up as the simplest case. In June 2002, the SEC charged it with an accounting fraud of "more than $3.8 billion", carried out by booking as investment, rather than as expense, the amounts paid to other carriers for the use of their networks [54]. The company filed for bankruptcy protection on July 21, 2002 [55].

The counterpoint

Not everyone sees a bubble in telecom. The Richmond economists, writing in 2003, called the cycle "in large part—a rational response to the changing fundamentals of technology and regulatory environment", with large forecasting errors in an environment of uncertainty: "our explanation for the telecommunications boom and bust does not involve fraud, irrationality, or a bubble. To be sure, as the bust became apparent, fraud did occur. But it is not clear that fraud played an important role in the boom and the early stages of the bust." [42]. The argument is serious, and it applies to AI: it is possible to get the demand forecast wrong without any fraud at all. Fraud appears later, when the numbers have to add up.

Doms adds a technical argument worth keeping in mind for data centers: "the marginal cost of fiber is small relative to the cost of laying a new fiber route; in other words, it would be much cheaper to lay additional strands of fiber today while trenches are open than to dig up the trenches again several years from now and lay some additional fiber lines. So, although the utilization rate is low, that does not imply that firms were irrational in the amount of fiber they installed." Even so, he acknowledges that technology investment in 1999 and 2000 was far above what a model predicted, because "expectations were too high, especially in two sectors of the economy, telecommunications services and the dot-com sector." [45].

Telecom leaves four signs to look for in 2026, and none of them requires assuming bad faith. Demand measured by contracted capacity. Sales financed by the seller itself. Cross-contracts between companies that are each other's customers and suppliers. And a technology that multiplies supply faster than usage. The next sections show that all four are present in the AI cycle, and on a larger scale.

Enron: the rule born of a scandal

Enron comes into this article for two reasons that have nothing to do with the fraud itself. The first is the economic motive that led the company to take debt off its balance sheet, because it is the same as that of any company that needs to invest heavily before it gets paid. The second is that the accounting rules that today decide whether or not a data center appears on a big tech company's balance sheet were born from the investigation of this case.

The report of the special committee of Enron's board, known as the Powers Report, describes the problem without adjectives: "Much of this growth involved large initial capital investments that were not expected to generate significant earnings or cash flow in the short term. While Enron believed these investments would be beneficial over a period of time, they placed immediate pressure on Enron's balance sheet. Enron already had a substantial debt load. Funding the new investments by issuing additional debt was unattractive because cash flow in the early years would be insufficient to service that debt and would place pressure on Enron's credit ratings. Maintaining Enron's credit ratings at investment grade was vital to the conduct of its energy trading business." [56].

The solution was to bring outside partners into separate entities, which could borrow in their own name, "although in many cases a guaranty or other form of credit support was required from Enron". And the report explains why management preferred this form: "Enron management preferred the latter treatment--known as "off-balance-sheet"--because it would enable Enron to present itself more attractively as measured by the ratios favored by Wall Street analysts and rating agencies." [56]. Large investment, distant returns and a rating to protect: the motive is identical to the one behind a $27 billion data center campus, and the section "Off the balance sheet" shows how it is dealt with today within the rules.

The 3% rule

Before 2003, the line between on and off the balance sheet was almost arithmetic. The Powers Report sums it up: a company could treat a special purpose entity (SPE) as independent if "(1) an owner independent of the company must make a substantive equity investment of at least 3% of the SPE's assets, and that 3% must remain at risk throughout the transaction; and (2) the independent owner must exercise control of the SPE." [56]. The SEC staff later described the practical effect: by giving voting control of the entity to a third party holding 3% of the equity, "a sponsor could generally avoid consolidation, despite the fact that the activities of the SPE could not be substantively changed by the "controlling" investor." [57].

Some of Enron's structures did not even meet the rule; others met it on paper but not in substance. The best known, the "Raptor" vehicles, were hedges against a decline in Enron's own investments, capitalized with Enron stock. The Powers Report points out the flaw: "If the value of the investments fell at the same time as the value of Enron stock fell, the SPEs would be unable to meet their obligations and the "hedges" would fail. This is precisely what happened in late 2000 and early 2001." [56]. In the second half of 2000, the "profits" from the Raptors accounted for more than 80% of the pre-tax earnings that Enron reported [56]. A guarantee that loses value exactly when it is needed has a technical name, wrong-way risk, and the term comes back in the section on 2008 and in the section on the AI circle.

Revenue that goes full circle

The Enron mechanism most similar to the subject of this article was not the Raptors but the "prepays". These were three-party transactions, involving an Enron affiliate, a bank and a vehicle created by the bank, in which the payments went full circle. The SEC staff describes the result: "The circular nature of delivery and payments with respect to the commodities had the effect of eliminating any material risk or any potential gain with respect to the changes in the price of the underlying commodity. Each party's apparent assumption of price risk was illusory. With the elimination of price risk, "prepays" were effectively debt." [57]. Enron booked the money as operating cash flow and the obligation as a risk management liability, not as debt: "these prepay transactions allowed Enron to hide debt, lower key financial ratios followed by analysts, and provide the illusion of cash flow from operations." [57].

The same passage provides the analytical tool that this article uses to read the AI deals, the so-called unit of account problem: "The economics of a transaction may look quite different depending on how broadly or narrowly one defines the boundaries of the transaction. That is, a particular contract may appear to have certain economic characteristics when viewed in isolation—and may be given a certain accounting treatment that corresponds to those economic characteristics—but if understood as a piece of a larger agreed-upon transaction may actually have quite different economics, and be properly accorded different accounting treatment." [57]. In plain terms: to understand a circular deal, one has to look at the whole loop, not at each leg in isolation. This holds as a method, not as an accusation. The AI deals discussed below are public, and there is no allegation of fraud against any of them.

How much was hidden

The size of what was kept out of the numbers is the most alarming figure in the case. Citing the examiner appointed by the bankruptcy court, the SEC staff noted: "For 2000, barring the use of these techniques, Enron's reported debt would have been $22.1 billion rather than $10.2 billion." What is more, on November 19, 2001, Enron reported about $13 billion of debt on its balance sheet, but on the same day told the banks that its debt was about $38 billion [57]. The Powers Report points out that much of this appeared in the notes to the financial statements, but in a useless form: "these disclosures were obtuse, did not communicate the essence of the transactions completely or clearly, and failed to convey the substance of what was going on between Enron and the partnerships." [56]. Available information and understandable information are different things.

The response: variable interest entities

The response from the body that sets U.S. accounting standards, the FASB, came in January 2003 with Interpretation No. 46, revised in December of the same year [57]. The central idea is that voting control is not enough to decide who consolidates an entity: "The voting interest approach is not effective in identifying controlling financial interests in entities that are not controllable through voting interests or in which the equity investors do not bear the residual economic risks." [58]. Entities without enough equity to finance themselves without support from third parties came to be called variable interest entities (VIEs), and the party that consolidates them is the "primary beneficiary": "the party that absorbs a majority of the entity's expected losses, receives a majority of its expected residual returns, or both" [58].

The rules did not put an end to the structures. The SEC staff observed, as early as 2005, that "many arrangements with potential VIEs were restructured such that the entity either would not be considered a VIE or such that no party would be required to consolidate the VIE." [57]. And one exception survived the reform, that of qualifying securitization vehicles (QSPEs): "Transferors to qualifying special-purpose entities [...] do not consolidate those entities." [59]. It was through this door that much of the leverage of 2007 passed.

The same SEC report measured another gray area, that of leases. With "all or nothing" rules, contracts were designed to fall just below the lines that required recognition on the balance sheet, and the staff estimated that U.S. companies had about $1.25 trillion in future operating lease obligations that did not appear on the balance sheet, only in the notes [57]. The diagnosis became famous: "lease structuring to meet various accounting, tax, and other goals, has become an industry unto itself in the last 30 years." [57].

Enron leaves three lessons for reading the AI data centers, all about structure and none about fraud. The motive for taking investment off the balance sheet is still to protect ratios and ratings. The rules that decide what goes on the balance sheet depend on judgments (who has the power, who absorbs the losses, how long the contract lasts) and can be designed around. And guarantees whose value falls together with the risk they were supposed to cover fail precisely when they are called upon.

2008: what leaves the balance sheet comes back in a crisis

The U.S. congressional commission that investigated the 2008 crisis, the Financial Crisis Inquiry Commission (FCIC), opened its conclusions with a sentence that should trouble anyone who says crises are unpredictable: "We conclude this financial crisis was avoidable. The crisis was the result of human action and inaction, not of Mother Nature or computer models gone haywire." [60]. This article uses 2008 as evidence in the strict sense. Not to say that AI will produce a crisis of the same size, but to show which mechanisms turned losses in one sector into losses across the system, and to look for those mechanisms in the current cycle.

The first mechanism was leverage that did not show. According to the FCIC, the five major investment banks operated with "extraordinarily thin capital. By one measure, their leverage ratios were as high as 40 to 1, meaning for every $40 in assets, there was only $1 in capital to cover losses. Less than a 3% drop in asset values could wipe out a firm." [60]. And part of it was out of sight: "the leverage was often hidden—in derivatives positions, in off-balance-sheet entities, and through "window dressing" of financial reports available to the investing public." [60]. The commission summed up the system in one sentence: "We had a 21st-century financial system with 19th-century safeguards." [60].

From Enron to commercial paper

The FCIC ties the post-Enron reform directly to the leverage of 2007. Banks put loans into asset-backed commercial paper (ABCP) programs, off the balance sheet, and so stopped setting aside capital for them. When the FASB, after Enron, made that way out harder, the market stalled, and bank regulators responded: in 2004 they allowed these assets to be left out of the capital calculation, with a charge of "up to 0.8%" on the liquidity lines provided to the programs. "Growth in this market resumed." [60]. The lesson is uncomfortable: the accounting rule tightened, the prudential rule loosened, and the structure kept growing.

The ABCP market reached $1,225.7 billion in the week ending August 8, 2007, according to Federal Reserve data. By the end of that year it stood at $774.5 billion [61]. Economists at the Fed itself described what happened as a bank run without banks: "more than 100 programs (one-third of all ABCP programs) were in a run within weeks of the onset of the turmoil and the odds of subsequently leaving the run state were very low." [62]. Structured investment vehicles (SIVs), a variant of these programs that invested mainly in AAA-rated securities, numbered 36, with almost $400 billion in assets, leveraged "an average of just under 14-to-1". Investors in one of them, Sigma, lost more than 95%. "As of fall 2010, not a single SIV remained in its original form." [60].

What was outside comes back in

The central point of 2008 for this article is what happens to commitments that are not on the balance sheet when the crisis arrives. Citigroup had sold "liquidity puts", promises to buy commercial paper issued by CDOs if no one else would, covering $25 billion. Before 2004, these promises required no capital at all; afterward, 0.16%. In 2007, the FCIC recounts, events would "catapult the entire $25 billion in commercial paper straight onto the bank's balance sheet, requiring it to come up with $25 billion in cash as well as more capital to satisfy bank regulators." [60].

The same happened with the SIVs the bank managed, through pressure rather than contract. On December 13, 2007, Citi announced a support facility for the vehicles and stated: "As a result of this commitment, Citi will consolidate the SIVs' assets and liabilities onto its balance sheet under applicable accounting rules." The SIVs' assets had fallen from $87 billion in August to $49 billion [63]. In November 2008, the bank brought the last $17.4 billion onto its balance sheet [60]. William Dudley, of the New York Fed, admitted to the FCIC that few people understood "the backstops that banks were providing SIV conduits either explicitly or implicitly." [60]. In bank regulation, this later came to be called step-in risk: the sponsor is not obliged to rescue the vehicle, but it does, because the alternative costs more in reputation and business.

Margin calls before the loss

The AIG case shows another mechanism: losses that arrive before the actual loss. The insurer's financial products subsidiary had sold protection against default, through credit default swaps, on $79 billion in CDOs [60]. The contracts required cash collateral if the market value of the securities fell or if AIG's rating was downgraded: "collateral calls could be triggered even if there were no actual cash losses in, for example, the super-senior tranches of CDOs upon which the protection had been written." [60]. As of September 12, 2008, the calls totaled $23.4 billion, and AIG had already paid $18.9 billion. The government ended up committing "more than $180 billion" to keep AIG's collapse from spreading [60]. When collateral is marked to market and the contract has rating triggers, the liquidity crisis arrives before any real loss. In the AI cycle, the closest parallel is loans with GPUs as collateral, whose size depends on the book value of the chips, as the section "The weakest links" shows.

The seal of approval

For the FCIC, the rating agencies were "essential cogs in the wheel of financial destruction". Between 2000 and 2007, Moody's gave its top rating to nearly 45,000 mortgage-related securities, and "83% of the mortgage securities rated triple-A that year ultimately were downgraded" (the year is 2006) [60]. The problem was not only one of method. Many of those securities were highly rated because the structure transferred another party's credit quality to them. The section "Off the balance sheet" shows that the largest bond ever issued for a data center is highly rated for the same reason: the rating belongs to the guarantee, not to the building.

Quarter-end window dressing

Lehman Brothers added a mechanism that completes the list. According to the bankruptcy examiner, Anton Valukas, the bank used transactions known internally as "Repo 105" to temporarily remove about $50 billion in assets from its balance sheet at the end of the first and second quarters of 2008. With them, the net leverage it reported at the end of the second quarter was 12.1, rather than 13.9 [64]. Lehman's own employees called the practice an "accounting gimmick" [64]. The report did not decide whether the transactions technically violated the accounting standard; the point is that a ratio closely watched by the market became a target for engineering.

Wrong-way risk

The concept that links all these cases is wrong-way risk. The Basel Committee defines the specific version as follows: "Specific wrong-way risk arises when the exposure to a particular counterparty is positively correlated with the probability of default of the counterparty due to the nature of the transactions with the counterparty." [65]. In its post-crisis review, the Committee noted that the credit insurers (monolines) were counterparties whose credit quality moved together with the exposure, and that these transactions "contributed to the losses during the crisis". The textbook example it gives is that of a counterparty that offers its own shares as collateral [66], exactly the flaw in Enron's Raptors. In October 2026, the IMF applied the same concept to AI's circular deals, as the next section shows.

The 2009 reform

The FASB responded in June 2009 with two standards, FAS 166 and FAS 167. They eliminated the exception for qualifying securitization vehicles and replaced the quantitative consolidation test with a qualitative one. Since then, a company consolidates a variable interest entity if it has "a. The power to direct the activities of a VIE that most significantly impact the VIE's economic performance b. The obligation to absorb losses of the VIE that could potentially be significant to the VIE or the right to receive benefits from the VIE that could potentially be significant to the VIE." [67]. Existing structures were not spared, and the effect was measured: in the week ending March 31, 2010, U.S. commercial banks had brought onto their balance sheets $377.8 billion in assets and liabilities of vehicles previously off the balance sheet, mostly in credit card securitizations [68]. It is under this rule, the test of "power" and of the "obligation to absorb losses", that today's data center joint ventures stay off the balance sheet.

Shadow banking then, private credit now

In 2008, the leverage was in what came to be called the shadow banking system. Today, the fastest expansion of credit outside the banks is in private credit funds. The IMF estimated the sector at about $2.1 trillion in 2023 and listed its vulnerabilities: "relatively fragile borrowers, increased exposure of pensions and insurers to the asset class, a growing share of semiliquid investment vehicles, multiple layers of leverage, stale valuations, and unclear interconnections between participants." [69]. In May 2026, the Federal Reserve recorded $1.4 trillion in private credit loans in the U.S. and something unprecedented: in the first quarter of 2026, accepted redemption requests at semiliquid credit funds (perpetual BDCs) exceeded inflows, "the first time this has occurred since these vehicles were created" [70]. The same report notes that the leverage of the largest life insurers was "well into the upper quartile of its historical distribution" [70].

Two bridges connect this market to AI. The first is where the money goes: BIS economists calculated that the share of technology companies in private credit's direct lending "doubled to over 40% between 2020 and 2025" [71]. The second is who buys the risk. The Financial Stability Board (FSB) warned in May 2026 that private ratings, sometimes from little-known agencies, are being used to attract insurers, which depend on ratings to meet capital rules: "The growing use of private ratings, sometimes from lesser-known providers, to facilitate investment by rating-reliant investors (including insurers) also warrants monitoring." [72]. And it summed up: "Private credit remains untested to a prolonged economic downturn and so warrants close attention." [72].

Where the analogy fails

There are important differences, and leaving them out would be doing to 2008 what the optimists do to AI. In 2008, the risk sat in banks funded with deposits and very short-term debt, and the asset was the mortgage of millions of families. Today's private credit is funded mostly with long-term capital, and the IMF itself wrote, in 2024, that "Private credit loans are funded largely with long-term capital, mitigating maturity transformation risks. The use of leverage appears modest, as do liquidity and interconnectedness risks." [69]. In April 2026, the IMF still judged that the hyperscalers "thus far have financed investments robustly without materially weakening their financial positions" [31]. The FCIC also had dissenting votes: some of the commissioners rejected the majority's narrative, and the dissent signed by Keith Hennessey, Douglas Holtz-Eakin and Bill Thomas has a section titled "The dangerous imprecision of the term 'shadow banking'" [60].

What 2008 teaches, then, is not that history repeats itself at the same size. It is a set of five mechanisms to look for. Off-balance-sheet commitments come back onto the balance sheet in a crisis, by contract or for reputational reasons. Triggers tied to market prices and to ratings bring losses forward. Ratings can measure another company's guarantee rather than the asset. Guarantees that fall together with the risk fail when they are called. And ratios watched by the market become targets for engineering. The following sections look for each of them in the financing of AI.

The circle: who invests, who sells and who buys

The section on telecom showed the mechanism: Lucent, Nortel and Cisco lent carriers money to buy their equipment, and the manufacturers' revenue grew with demand they themselves were financing. In 2025 and 2026, the mechanism came back bigger and in more forms. Chip and cloud suppliers buy stakes in AI labs, which sign purchase commitments with those same suppliers. There are stock bonuses for buyers, lease guarantees, commitments to buy whatever capacity is left over and financing released as the product is delivered.

The generic design fits in a loop:

chip or cloud supplier
   |  1. invests in the lab or guarantees its debt
   v
AI lab
   |  2. signs a purchase commitment with the supplier
   v
supplier's contract backlog (RPO)
   |  3. future revenue and market value rise
   v
supplier raises more money and invests again
   |
   +--> back to step 1

The deals

The table brings together the main documented deals. The amounts are those announced or filed by the companies, many of them "up to" figures, conditional on deliveries or not yet final:

The sources are in the paragraphs that follow, which detail each case.

NVIDIA and OpenAI

In September 2025, NVIDIA and OpenAI announced a letter of intent for at least 10 GW of NVIDIA systems: "To support the partnership, NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed." [73]. The supplier's money would be released as the customer installed the supplier's products. In its next quarterly report, NVIDIA warned: "There is no assurance that we will enter into definitive agreements with respect to the OpenAI opportunity" [74]. The $100 billion deal did not materialize. In February 2026, NVIDIA put $30 billion into OpenAI's $110 billion round, "at a $730B pre-money valuation", and OpenAI secured "3GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems" [75]. In March, Jensen Huang said the $100 billion opportunity was probably "not in the cards", because OpenAI was going to go public [76].

NVIDIA does not invest only in OpenAI. In July 2026, the company reported "equity investments of $99 billion and equity investment commitments of $25 billion", and the book value of its stakes in private companies had gone from $3.8 billion to $47.9 billion in one year [77]. In its annual report, the company itself describes some of these investments: "These investments include AI model makers that purchase our products directly or through CSPs." [78]. Unrealized gains on these stakes flow into earnings. When the customers' valuations rise, so do the earnings of the company that sells to them.

In the same quarter, NVIDIA created what it called "a new business model": AI clouds buy its infrastructure and NVIDIA commits to renting capacity from them, under contracts that "the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates". These commitments totaled $36 billion [77]. In practice, the supplier sells the equipment and guarantees a demand floor for whoever bought it. The section "Off the balance sheet" shows the next step: the guarantees of up to $105 billion on OpenAI's leases.

AMD and OpenAI

AMD chose a different form. In October 2025, it gave OpenAI the right to buy up to 160 million AMD shares at $0.01 each, vesting in tranches as OpenAI bought AMD GPUs, up to 6 GW, and as AMD's stock reached targets that "escalate to $600 per share for the final tranche" [79]. By my calculation, 160 million shares were about 9.8% of AMD's shares outstanding at the end of October 2025 [80]. AMD's chief financial officer predicted that the partnership would bring "tens of billions of dollars in revenue for AMD" [79]. In the accounting, the stock bonus becomes a discount on revenue: "The grant date fair value of the warrant will be recognized as a reduction to revenue when revenue is recognized." [80]. In February 2026, AMD repeated the model with Meta, at the same size and with the same targets [81]. And it began guaranteeing the lease of a partner's data center, with a maximum exposure of $4.1 billion; guarantees of this kind, the company says, "may be issued in exchange for warrants" [82].

Microsoft and OpenAI

Microsoft is at once OpenAI's shareholder, its supplier and a beneficiary of its revenue. After the October 2025 restructuring, its stake was valued at about $135 billion, "roughly 27 percent", and OpenAI "has contracted to purchase an incremental $250B of Azure services" [83]. In January 2026, the chief financial officer reported that "Approximately 45% of our commercial RPO balance is from OpenAI" [84]; by my calculation, about $280 billion. In the fiscal year ended June 2026, Microsoft recorded $24.1 billion in revenue from OpenAI and had $6.0 billion receivable from it, against $11.9 billion invested since the beginning. The year's earnings also included $6.5 billion in net gains on the OpenAI stake, which according to the company "primarily relate to the dilution gain from the OpenAI Recapitalization" [3]. Being diluted at a higher valuation generated an accounting profit for the shareholder.

Amazon, OpenAI and Anthropic

With Amazon, money moves in both directions between the same two companies. In February 2026, Amazon committed to invest $50 billion in OpenAI, and OpenAI expanded its $38 billion contract with AWS "by $100 billion over 8 years", with about 2 GW of capacity on Trainium chips [85]. With Anthropic, Amazon invested $8 billion in convertible notes between 2023 and 2025 and another $5 billion in equity in the second quarter of 2026, and opened a financing facility of up to $20 billion [7]. On the other side, Anthropic announced: "We are committing more than $100 billion over the next ten years to AWS technologies, securing up to 5GW of new capacity to train and run Claude." [86].

Amazon's balance sheet shows the valuation effect. In the second quarter of 2026, the company recorded "upward adjustments of approximately $50.5 billion" in the value of its Anthropic shares, because of Anthropic's new funding rounds [7]. By my calculation, that is about three times AWS's operating income in the same quarter, which was $16.6 billion. Anthropic, for its part, raised $65 billion in May 2026 at a valuation of $965 billion and said its annualized revenue had passed $47 billion [87]. The growth is real; the investor's gain, for now, is an accounting one.

Anthropic, Google and the warning in Alphabet's report

Anthropic repeats the design with several suppliers. In November 2025, the company announced that it "has committed to purchase $30 billion of Azure compute capacity", while "NVIDIA and Microsoft are committing to invest up to $10 billion and up to $5 billion respectively in Anthropic" [88]. Google is the other major supplier. In October 2025, Anthropic announced that it would use up to one million Google TPUs, an expansion "worth tens of billions of dollars" [89]. Alphabet's annual report contains the clearest warning about the risk of this kind of deal, without naming companies: "Our financial instruments also include debt and equity investments in companies with which we also entered into commercial arrangements at or near the same time." And it continues: "if our investment in such companies becomes impaired, we may need to re-evaluate the accounting for the commercial arrangement, including reducing any remaining performance obligations." [4]. In other words, if the investment loses value, the backlog can shrink too. The two fall together.

CoreWeave and SB Energy

The same design appears among smaller companies. When it closed its first contract with CoreWeave, in March 2025, OpenAI received $350 million in CoreWeave's own shares [90]; the contracts between the two reached about $22.4 billion in September [91]. NVIDIA, which supplies all of CoreWeave's GPUs, bought $2 billion of the company's shares and committed to buying any capacity left without a customer through 2032 [92], [93]. In January 2026, OpenAI and SoftBank invested $1 billion in SB Energy, the SoftBank group's data center developer; OpenAI signed a 1.2 GW lease with it, and SB Energy "will become a major customer of OpenAI" [94]. In a single transaction, OpenAI invests in its own landlord, which becomes its customer and belongs to the group of one of its largest shareholders.

What the backlog is worth

These deals show up in the suppliers' remaining performance obligations. In mid-2026, Microsoft's commercial backlog was $678 billion [3], Oracle's $664 billion [18], Google Cloud's $514 billion [14], AWS's about $496 billion [7] and CoreWeave's $103.7 billion [93]. By my calculation, the total exceeds $2.4 trillion. On the other side are, for the most part, the same few labs: OpenAI alone speaks of commitments of about $1.4 trillion over eight years, against annualized revenue of about $50 billion at the end of September 2026 [95], [96].

The BIS economists measured the circularity systematically. Between 2021 and 2025, "55.2% of incoming investments in AI firms came from other AI firms", and 46.4% of the value of investments between AI companies also involved a supply relationship between investor and investee. The consequence, according to them: "circular relationships make reported demand partly endogenous to firms’ own financing decisions. For example, when a supplier finances a customer, part of the supplier’s revenue growth reflects its own capital investment, rather than organic final demand." And the comparison with telecom is theirs: "The parallel with the telecommunications boom of the late 1990s is instructive: upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors’ equipment" [51].

The IMF, in its October 2026 report, put numbers on the circle: "Fourteen publicly disclosed transactions since 2023 account for approximately $300 billion of aggregate equity investments led by hyperscalers in financially weaker AI firms, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments through services or chip orders". About $320 billion of that debt sits with the investee companies, "including debt backed by investor guarantees, contingent financing, and other off-balance-sheet debt". And it named the risk: "distress at key investee firms could present significant "wrong-way risk" to the AI ecosystem: Equity and debt investments in the firms depreciate in value precisely when future revenue streams tied to them are impaired." [97]. Jason Wu, of the IMF, summed it up at the press conference: "If there are questions about the monetization in AI, the equity investments will deplete in value, the debt may be difficult to service, and the revenue commitments in the future may not materialize", which he called a "triple whammy" [98]. In April, the Fund had already estimated that, of a rise of about 12 percentage points in the shares of the companies in the circle between September and December 2025, 7 points came from mutual reinforcement among them [31].

Not every deal is circular

The argument has limits. OpenAI's deal with Broadcom, for 10 GW of its own accelerators, involves no investment by the supplier in the customer; the companies signed a "term sheet" and did not disclose a value [99]. Many commitments are conditional on deliveries, and some never became contracts, such as NVIDIA's $100 billion. This reduces the effective risk, but it also shows that the headline numbers overstate firm demand. And Microsoft says that, in the April to June 2026 quarter, all of the growth in its backlog came from customers that are not frontier labs, as the section "What could disprove the thesis" shows.

The circle is not fraud, and each deal has a commercial logic: the labs need guaranteed capacity, and the suppliers need customers for the capacity they build. The problem is what the whole does. It inflates the suppliers' revenue and backlog with demand they financed themselves. It turns the labs' private valuations into accounting profit for the investors. And it concentrates risk in a handful of customers that are still burning cash. If one of them fails, the others lose on all three fronts at once: on the stake, on the credit and on the promised revenue.

Off the balance sheet: Hyperion, xAI and the engineering of data centers

The sections on Enron and on 2008 showed how off-balance-sheet vehicles work and how they come back onto the balance sheet in a crisis. In 2025 and 2026, the same toolbox was used again to finance AI: special purpose vehicles, joint ventures in which the company keeps a minority stake, residual value guarantees and renewable short-term leases. The difference from Enron is important: everything is disclosed, audited and, as far as is known, within the rules. So is the similarity: the debt stays off the balance sheet of the company that uses the asset and that, in practice, supports the payments.

Hyperion: a $27 billion data center off Meta's balance sheet

On October 21, 2025, Meta announced a joint venture with funds managed by Blue Owl Capital to build the Hyperion campus, in Louisiana. The Blue Owl funds took 80%, Meta 20%, and the total development cost was estimated at about $27 billion. Meta leased the entire campus under contracts with "a four-year initial term with options to extend" and gave the joint venture "a residual value guarantee for the first 16 years of operations" [100]. It also received a one-time distribution of $2.55 billion, according to the annual report; the announcement spoke of about $3 billion [8].

The money came from a bond. According to Bloomberg, Morgan Stanley structured "over $27 billion of debt and about $2.5 billion of equity into a special purpose vehicle", with PIMCO as the main buyer, and the notes received an A+ rating from S&P, according to a rating report. Bloomberg summed up the design: "Under the SPV structure, Meta is not borrowing the capital itself, the financing entity is." [101]. The issuer appears in a document that one of the Blue Owl funds filed with the SEC: Beignet Investor LLC, which holds the 80% stake in the joint venture, issued "$27,293,849 in senior secured notes with an interest rate of 6.581% and a maturity date of May 30, 2049" (amounts in thousands of dollars) [102]. The morning after the issue, the bonds were already being offered at 105% of face value [101].

Meta's annual report explains why none of this goes on its balance sheet. The leases begin in 2029, with an initial commitment of about $12.31 billion. The residual value guarantee has an aggregate limit of about $28 billion, which declines over time, and "As of December 31, 2025, RVG payments are not probable and therefore, no liability has been recorded." On the joint venture: "We do not have the power to direct the activities that most significantly impact the Venture's economic performance. Therefore, we are not the primary beneficiary and do not consolidate the variable interest entity (VIE). Our maximum exposure to loss related to the Venture was $45.95 billion as of December 31, 2025" [8]. In June 2026, the maximum exposure was $46.03 billion [9].

The argument is in the report for the first quarter of 2026. According to Meta, the decisions that most affect the joint venture's performance are "decisions pertaining to remarketing the data center campus, including but not limited to, negotiations with future lease tenants and individual property sales", and those decisions are not Meta's [103]. It is one possible reading of the accounting rule, which requires the power to direct the most significant activities and significant exposure to losses for consolidation [104]. But decisions about putting the campus back on the market matter only if Meta leaves, and the one who decides whether to leave, every four years, is Meta. Meta also builds and manages the campus. The auditor noticed how delicate this was: Ernst & Young included the consolidation of the joint venture among its critical audit matters and wrote that "Auditing the Company's determination of the primary beneficiary of the VIE was especially challenging" [8]. In the Wall Street Journal, columnist Jonathan Weil was more direct in the headline, "AI Meets Aggressive Accounting at Meta's Gigantic New Data Center", and in the subheadline: "Favorable treatment off the balance sheet hinges on some convenient assumptions." [105].

The rule on guarantees helps explain the design. Under the U.S. lease accounting standard, the lease liability includes only the residual value guarantee amounts that the company will probably pay. Deloitte notes that, for this reason, contracts in which much of the economics lies in a residual value guarantee "may continue to yield favorable accounting results (e.g., reduced leverage) under ASC 842" [106]. And leases that have not yet commenced go on the lessee's balance sheet only at the commencement date [107].

By my calculation, the declared maximum exposure of $46 billion was equivalent to 78% of Meta's total long-term debt at the end of 2025, which was $58.7 billion [8]. And the model is being repeated. In July 2026, Meta signed an exclusivity agreement for a second joint venture, in El Paso, Texas, again with 20%, and said it will provide "residual value guarantees with a maximum aggregate exposure of approximately $13 billion" [9].

Who ended up with the risk

With a private bond, finding out who owns it is difficult. Part of the ownership shows up in the reports that funds registered in the United States file with the SEC, the N-PORT forms. By my analysis of this data, with positions almost all as of March 31, 2026, 317 funds held $9.44 billion of the Beignet notes, about 35% of the issue. Funds of registrants with PIMCO in their name accounted for $6.57 billion, and the PIMCO Income Fund alone held $4.11 billion, about 15% of the issue. Funds offered only through insurance products, such as variable annuities, held about $473 million, and five closed-end funds had taken out loans pledging the notes as collateral [108]. The average price reported by the funds fell from 105.7% of face value in December 2025 to 102.9% in March 2026 [109], [108]. The other two-thirds of the issue are with holders who do not appear in this data, such as insurers' own portfolios, pension funds and private funds.

The equity also reached individual investors. The Blue Owl fund that holds 63% of Beignet Investor, with a commitment of $1.53 billion, is a non-traded real estate fund, sold in a private offering to "accredited investors", individuals and institutions with high net worth [102]. The risk of a single campus of a single company was spread across bond funds, ETFs, retirement plans and wealthy investors, as the mortgages of 2006 were spread across thousands of portfolios.

xAI: the GPUs in a vehicle of their own

Elon Musk's xAI used another version. According to Bloomberg, in October 2025 the company put together a $20 billion capital raise that included a special purpose vehicle "that will buy Nvidia processors and rent them to xAI", with about $7.5 billion in equity and up to $12.5 billion in debt; NVIDIA would invest up to $2 billion in the equity, and xAI would lease the chips for five years [110]. In an explainer, Bloomberg summed it up: "And xAI's exposure? Just a 5-year lease; nothing else." [111]. In January 2026, Apollo led $3.5 billion in financing for a Valor Equity Partners vehicle to buy $5.4 billion in infrastructure, including GB200 GPUs, and lease it to an xAI subsidiary. Apollo stated that "NVIDIA invested in VCI as an anchor Limited Partner" [112]. The chip supplier is a partner in the vehicle that buys the chips to lease them to the customer.

The same Bloomberg piece offered the comparison with the past in the words of a UBS strategist, Matthew Mish: "During the dot-com era, most of the growth was financed with equity not debt. So when it burst, the impact in the economy was manageable," and "Now, there's a rapid growth of capex in AI firms driven by debt and it's starting to be kept off balance sheet." [111].

Guarantees from the seller

The newest form of the design is the guarantee given by the supplier. In August 2026, NVIDIA reported to the SEC, under the form item that covers direct and off-balance-sheet financial obligations, residual value guarantees of up to $105 billion on SB Energy's leases to an OpenAI affiliate, for about 4.25 GW of capacity at a campus in Ohio. If OpenAI becomes insolvent or stops paying, NVIDIA covers "any shortfall between the guaranteed minimum value of a lease (as set forth in the Agreements) and amounts recovered through a replacement lease or sale". The guarantees end, among other cases, when OpenAI obtains "a satisfactory credit rating" [113]. According to CNBC, talks had begun in July with a guarantee of up to $250 billion, which would allow OpenAI to take on debt for a 10 GW campus "on the strength of Nvidia's credit" [114]. In return, the campus will house only NVIDIA infrastructure [77]. There are counterweights: the guarantees take effect in phases, the first expected in fiscal year 2029, decline over the 20 years of the leases, and OpenAI has committed to reimburse NVIDIA [77]. Jensen Huang described the goal: "We are securing long-lived infrastructure for NVIDIA compute so OpenAI can deploy the most productive AI factories that can be upgraded repeatedly with each new generation delivering more intelligence and better economics." [115]. The building lasts 20 years, the chips change with every generation, and the one guaranteeing the lease is the one selling the chips.

Google did the same with former bitcoin miners converted into data centers. It guaranteed the leases of the neocloud Fluidstack so that the miners could obtain financing: about $3.2 billion in the case of TeraWulf, with warrants that would bring Google's stake to about 14%; $1.4 billion in the case of Cipher Mining; and, in the case of Hut 8, a guarantee that "covers the lease payments and related pass-through obligations" of a $7.0 billion, 15-year contract, with banks financing up to 85% of the project cost [116], [117], [118]. According to Bloomberg, Google records these guarantees as credit derivatives [111].

In August 2026, NVIDIA took the idea to another scale. It signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing platforms able to "mobilize over $500 billion of third-party capital" and presented its chips as an investment asset: "NVIDIA compute is an investable asset" [119]. According to CNBC, NVIDIA will have the option to guarantee 25% of each loan. BlackRock's chief executive, Larry Fink, made the comparison this article has been making: "This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s". And the KKR executive in charge of digital infrastructure added: "you can securitize it or effectively divide that risk and sell it to investors who want to participate anywhere in that stack" [120]. The agreements were memoranda of understanding, with no final contracts.

What does not appear on the balance sheet

Added together, the leases signed but not yet commenced by the largest companies in the cycle exceed $1 trillion. None of them is on the balance sheet, because the accounting rule recognizes a lease only when it commences:

The sources are the companies' quarterly and annual reports [3], [18], [9], [7], [5], [93]. By my calculation, the total, about $1.15 trillion, is more than double the combined financial debt of Microsoft, Alphabet, Amazon, Meta and Oracle in mid-2026, about $480 billion. These are nominal amounts, payable over as long as 20 years, and part of them depends on contractual conditions; they are not debt. But they are fixed commitments, and Moody's says it treats them as debt equivalents [17].

The line between investment and leasing also moves. In July 2026, Microsoft raised the estimated useful life of its data centers from 15 to 25 years and explained the effect: "more of our future datacenter leases will shift from finance leases to operating leases as a result of this update. Finance leases are included in capital expenditures while operating leases are not." As a result, the capex forecast for the year fell from about $190 billion to about $175 billion, with investment plans "unchanged" [13]. The investment is the same; the number the market tracks got smaller.

Regulators have given the phenomenon a name. In March 2026, the BIS wrote: "These arrangements amount to "shadow borrowing": obligations that are economically akin to debt but largely reside outside corporate balance sheets. By channelling sizeable private credit into AI infrastructure, these structures strengthen links between hyperscalers and non-bank investors such as private credit vehicles and insurers." [121]. In its June annual report, it added that data center construction is increasingly outsourced to parties that lease back to the hyperscalers "on long-dated contracts with embedded exit clauses. The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times." [122]. The IMF, in October 2026, pointed to "the eventual realization of sizable off-balance-sheet commitments" as a future source of financing needs [123]. And it noted that data center securitizations, $46 billion since 2018, are expected to reach $150 billion in 2028 [31].

None of this is illegal, and today's accounting is much more demanding than that of 2001 or 2007. The question that 2008 left behind is a different one: what happens to these commitments when the asset loses value. Meta's residual value guarantee is paid only if Meta does not renew the leases and the campus is worth less than the limit, that is, if demand for data centers falls. NVIDIA's guarantee is called only if OpenAI does not pay. Both triggers depend on the same event, an AI correction, which at the same time knocks down the value of the guarantor's shares and of its cash flow. The debt has not disappeared. It has changed address, and it remains tied to the same risk.

The physics of depreciation: a GPU is not real estate

Every debt has some backing, and that backing determines how much the creditor recovers if something goes wrong. In an AI data center, the backing has three layers with very different lives. The building and the land last decades: Alphabet depreciates its data centers and offices "over periods of seven to 40 years" [4]. Power and cooling equipment lasts about a decade: CoreWeave depreciates "Data center equipment" over 8 to 12 years, and Amazon raised the useful life of the "heavy equipment" that supports its distribution centers and data centers from 10 to 13 years [124], [125], [6]. And the servers, where most of the money is, hold their value for only a few years. The question for this section is how many.

The supplier's rhythm

In June 2024, NVIDIA's chief executive, Jensen Huang, announced the pace of launches: "Our company has a one-year rhythm. Our basic philosophy is very simple: build the entire data center scale, disaggregate and sell to you parts on a one-year rhythm, and push everything to technology limits" [126]. Until then, the leap between generations took two years: the Blackwell launch announcement, in March 2024, describes the previous architecture, Hopper, as "launched two years ago" [127]. At the announced pace, whoever buys a server today and depreciates it over six years will still have it on the balance sheet after five or six annual launches by the supplier.

The supplier itself says how far each older generation falls behind. At the Blackwell launch, NVIDIA stated that the new platform made it possible to run trillion-parameter models "at up to 25x less cost and energy consumption than its predecessor" [127]. On the GB200 NVL72 system page, the comparison with the H100 is "30x" in language model inference and "25x" in energy efficiency [128]. In March 2026, the next generation, Vera Rubin, promised "one-tenth the token cost relative to NVIDIA Blackwell" [129]. And in August 2026 the company cited tests by the consultancy SemiAnalysis according to which the Vera Rubin NVL72 delivers "30x higher throughput per megawatt and 45x lower token costs than NVIDIA GB300 NVL72" [130]. These are the seller's numbers, measured under conditions chosen by the seller, and they should be read that way. But the signal is what matters: the seller is telling the market that equipment bought the year before produces much more expensive tokens than the new equipment.

The H100 page on NVIDIA's website even puts a price on the difference. According to the page, with SemiAnalysis data from April 2026, the H100 produces inference at about $0.09 per million tokens, and the B200 at about $0.02: "roughly 4.5x cheaper than H100" [131]. The comparison uses different software on the two chips and different per-user speeds, so it is not exact. The same site records the other side of the coin: the cost per token of Blackwell itself fell fivefold after launch, through software alone, "driven by continuous software optimization" [132]. The owner of a GPU gains from software optimization, but loses when the next generation arrives and the price of tokens falls for everyone.

The economic mechanism is simple. When supply keeps up with demand, the rental price of a GPU-hour tends to be pulled toward the cost of producing tokens with the most efficient equipment available. An old GPU keeps working, but it can only collect what is left between the market price of a token and its own cost, which includes the extra energy it consumes. That is why the economic useful life of a GPU is not the same as its physical life.

Physical wear

Physical life is not long either when the machine runs at its limit. Meta published one of the rare detailed accounts of failures in large-scale training. During 54 days of pre-training of Llama 3 405B, on a cluster of up to 16,000 H100 GPUs "each running at 700W TDP with 80GB HBM3", there were 466 interruptions, 419 of them unexpected: "Approximately 78% of the unexpected interruptions are attributed to confirmed hardware issues, such as GPU or host component failures, or suspected hardware-related issues like silent data corruption and unplanned individual host maintenance events." [133]. In the table of causes, 148 interruptions came from faulty GPUs and 72 from HBM3 memory failures [133]. Automation kept up: effective training time stayed "higher than 90%". But a fleet that loses components at this rate ages faster than a database server.

The building ages too

The idea that the building lasts decades and only the chip gets replaced runs into power. An H100 draws up to 700 W; a DGX B200 server, with eight Blackwell GPUs, "~14.3 kW max" [131], [132]. GB200 NVL72 racks are already liquid-cooled [128]. NVIDIA engineers wrote in 2025 that the power distribution used inside racks today "begins to hit physical limits" above 200 kW per rack, and announced an 800-volt direct current architecture "to support 1 MW IT racks starting in 2027". Under the current standard, a single 1 MW rack "requires up to 200 kg of copper busbar" [134]. The International Energy Agency (IEA) calculates that the power density of AI servers "increased by 11 times" between 2020 and 2025 and is expected to quadruple again by 2027 [22]. A building designed for one generation of chips may need major construction work, on power and cooling, to accommodate the next one.

Accounting went the other way

While the pace of chips accelerated, the big companies' accounting did the opposite: it lengthened the useful life of servers. A longer useful life spreads the cost over more years and raises each year's profit. The changes are in the annual reports:

The justifications vary. Microsoft cited "investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology" [136]. Meta, in 2022, "expected longer refresh cycles in our data centers" [137]. Amazon is the exception, and the most revealing one. In 2025, it shortened the life of some servers and gave the reason: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." [6]. At the end of 2024, the company had already decided to retire some servers early and booked "approximately $920 million of accelerated depreciation and related charges" [125]. Microsoft, in its 2026 annual report, began describing the useful life of servers and network equipment as "two to six years" [3].

The amounts involved grow with the investment. At Meta, depreciation of servers and network equipment was $13.36 billion in 2025, against $7.32 billion in 2023, and the gross value of these assets went from $68.4 billion to $98.0 billion in one year [8]. The IMF ran the numbers for the whole sector. Based on the depreciation disclosed by ten hyperscalers, the implied useful life of all property and equipment, buildings included, is about seven years, but GPUs and advanced chips "could face obsolescence within two years". The Fund also tested a shorter useful life: "Under a scenario of three-year useful life of capital and assuming hyperscalers fulfill these additional financing needs by issuing debt, their debt levels could rise from the present $800 billion to more than $1 trillion." Under the IMF's model, these companies' CDS spreads would rise by about 60 basis points, "a significant increase considering that these spreads are currently in the range of 20 to 160 basis points" [31]. The IMF treats obsolescence as a business risk, with no first-order effect on financial stability. For those who lend against these assets, however, it is the central question.

Burry, NVIDIA and the rental price

In November 2025, the investor Michael Burry brought the debate to the center of the market. He wrote that "Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era" [138] and that "Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment. Yet this is exactly what all the hyperscalers have done" [139]. According to Business Insider, he estimated that Meta, Google, Oracle, Microsoft and Amazon would understate depreciation by about $176 billion between 2026 and 2028 and that, in 2028, Oracle's earnings would be inflated by about 26% and Meta's by about 20%, "without elaborating on his calculations". Jim Chanos, the investor who bet against Enron, took the same line: "the chips basically depreciate over two years" [139].

The word "fraud" is Burry's and must be read with care. The useful-life changes are in the annual reports, with the effect in dollars, as the table above shows; the debate is about the estimate, not about concealment. NVIDIA responded in a memo to analysts. According to CNBC, the company argued that customers depreciate GPUs over four to six years "based on real-world longevity and utilization patterns" and that the A100s, launched in 2020, "continue to run at high utilization rates and retain meaningful economic value well beyond the two to three years claimed by critics". Burry shot back: "I am not claiming Nvidia is Enron. It is clearly Cisco." [140].

So far, the rental market has partly proved both sides right. Through 2025, prices fell. In June of that year, AWS cut the price of instances with NVIDIA GPUs by up to 45% [141], and the hourly price of a machine with eight H100s on the company's public price list went from $98.32 in August 2023 to $55.04 in October 2026 [142], [143]. In 2026, however, the rental price of reserved capacity rose. In AWS capacity blocks, the H100 went from $3.93 per GPU-hour in June 2025 to $5.97 on October 13, 2026, an increase of 52% [144], [145]. At Lambda, a neocloud, the on-demand price of the H100 went from $2.99 per hour in 2025 to $3.99 starting in April 2026, and the price of clusters for a few weeks went from $2.40 in November 2025 to $5.54 to $6.16 in October 2026 [146], [147]. According to CNBC, a batch of H100s from 2022 that became available when a contract at CoreWeave ended was rented out again "at 95% of their original price" [148].

The most cited explanation is scarcity. The Verge reported that "The price to rent old chips has been rising, and Silicon Data projects that it will continue rising through 2028", and an analyst interviewed by the site attributed the increase to "a major shortage of inference chips" [149]. The HBM memory shortage flagged by the IEA points in the same direction [22]. In August 2026, the derivatives exchange CME announced futures contracts on H100 and B200 rental indexes, citing "sharp swings in compute prices" [150]. In an October 2026 piece defending the durability of GPUs, NVIDIA cited third-party estimates, based on resale prices, of five to six years of useful life for an H100 system and nine to ten years for the GB300 NVL72 [151].

It is the best argument against rapid obsolescence, and it must be taken seriously. But a price sustained by scarcity is not the same as long-term value. The same NVIDIA piece states: "Power is the binding constraint on an AI factory. This makes tokens per second per megawatt the number that governs earning capacity." And it adds that Vera Rubin delivers "over 30x higher throughput per megawatt" than the GB300 [151]. With power limited, whoever swaps an old rack for a new one multiplies the output of the same megawatt. When the chip shortage passes, the pressure to retire old machines returns, and with it Burry's question.

Why this matters for lenders

The table below is just arithmetic, on a $100 investment in servers depreciated on a straight-line basis. It shows how much of the cost is still on the balance sheet after three years, depending on the useful life chosen:

With a six-year useful life, half the cost of a GPU is still on the balance sheet when it turns three, and at least two newer architectures are already on sale. If the market value of that three-year-old GPU is well below half the original price, the difference is a loss that has not yet shown up in earnings. For anyone holding the GPU as collateral for a loan, it is smaller collateral than it seems. CoreWeave, whose business is renting out GPUs, describes this risk in its own prospectus: "Any changes to the significant assumptions underlying our estimates or to the estimates of our components’ useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and future prospects." [124].

There are good arguments on the other side. GPUs that are no longer useful for training frontier models can spend years running inference or less demanding tasks. Software optimization increases the output of the same chip, as NVIDIA itself shows. And Alphabet notes that the choice of useful life draws on "historical asset performance, expected technology advancements, and our future infrastructure deployment plans" [4]: it is not an arbitrary number. The conclusion is not that the companies got the math wrong. It is that the collateral backing much of AI's debt loses economic value faster than the buildings, faster than the transmission lines and, in several structures, faster than the debt itself. It is the fifth marker in the bubble table, and it is confirmed by those who sell the chips and by those who account for them.

Energy: the bottleneck that may become a stranded asset

An AI data center is, above all, a grid connection. The chips change every year, but the substation, the transmission line and the power plant that feed them are designed to last 20, 30 or 40 years. It is the same mismatch of time horizons as in the previous section, only now off the balance sheets of the technology companies. Whoever builds the energy infrastructure for a data center takes on the risk that demand will disappear before the project pays for itself. When that happens, what is left is called a stranded asset: an investment that is still standing but does not generate the revenue that justified building it.

How much AI consumes

The IEA estimates that data centers consumed 415 TWh in 2024, "around 1.5% of the world’s electricity consumption". The United States accounted for 45% of that total, China for 25% and Europe for 15%. In the agency's base case, consumption "is set to more than double to around 945 TWh by 2030", slightly more than Japan's total electricity consumption today. In the United States, data centers account for almost half of the growth in demand through 2030, and by the end of the decade the country "is set to consume more electricity for data centres than for the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined." [152]. The April 2026 update confirmed the trajectory: data center consumption grew 17% in 2025, and that of data centers dedicated to AI, 50% [22].

For the United States, the Lawrence Berkeley National Laboratory (LBNL) calculated that data centers consumed 176 TWh in 2023, 4.4% of the country's electricity. For 2028, the projection ranges from 325 to 580 TWh, from 6.7% to 12.0% of projected consumption [153]. The high scenario is almost double the low one. Anyone who builds a power plant or a transmission line for this market has to pick a number within that range and live with it for decades. The same report assumes that AI servers last as long as conventional ones, about 5.7 years in 2023: "For AI accelerated and nonaccelerated servers, the same lifetime assumptions are used." [153]. Even the energy projections carry the useful-life assumption questioned in the previous section.

The grid queue

The limit lies not only in generation but in the connection. The IEA calculates that, without measures to unblock the grid, "around 20% of planned data centre projects could be at risk of delays". In advanced economies, a new transmission line takes four to eight years, and "wait times for critical grid components such as transformers and cables have doubled in the past three years" [152]. The IMF speaks of interconnection times "now stretching from two to seven years" [31].

The U.S. interconnection queue shows the pace of the grid. According to LBNL, "As of the end of 2025, over 2,060 gigawatts (GW) of total generation and storage capacity were actively seeking connection to the grid." For projects completed in 2025, the median time between the request and the start of operations was more than five years. And "Only 13% of capacity that submitted interconnection requests from 2000-2020 had reached commercial operations by the end of 2025; 75% of that capacity had been withdrawn" [154]. That queue is made up of power plants and batteries, not data centers, but it shows what happens to connection requests: most never get built. NVIDIA itself acknowledged the problem in March 2026: "Energy is now the biggest bottleneck for AI infrastructure buildouts, with over $300 billion in equipment backlogs and more than 200 gigawatts of projects waiting in U.S. interconnection queues." [155].

Turbines and reactors

With no room on the grid, many projects turn to their own generation, almost always gas-fired. GE Vernova, one of the largest turbine manufacturers, reported that its gas equipment backlog, together with production reservations, grew from 62 to 83 GW in the fourth quarter of 2025 and from 100 to 116 GW in the second quarter of 2026, and that it expects to reach at least 125 GW by the end of the year. The company's output reaches a rate of 20 GW a year in the third quarter of 2026, is expected to rise to 24 GW in 2028 and has a target of 30 GW in 2030 [156], [157]. By my calculation, the 116 GW under contract or reserved amount to almost six years of current production. The IEA records "A 70% surge in gas turbine orders in 2025" and calculates that gas generation on the data center site needs to be oversized by 30% to 70% relative to demand to guarantee supply [22].

The other bet is nuclear, with contracts that run far longer than any chip:

The Crane restart also secured a $1 billion loan from the U.S. Department of Energy [162], which puts part of the cycle's risk on the government's balance sheet. For the seller of the power, the 20-year contract is the guarantee that underpins the investment. For the buyer, it is a fixed cost for two decades, taken on to power equipment depreciated over five or six years.

Who pays if demand disappears

Some U.S. utilities already require guarantees with this risk in mind. In Ohio, AEP reached a settlement in 2024 with the staff of the state regulator (PUCO), the consumer advocate and other parties. The settlement "requires large new data center customers to pay for a minimum of 85% of the energy they say they need each month – even if they use less – to cover the cost of infrastructure needed to bring electricity to those facilities." [163]. The regulator approved the tariff in July 2025, with minimum usage bands capped at 85% of contracted capacity. The initial contract lasts for the ramp-up period, of up to four years, plus eight years. A customer without a high credit rating must post collateral for 50% of all the minimum payments under the contract. Only after the fifth year following the ramp-up can it exit, by paying 36 months of minimum charges. And the letter of commitment "requires customers to reimburse Ohio Power 100% of the buildout costs if the customer cancels — or delays the project by more than 12 months — prior to the target energization date." [164]. In Texas, state law SB 6, on the connection of large loads, aims, according to AEP, at "minimizing the potential for stranded infrastructure costs" [165].

It is the logic of any prudent lender: the utility cannot verify whether the promised demand will materialize, so it demands collateral from whoever made the promise. The risk does not disappear. It passes to the owner of the data center, which may be a special purpose vehicle like those in the section "Off the balance sheet". If the utility cannot collect, the cost lands back on the electricity bills of other consumers. The IEA describes this risk: data centers "often oversize their grid connections initially", and the mismatch between fast-moving data centers and slow-moving energy investments "could raise electricity prices in some places if unmitigated." [22].

The agency also links the pace of energy to that of credit: "Data centre investments have grown too large to be funded from company balance sheets alone, and large amounts of funding from capital markets will be critical for their buildout. As a result, the pace of data centre growth, and the resulting increase in energy consumption, will be sensitive to market sentiment, including expectations for returns on investment in data centres and AI deployment, as well as to broader macroeconomic and financing conditions." [22]. In other words, the energy demand that justifies turbines and transmission lines depends on the mood of the credit market, which can change within weeks.

The physics on the grid side

There is also an engineering problem that does not show up in spreadsheets. A training cluster does not draw power steadily. Meta reported that, while training Llama 3, tens of thousands of GPUs could increase or reduce their consumption at the same time, for example while waiting for a checkpoint, and that this "can result in instant fluctuations of power consumption across the data center on the order of tens of megawatts, stretching the limits of the power grid." [133]. At the same time, the IEA points out that an AI data center "is 10 times more capital-intensive than an aluminium smelter, which means curtailing its operations to provide flexibility to the grid is very costly." [152]. The data center has to run almost nonstop to pay back its capital, and the grid has to absorb swings that other industries do not produce.

Efficiency cuts both ways. According to the IEA, "Software and hardware advances have resulted in the energy use per AI task dropping by at least an order of magnitude annually in recent years." [22]. Each task becomes cheaper, which can multiply usage. But the same efficiency reduces the energy needed for a given volume of work and, as seen in the previous section, drives down the value of older machines.

Energy is the part of the AI cycle that most resembles the railways and fiber: long-lived assets, built ahead of demand, that go on existing even if investors lose money. In 2002, dark fiber became cheap bandwidth for the following decade. Idle substations and turbines would also find another use. The question is who pays the bill in the meantime: the data center owner, if the minimum-usage tariffs work; that owner's creditors, if it goes bust; or the other electricity consumers, if the rules fail.

The weakest links

Credit crises rarely start with the strongest institutions. In 2007, the first to fall were the vehicles and funds that depended on short-term funding and on collateral that was losing value [60]. In the AI cycle, the big hyperscalers have solid balance sheets: according to the Bank of England, "the AI hyperscalers all have a credit rating of AA- or higher" [166]. The risk is concentrated in three types of company: the AI lab that signs commitments far larger than its own revenue, the traditional company that took on debt to become a capacity supplier, and the neoclouds, which rent out GPUs bought with debt secured by the GPUs themselves. The best-documented examples are OpenAI, Oracle and CoreWeave.

OpenAI: trillion-dollar commitments, revenue in the tens of billions

OpenAI's growth is real. Its chief financial officer, Sarah Friar, wrote in January 2026 that annualized revenue had gone from "$2B ARR in 2023, $6B in 2024, and $20B+ in 2025", keeping pace with computing capacity, which went from 0.2 GW to about 1.9 GW over the same period [167]. Annualized revenue is the run rate at the end of the year, not the revenue for the year: according to CNBC, 2025 revenue was $13.1 billion. In March 2026, the company said it was "generating $2B in revenue per month" [168], [169]. Usage is real too: in February 2026, ChatGPT had "more than 900M weekly active users" and more than 50 million subscribers [75]. By my calculation, about 5.5% of weekly users were paying.

The problem is the scale of the commitments. In November 2025, Sam Altman wrote: "We are looking at commitments of about $1.4 trillion over the next 8 years" [95]. By my calculation, that works out to an average of $175 billion a year. At the end of September 2026, according to CNBC, OpenAI reported annualized revenue of about $50 billion to investors, less than a third of that; the $68 billion that circulated the previous month included gross revenue from partners [96]. OpenAI does not publish audited financial statements. What is known about its cash comes from documents sent to shareholders and reported by the website The Information: the company projected burning $115 billion through 2029 [170], then forecast "$111 Billion More Cash Burn Through 2030" [171] and, in April 2026, was expected to burn "more than $200 billion before it starts generating cash" [172]. In the first quarter of 2026, the company burned $3.7 billion, "more than half its $5.7 billion in revenue" [173].

The money came from ever-larger rounds. OpenAI raised $40 billion at a $300 billion valuation in March 2025 and, a year later, closed $122 billion at $852 billion. For the first time, it opened the round to individuals through banks, "raising over $3 billion from individual investors", and announced its inclusion in exchange-traded funds from ARK Invest [169]. In June 2026, it confidentially filed with the SEC to go public [174]. Part of the shareholders' money also came from debt. SoftBank, which invested a total of $64.6 billion for about 13% of OpenAI, paid for its 2026 contribution of $30 billion first with a bridge loan from banks and, for the last installment, with bonds of about $11.1 billion, rated BB+, below investment grade, and with interest of up to 9.75% a year [175], [176], [177]. Strategists interviewed by CNBC in May were already reading the lineup of mega-offerings, with SpaceX, OpenAI and Anthropic, as a sign of a market top. One of them drew the comparison: "Back in 1999, we saw the same kind of thing where people were just rushing to get these IPOs out." [178].

OpenAI's balance sheet is light. According to The Information, which read the company's financial statements, on March 31, 2026 it had "zero debt" [179], and its bank credit line of about $4.7 billion was untouched [169]. The debt exists, but it sits on other balance sheets: in the data center leases signed by Oracle and developers, in the neoclouds' contracts and in NVIDIA's guarantees, which the section "Off the balance sheet" details. NVIDIA itself explained why it needs to guarantee its customers: "AI model makers have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity to secure the AI infrastructure necessary to grow." [77]. That is the definition of vendor financing: the supplier gives the customer the credit that the market will not.

In November 2025, the CFO went as far as mentioning a federal "backstop" to finance chips and walked it back the same day: "I used the word 'backstop' and it muddied the point" [180]. Altman said that OpenAI neither has nor wants government guarantees for its data centers and wrote: "But we of course could be wrong, and the market—not the government—will deal with it if we are." [181]. Analyst Gil Luria, of D.A. Davidson, drew the conclusion: "If sanity prevails and this federal guarantee doesn't materialize, it could become clear that OpenAI cannot live up to its extravagant commitments. In which case we would view the companies propped up by OpenAI promises as the most vulnerable," and named Broadcom, AMD, Oracle, CoreWeave and Oklo [182]. In April 2026, the Wall Street Journal reported, according to CNBC, that OpenAI had missed its own revenue and user targets and that the CFO was worried about the ability to pay for future computing contracts. The company called the report "ridiculous", but shares of Oracle and of chipmakers fell the same day [183]. On October 8, 2026, the day CNBC confirmed the annualized revenue of $50 billion, NVIDIA shares fell 3%, Oracle's almost 6% and CoreWeave's almost 8% [96]. The transmission channel is already in sight.

Oracle: a software company that became AI's creditor

No large company has changed its profile as much. At the end of the fiscal year that closed in May 2025, Oracle's backlog of contracts yet to be fulfilled (remaining performance obligations, or RPO) stood at $137.8 billion. Three months later, it stood at $455 billion, "up 359%". Then-CEO Safra Catz said: "It was an astonishing quarter" [184]. The table shows what came with it, in billions of dollars:

The figures come from the company's reports and releases to the SEC; those for the March to May 2025 quarter are derived from the difference between the fiscal year and the preceding nine months [185], [12], [10], [186], [18]. In fiscal year 2026, free cash flow was negative $23.7 billion, in Oracle's own words [187]. And only about 13% of the $664 billion in RPO is expected to turn into revenue in the following 12 months [18].

Oracle does not say how much of the RPO comes from each customer, but it warns in its annual report: "The economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations." [10]. Analysts did the math: according to D.A. Davidson, at least $300 billion of the $523 billion in RPO in November 2025 came from OpenAI [188]; according to Bank of America, more than half of the RPO in June 2026 [189]. Asked whether OpenAI would be able to pay $60 billion a year for cloud, co-CEO Clay Magouyrk replied "of course" [190].

To pay for the buildout, Oracle went to the market. In September 2025, it issued $18 billion in bonds in six tranches, the longest maturing in 2065, with a 6.10% coupon [191]. In February 2026, another $25 billion in eight tranches, the longest running to 2066 with a 6.85% coupon, plus $5 billion in convertible preferred stock [192]. From June to August 2026, it sold $20 billion in stock on the market [186]. The ratings stood at Baa2 and BBB, both with a negative outlook, "just two notches above junk status", in CNBC's words about the Moody's rating [17]. According to the company's own term sheets, the spread over Treasuries rose between the two issues: from 1.05 to 1.45 percentage points on the ten-year bonds and from 1.37 to 1.95 points on the 40-year ones [191], [192].

Off the balance sheet, the commitment is even larger. Leases already signed but not yet commenced, "substantially all related to data center arrangements", with terms of 15 to 19 years, went from $43.4 billion in May 2025 to $288 billion in August 2026 [12], [18]. That is more than double all of the company's financial debt. Oracle also guaranteed up to $3.3 billion of a loan taken out by one of its landlords [10].

Part of the investment began to be paid for by customers. In the quarter from June to August 2026, Oracle received $11.4 billion in prepayments "that included a significant financing component", that is, with interest built in [18]. The company then created a new metric, outside the accounting rules, the "net cash outlay for capital expenditures", which deducts from capex those prepayments and part of the flows with equipment manufacturers classified as financing. By that measure, investment in fiscal year 2026 was $47.7 billion, against $55.7 billion in capex under the accounting rules; in the following quarter, $18.0 billion against $28.5 billion [186]. Oracle does not identify who advanced the money. The cloud customers it names itself include AMD, Meta, NVIDIA, OpenAI, TikTok and xAI [193], and OpenAI depends on new investment rounds to pay its own bills.

The market reacted. In June 2026, the stock was down about 55% from its September 2025 peak and had "the steepest weekly drop since a 20% plunge in August 2001, during the depths of the dot-com bust" [194]. In July, Torsten Slok, of the asset manager Apollo, told CNBC: "CDS for Oracle is at the same level it was at in 2008." [195]. Buyers of the September 2025 bonds sued the company, alleging that it did not disclose that it would need much more debt, an allegation that has not yet been ruled on [196]. In September 2026, Oracle sent a force majeure notice to the developer of Project Jupiter, in New Mexico, to protect itself from higher costs; according to the Financial Times, cited by CNBC, the $18 billion in debt tied to that data center was already trading "at stressed levels" [197].

CoreWeave and the neoclouds: the GPU as collateral

CoreWeave rents out NVIDIA GPU capacity to labs and hyperscalers. On June 30, 2026, it had $35.1 billion in debt against $5.0 billion in equity, about seven times, plus $16.3 billion in lease liabilities and $35.5 billion in leases signed but not yet commenced, off the balance sheet [93]. In the second quarter of 2026, revenue was $2.6 billion, up 112% from a year earlier, but net interest came to $640 million, about a quarter of revenue, and the loss was $626 million. In the first half, the company invested $14.1 billion and generated $3.7 billion in cash from operations [93].

The detail that links CoreWeave to the section on depreciation lies in its loan agreements. The credit limit is calculated "based upon the depreciable cost of computing equipment", and in one of the facilities, "based upon the depreciable cost of graphics processing unit servers" [93]. In other words, how much the company can borrow depends on the book value of the GPUs, which depends on the useful life chosen. Effective rates range from 7% to 15% a year, and part of the debt sits in bankruptcy-remote special purpose subsidiaries, which had pledged $18.2 billion in long-term assets as collateral [93].

In 2026, this credit became an asset class. In March, an $8.5 billion facility received an A3 rating from Moody's; CoreWeave described it as "the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract" [198]. In August, another facility came with an explicit mismatch: "the DDTL 5.5 Facility carries an approximate five-year maturity while its underlying customer contracts average approximately three years in length. By financing these shorter-date commitments, lenders are signalling confidence in the long-term value of NVIDIA GPUs running on CoreWeave’s cloud platform and a willingness to underwrite renewal risk." [199]. The lender lends for five years against three-year contracts and bets that the GPU will still have a customer when they end.

Revenue is concentrated. Microsoft accounted for about 67% of 2025 revenue; in the second quarter of 2026, the three largest customers accounted for 36%, 26% and 10% [200], [93]. The company warns that, if a large customer fails, it may be left with idle capacity and with the expenses for equipment, leases and financing, and adds: "We do not currently maintain credit insurance to insure against customer credit risk." [93]. On the other side is NVIDIA, which supplies all of CoreWeave's GPUs, bought $2 billion of the company's shares in January 2026 and committed, in September 2025, to buy any capacity left without a customer through 2032, under an initial contract of $6.3 billion [92], [93]. The section "The circle" already covered this contract.

Nebius, another neocloud listed in the United States, shows the same pattern. In the first half of 2026, it invested $8.1 billion and had revenue of $0.98 billion; operating cash flow was positive only thanks to $4.4 billion in customer prepayments. In January, it extended the useful life of servers and network equipment from four to five years [201]. The $17.4 billion contract with Microsoft only takes effect once Nebius confirms "that it has secured any additional financing required in respect of the capital expenditures under the Agreement" [202]. The contract depends on the financing, and the financing depends on the contract.

The cost of money has gone up. At the end of September 2026, the ten-year U.S. Treasury was yielding close to 5.17%, according to CNBC, and lenders reported greater selectivity: "Instead of a roster of 50 neoclouds, there's probably 20 that the market's truly interested in." [203]. CoreWeave calculates that each percentage point increase in interest rates adds about $30 million per quarter to its interest expense on floating-rate debt [93]. Analysts at Mizuho were already talking in July about "neoclouds generating negative free cash flow, widening credit spreads and capital raise concerns" [195].

The three links are bound to one another. OpenAI's commitments are a significant part of Oracle's and CoreWeave's backlogs; Oracle's and CoreWeave's debt rests on those contracts; and NVIDIA guarantees pieces of the chain. If OpenAI's revenue disappoints, the pressure reaches Oracle and CoreWeave first, then their creditors and, finally, NVIDIA's guarantees. It is the wrong-way risk described in the section on 2008: the protection loses value precisely when it is called upon.

Macroeconomics and the official warnings

The bust of 2000 sank the stock market and brought on an eight-month recession. The bust of 2008 broke banks and brought on the longest recession since World War II, lasting 18 months [40]. The difference lay in the channel: stocks in one case, credit in the other. This section measures the weight of AI in the U.S. economy and shows what central banks and international bodies have said about the channel through which a correction would pass.

How much of the growth comes from AI

In October 2025, Fortune magazine published a calculation by Harvard economist Jason Furman: "Technically, as Furman notes, investment in information-processing equipment and software was only 4% of U.S. GDP for the first half of 2025, yet it also accounted for fully 92% of GDP growth over that period." Furman himself cautioned that, without the AI boom, interest rates and electricity would probably be cheaper, which would give some growth to other sectors [204]. I redid the calculation with the data available at the time and it checks out: the 92% comes from subtracting the chained-dollar values of technology investment from those of GDP [205]. The problem is the method. The BEA, which calculates U.S. GDP, warns that, for products with rapidly falling prices, such as computers, "the chained-dollar levels (as distinct from chain-weighted indexes and percent changes) will overstate their relative importance to GDP growth." [206].

According to the BEA's own table of contributions, in the revised version of September 2026, information processing equipment and software accounted for about 56% of growth in the first half of 2025, 30% of growth for the full year (0.70 of 2.3 percentage points) and about 41% in the first half of 2026, by my calculation [23]. These are gross figures. Much of the equipment is imported, and GDP adds investment but subtracts imports. U.S. imports of computers, peripherals and parts went from $217 billion in 2024 to $362 billion in 2025 and reached an annual rate of $601 billion in the second quarter of 2026 [23]. Official estimates vary widely depending on the method:

The sources for the table are, in order, [204], [207], [166], [208], [23] and [31]. In gross terms, technology investment accounted for somewhere between half and two-thirds of U.S. growth in the first half of 2025. Net of imports, the effect was much smaller. The gap between the estimates already says something: the weight of AI in the U.S. economy depends on the method, and anyone who asserts a single number is choosing one.

The effect does not stop at investment. Data center construction in the United States reached an annual rate of $84.95 billion in August 2026 (preliminary data), 73% higher than a year earlier, while total construction fell 1.7% [209]. And the rise in stock prices also affects consumption: Bank of England staff estimated that the appreciation of U.S. stocks accounted for "around 10% of the increase in US real consumption since the start of 2024, with AI-related equities the main driver during this period." [166].

Economists at the BIS, the central banks' bank, put the cycle in historical perspective. By their calculations, spending on IT equipment factories and data centers was equivalent to about 1% of U.S. GDP in mid-2025, "similar in size to the US shale boom of the mid-2010s and half as large as the rise in IT investment during the dot-com boom of the 1990s." And they added: "The end of previous investment booms was associated with a slowdown in GDP growth of more than 1 pp on average (Graph 4.B). Notably, the largest contraction was after the US dot-com boom, even though the boom was small relative to GDP." [208]. The size of the cycle, on its own, does not say how much it can hurt.

Concentration

The second channel is the stock market. In October 2025, the Bank of England's Financial Policy Committee noted that the five largest companies in the S&P 500 made up about 30% of the index, a share "higher than at any point in the past 50 years", and that, as a result, "any AI-led price adjustment would have a high level of pass-through into the returns for investors exposed to the aggregate index." [210]. The IMF calculated the same month that the seven largest technology companies were 33% of the index and that a concentration measure based on the Herfindahl-Hirschman index was "substantially higher than during the dot-com bubble" [207]. In July 2026, the Bank of England estimated that AI-related companies already made up about half of the S&P 500, compared with a quarter in 2022, and that the share could grow with the OpenAI and Anthropic stock offerings [1].

The BIS applied a statistical test used to detect bubbles, which measures whether a price has entered an explosive trajectory. In December 2025, both the S&P 500 and gold had "entered explosive territory", and it was "the only time in at least the last 50 years in which gold and equities have entered this territory simultaneously". The BIS itself points out the limit of the test: "it provides no information on when bubbles may burst." [211].

From equity to debt

What worries central banks most is not stock prices but the change in how the bill is paid. The European Central Bank wrote in May 2026 that AI companies and infrastructure "have started to rely increasingly on credit financing" and recalled a historical pattern: "Strong growth in either equity prices or business debt in isolation has not been associated with financial crises in the past. That said, 15% of those years that showed particularly strong growth in both equity prices and business debt have been followed by a financial crisis within two years". The chart accompanying the text has a blunt title: "Greater debt financing for AI is concerning for financial stability" [212]. The same report adds the caveat that corporate debt growth, in aggregate, is still low in the United States and the euro area.

The Bank of England's figures, from July 2026, show the speed. The five largest hyperscalers had only 3% of U.S. investment-grade corporate debt at the end of 2025, but accounted for more than 15% of issuance in 2026 through early May. AI-related issuers made up 41% of U.S. high-yield issuance (excluding refinancings) in the year, and, according to the OECD, the share of private credit in AI financing rose from 9% in 2024 to 34% in 2025 [1]. In September, the bank's committee cited a Morgan Stanley estimate: about $450 billion in AI-related debt issued in 2026 through the beginning of the month, more than double the whole of 2025 [213]. In the Fed's survey of market participants, AI went from 30% of mentions among risks in 2025 to 50% in 2026, and respondents noted that the investments "are increasingly funded by debt, creating leverage in the system" [214], [70].

The table summarizes how the official warnings have evolved:

The BIS annual report, from June 2026, makes the historical connection that this article proposes. The manias for canals in the 1830s, railways in the 1840s, electrification in the 1920s and the internet in the 1990s "all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify. These episodes ended with an eventual reversal in investment, inducing economy-wide recessions." And on the current cycle: "The scale and pace of the current AI investment boom accompanied by expectations of large productivity payoffs bear resemblance to these precedents, highlighting potential downside risks in the near term." [122].

July 2026: a rehearsal

The market has already had a rehearsal. According to the Bank of England, "Equity valuations for AI companies fell sharply in July. The scale of the adjustment was amplified by an unwinding of stretched positions and associated deleveraging activity." There were large losses for leveraged investors, but "there was no spillover to core markets." [213]. The BIS recorded the collapse of "a highly leveraged AI-focused hedge fund", a drop of almost 15% in the price-to-earnings ratio of the hyperscalers and of 40% in that of semiconductor makers, and a rise in the CDS spreads of those companies, while those of the rest of the S&P 500 held steady [217]. On July 27, Apple overtook NVIDIA as the most valuable company in the world; according to CNBC, investors "rewarded its reluctance to spend heavily on capital expenditures for AI, preferring to rent capacity instead of building its own." [218].

The episode showed both things at once. The market already demands a premium from those who invest and borrow the most. And, as long as most of the cycle is paid for with cash and equity, a sharp drop in the stock market does not yet turn into a credit crisis.

What the regulators do not say

None of these documents predicts a crisis. The Bank of England wrote in December 2025 that "The predominantly cash and equity-financed nature of AI development to date limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term." [166]. The IMF, in April 2026, judged that "financial stability risks remain contained" [31]. And the ECB noted that, unlike in 2000, the hyperscalers "combine high profit margins, strong earnings growth, little debt and diversified underlying businesses beyond AI" [215].

What they all point to is the mechanism: the shift from financing through cash and equity to financing through debt, on and off the balance sheet. It is this shift that can turn a stock market correction, like the one in 2000, into a credit problem, like the one in 2008. The Bank of England simulated a hypothetical AI-related correction scenario, with a 45% drop in U.S. stocks over six quarters and a 350-basis-point rise in credit spreads, which would take 2.2 percentage points off UK GDP [1]. And it pointed out, in September, that growth projections and public finances also depend on AI: "A reassessment of those expectations could therefore affect not only AI-related asset valuations but also sovereign debt markets." [213].

What about Brazil?

Brazil is not at the center of the circle described in this article. No Brazilian company sells GPUs, trains frontier models or issues tens of billions of dollars in bonds to build data centers. But the country is linked to the cycle through three channels: public policy, which is betting on attracting data centers; energy, which is the country's main selling point and the most durable part of the investment; and investors, who are exposed to U.S. stocks and to global financial conditions.

Redata: the incentive that arrived late

The main federal bet is the Special Tax Regime for Data Center Services, or Redata. Provisional Measure 1,318, of September 17, 2025, suspended PIS/Cofins, IPI and Import Tax on purchases of technology equipment by qualifying data centers. In return, a company had to offer the domestic market at least 10% of the capacity installed with the benefit, meet all of its energy demand with clean or renewable sources, have a water usage effectiveness (WUE) of no more than 0.05 liters per kWh and invest 2% of the value of the products in research and development [219]. The explanatory memorandum states that "about 60% of national digital workloads" are processed abroad and that operating in Brazil costs, on average, 30% more, "mainly due to taxation on Information and Communication Technology (ICT) equipment". The Federal Revenue Service estimated the forgone tax revenue at R$5.20 billion in 2026, R$1.00 billion in 2027 and R$1.05 billion in 2028 [220]. The most repeated figure, however, is not in any technical document. At the signing ceremony, Vice President Geraldo Alckmin said: "There is an expectation that Brazil could attract R$2 trillion in investments, over 10 years" [221]. No methodology has been published for that figure.

The provisional measure expired on February 25, 2026, without being voted on [222]. According to Agência Brasil, the President of the Senate, Davi Alcolumbre, decided not to put the text to a vote before the deadline [223]. Its content came back as a bill, passed by the Chamber of Deputies on February 24 and by the Senate only on September 1 [224], [225], and was signed into law as Law 15,504, of September 15, 2026. The law replaced "clean or renewable sources" with "renewable or low-emission sources" and kept one decisive detail: the PIS/Cofins and IPI benefits "shall have effect until December 31, 2026", because of the transition to the CBS and the IBS under the tax reform [226]. On the same day, a complementary law created an exception to the fiscal rules so that the benefit could be granted [227]. When the law was signed, Agência Brasil noted that some of the rules "will still depend on regulations to define procedures for the qualification of companies, technical criteria and other conditions necessary for the effective use of the tax benefits" [228]. As of October 13, 2026, a search of the Official Gazette had not found the implementing decree [229]. If it is delayed, the window for the main tax benefits in 2026 will be short.

Announcements and official acts

The amounts announced for data centers in Brazil are enormous, but they change in nature once you read the official document:

The ByteDance case is the most instructive. In its approval by the council of the Export Processing Zones, the project "provides for total investments of R$108 billion through 2035 in high-tech equipment" and, "For the decade after 2035 (36-46), the company expects to invest another R$135B in technological upgrades" [232]. Almost all of the headline number is equipment and its replacement over two decades, the same short-lived asset discussed in the section on depreciation. The first concrete milestone is much smaller: "The expectation is that, by the third quarter of 2027, at least 40MW will be in full operation." [234].

In Rio Grande do Sul, the state government announced that the Ministry of Mines and Energy had published "the resolution that guarantees the connection of 5 gigawatts (GW) of power for the data center city in Eldorado do Sul" [236]. The published ordinance does not mention any capacity at all. It recognizes that the grid access option "meets the criteria of minimum global cost", makes the connection conditional on "system availability verified by the ONS" and is revoked if there is no transmission system use contract by December 31, 2033 [237]. Elea spoke of 1.5 GW of renewable energy "to be delivered immediately" at Rio AI City [238], but the concrete phases described by the company itself add up to about 200 MW [239]. In April 2026, the announcement of Elea's sale to I Squared Capital mentioned "more than 1 GW of planned capacity and more than 300 MW on land with guaranteed power" [240]. To compare scales: Equinix's SP6, a real data center in operation, cost $114 million, and the company says it has invested in the country "on average, about R$1 billion a year over the last decade" [241].

None of this indicates fraud. Announcing potential is common in infrastructure. The problem, in a cycle of euphoria, is that the headline number circulates as if it were investment already made and ends up serving as a basis for public policy. Electricity planning shows the size of the gap. In a September 2025 presentation, the Energy Research Office (EPE) showed that the data center requests tracked by the ministry added up to 10,721 MW by 2029 and 20,986 MW by 2038, of which 13,399 MW were in the Southeast. EPE's own forecast for new data center load on the main grid in 2029 is around 1,000 average MW, reading off the chart [242]. The units are different, requested capacity versus average energy, so the comparison holds as an order of magnitude. A search of the Official Gazette found at least 15 ordinances from 2024 and 2025 recognizing data centers' access to the main grid, and none of them states a capacity [243]. To filter out speculative requests, ANEEL began requiring financial guarantees from those who request access to the transmission grid [244], the same logic as the Ohio tariff described in the section on energy.

Energy to spare, grid in short supply

Brazil's selling point is clean energy. Renewables made up 86.8% of the electricity mix in 2025 [245]. And there is energy to spare. By my calculation using the ONS open data, the wind and solar plants monitored by the operator were prevented from generating 37.2 TWh in 2025 by curtailment orders, 20.7% of what they could have produced. The peak was in October 2025, at 30.4%, and 84% of the volume curtailed that year was in the Northeast. From January to October 12, 2026, the preliminary data add up to 31.3 TWh, or 22.1% [246], [247]. Under Law 15,269/2025, the costs of curtailment due to "oversupply of electricity, referring to events caused by the impossibility of allocating electricity generation to the load" can no longer be covered by the charge provided for in the sector's law [248]. In practice, the generator is not compensated for this type of curtailment.

Data centers are presented as buyers of this surplus, but the numbers do not add up on their own. The energy curtailed in 2025 is about four times the new data center load that EPE projects for 2029 (1,000 average MW is equivalent to about 8.8 TWh a year). And the surplus is concentrated in the Northeast, while most of the data center requests are in the Southeast. Brazil's bottleneck, like America's, is the grid and location, not generation.

The debt that links Ceará to AI

The most direct link between the AI cycle and long-term debt in Brazil lies in power generation. In June 2026, Casa dos Ventos completed a $1.1 billion private placement in the United States, in two issues: $825 million with a 24-year term and $252 million with a 17-year term. The plants financed "are part of long-term energy sale contracts, the PPAs, signed with companies such as Ascenty, Omnia and ByteDance, owner of TikTok." The debt is in dollars because "the energy sale contracts are also in American currency", which, according to the company, "eliminates the currency risk of the transaction" [249]. The currency risk goes away, but counterparty risk remains: debt of 17 to 24 years, backed by contracts with buyers whose business depends on equipment depreciated over five or six years. It is the same mismatch as in the American nuclear contracts. On the other side are, according to the company, "23 institutional investors from various segments, such as insurers, asset managers and pension funds" [249].

Investors and the Central Bank

Brazilian investors are exposed too. According to B3, the outstanding balance of BDRs, the receipts for foreign shares traded on the Brazilian exchange, fell from R$34 billion in 2021 to R$18 billion in 2022 and reached R$50 billion in 2025, with about 1 million investors. NVIDIA (NVDC34) was the third most traded BDR in the period covered by the report, with R$116 million a day [250]. The 2022 drop coincided with the fall in technology stocks in the United States [46], but the exchange rate and redemptions may also have played a part, and I found no study that separates each factor. Funds aimed at the general public may invest up to 20% of their net assets in assets abroad, and those for qualified investors, up to 40% [251]. Closed pension funds may hold up to 10% abroad plus up to another 10% in BDRs and international ETFs traded on B3 [252]. These are ceilings, not actual exposure, and I found no consolidated data on the real allocation.

The Central Bank of Brazil is following the issue. In its Financial Stability Report of May 2026, it wrote that "high valuations and greater concentration of investment in certain segments increase markets' sensitivity to disappointed expectations, with a risk of sharper price corrections". On the American private credit that finances data centers, it pointed to "significant risks associated with concentration and opacity". And it noted that emerging economies "remain exposed to currency pressures and capital outflows". But it concluded that, "Given the reduced foreign-exchange exposure and the low dependence on external funding, the exposure of the SFN to financial fluctuations originating abroad is low." [253].

In a correction, Brazil would feel the AI cycle indirectly: through capital outflows and the exchange rate, through announced projects that would shrink or never get off the drawing board, through generators whose debt depends on contracts with data centers, and through retail investors in BDRs, who already saw the balance fall by almost half in 2022. Redata's estimated fiscal cost, R$7.25 billion over three years, is small next to the announced amounts. The country does not carry the cycle's debt. It carries part of the narrative and part of the energy.

What could disprove the thesis

An argument that cannot be refuted is not analysis. This section brings together the best arguments against the idea that the AI cycle has the shape of a credit bubble, with the data that support them and what they do not prove.

Demand exists and is growing

The strongest argument is revenue. Microsoft reported that Azure surpassed $100 billion in fiscal year 2026, up 41%, and that "Customer demand continues to exceed available capacity" [13]. Google Cloud revenue grew 82% in the second quarter of 2026, to $24.8 billion, and its backlog reached $514 billion [254], [14]. AWS grew 37% in the same quarter, to an annual run rate of $169 billion, and said its AI business had passed $25 billion a year [25]. At Meta, revenue rose 28% and Mark Zuckerberg attributed the result to AI: "AI is accelerating our core business today" [16]. On the usage side, the Gemini app reached 950 million monthly users [254], and OpenAI speaks of "our collective 1.2B weekly users" [255].

What this does not prove is that revenue covers the investment, or that demand is independent of the financing. In January 2026, Microsoft reported that "Approximately 45% of our commercial RPO balance is from OpenAI" [84]. At Amazon, the average term of long-term contracts not yet recognized rose from 4.0 to 6.4 years in one year, with the expanded contracts with OpenAI and Anthropic [7]. Microsoft itself offers the counterpoint: in July 2026, it said that all of the growth in its backlog in the previous quarter came from customers outside the frontier labs and that "nearly 90%" of the year's cloud revenue came from these other customers [13]. That is the most important data point to watch. If demand from those who are not AI labs keeps growing at this pace, the circle described in this article weighs less than it seems.

The balance sheets are strong

The four largest together held about $533 billion in cash and short-term investments in June 2026: $76.8 billion at Microsoft, $242.5 billion at Alphabet, $123.0 billion at Amazon and $90.3 billion at Meta, by my calculation from the balance sheets filed with the SEC [3], [5], [7], [9]. The ECB and the BIS pointed out that, unlike in 2000, the rise in stocks comes with profits: according to the BIS, this is "dissimilar to the dotcom bubble of the late 1990s, which was largely fuelled by over-optimistic expectations that were not underpinned by realised earnings growth." [211], [215].

What this does not prove is that the cash comes from operations. Alphabet raised a net $49.6 billion in equity in June 2026, "including capital expenditures to scale AI infrastructure", issued $56.2 billion in debt in the first half and recorded a gain of $98.0 billion, "primarily the result of net unrealized gains on our equity securities" [254]. The chief financial officer summed up the company's debt: "a year ago, we were at about $16 billion; and we went to about $100 billion across multiple currencies and geographies." [14]. At Meta, free cash flow in the second quarter of 2026 was $784 million, and its cash of $90.3 billion was already almost equal to its long-term debt of $83.7 billion [16]. The balance sheets remain strong, but the cushion has shrunk fast.

The debt is still small relative to the system

The Bank of England wrote that financing predominantly with cash and equity "limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term" [166]. The ECB noted that corporate debt growth, in aggregate, is still low [212], and in November 2025 it calculated that the technology sector made up 1.8% to 8.3% of the main corporate bond indices [215]. The section on 2008 showed another point in favor: private credit is funded mainly with long-term capital, without the dependence on very short-term money that brought down the vehicles of 2007.

What this does not prove is that the risk has been measured. Debt statistics do not include leases that have not yet commenced, which exceed $1 trillion at the largest companies, nor the guarantees and purchase commitments described in the sections "The circle" and "Off the balance sheet". And speed matters as much as the stock: AI-related issuance in 2026 had already more than doubled that of 2025 by September [213].

Productivity will show up

Task-level studies show real gains. In a study of 5,172 support agents, published in the Quarterly Journal of Economics, access to an AI assistant "increases worker productivity, as measured by issues resolved per hour, by 15% on average" [256]. Researchers at the St. Louis Fed estimated that generative AI users saved on average 5.4% of their work hours, which would correspond to an increase of about 1.1% in aggregate productivity [257]. Adoption by businesses is also growing: in the U.S. Census survey that closed on October 4, 2026, 23.8% of businesses said they had used AI in the previous two weeks, compared with 17.3% in November 2025, with the same question already in use [258]. The BIS sums up the gap between the micro and the macro: task-level studies report gains "between 20 and 50% in time savings", but aggregate productivity estimates "tend to be more conservative at less than 1% over a long horizon" [122].

What this does not prove is that those who finance the infrastructure will make money. That is the lesson of the railways and of fiber: the technology can transform the economy and the first investor can still lose money. Satya Nadella, Microsoft's chief executive, seems to agree with the parallel. Asked in July 2026 about the risk of overcapacity, he cited precisely the year of the American railway panic: "All of us are reading this 1873 as the book to be read." The chief financial officer, Amy Hood, gave the practical answer: capex has shifted to short-lived assets, and "if the demand environment changes, you just slow down what is, in fact, the largest component and the driver of COGS." [13]. It is a good argument for Microsoft. But the flexibility of those who buy GPUs is the risk of those who financed them and of those who built the building for them.

The price of tokens can create its own demand

The cost of using AI is falling very fast. NVIDIA says that Blackwell's cost per token fell fivefold after launch, through software alone [132], and the IEA calculates that energy per AI task has been falling "by at least an order of magnitude annually" [22]. If demand is elastic enough, each price drop multiplies usage, and the installed capacity is absorbed. That is the bet of those who build. The problem is timing: if demand takes longer than the debt, whoever financed it goes bust before the capacity fills up, as happened in telecom.

What I would consider proof to the contrary

Based on these arguments, these are the signs that would make me abandon the thesis:

• revenue from AI products, not from cloud in general, covering a growing share of capex for several consecutive quarters;
• OpenAI reaching positive cash flow earlier than projected in the documents reported in the press;
• rental prices for previous-generation GPUs stable for more than three years, consistent with useful lives of six years;
• credit spreads for Oracle and the neoclouds falling while they refinance the debt maturing in 2027 and 2028;
• hyperscaler backlog growth sustained by customers that are not AI labs.

If these signs appear together, the AI cycle will have been expensive, but not a credit bubble. As of October 2026, the picture was mixed. At Microsoft, annualized AI revenue was growing 123% a year [20], faster than capex, which rose about 80% in fiscal year 2026, and in the last quarter the backlog grew only with customers from outside the labs. The rental price of H100s, chips from 2022, rose again in 2026, another point in favor of durability, as the section on depreciation shows. The spreads of Oracle and the neoclouds, on the other hand, rose, and OpenAI was still burning cash.

Scenarios and warning signs

Bubbles do not set a date to burst. Greenspan spoke of irrational exuberance in 1996, and the stock market rose for three more years; the BIS statistical test, as the BIS itself warns, does not say when the bubble ends. That is why this section does not predict a date. It describes what a correction would look like, based on two rehearsals that have already happened, and lists the indicators that show which direction the cycle is heading in.

Two rehearsals

The first came from technology. In late December 2024, China's DeepSeek published the technical report on its V3 model and put the cost of the final training run at $5.576 million, assuming a rental price of $2 per H800 GPU hour. The report itself warned that the figure excluded "the costs associated with prior research and ablation experiments on architectures, algorithms, or data" [259]. On January 27, 2025, the market reacted: "Nvidia lost close to $600 billion in market cap on Monday, the biggest drop for any company on a single day in U.S. history." Broadcom fell 17% and Oracle, 14%. Analysts at Cantor spoke of "great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs." [260]. The recovery was quick: in October 2025, NVIDIA became the first company to surpass $5 trillion in market value [261]. The episode showed that a change in efficiency can reprice installed capacity, and those who sell it, overnight.

The second came from finance. In July 2026, as the section on macroeconomics shows, AI-linked stocks fell sharply, a leveraged fund collapsed and hyperscaler shares began trading at lower multiples, but there was no contagion to the rest of the market. Both rehearsals were stock market events. Neither reached credit. The question is what would make it get there.

Three scenarios

In the first scenario, demand catches up with capacity. Revenue from AI products grows to the point of covering a significant part of the investment, GPU rental prices hold firm even after the shortage ends, and companies refinance the debt maturing in 2027 and 2028 without difficulty. The losses fall on a few neoclouds and on investors who came in late. This is the scenario the companies describe, and the cloud data in the previous section lend it some support.

In the second, the correction is in the stock market, as in 2000. AI monetization disappoints, stocks fall, and the hyperscalers do what Microsoft's chief financial officer described: they slow down purchases of short-lived assets [13]. Chip and memory suppliers lose orders, the most indebted neoclouds are restructured, and growth slows. The Bank of England simulated such a case, with a 45% fall in U.S. stocks and a 350-basis-point rise in spreads [1]. The IMF calculated that an unwinding of AI investment could reduce global GDP, investment and consumption "by as much as 0.4, 1.3, and 0.6 percent, respectively, by 2028" [97]. And the BIS points out that the end of previous investment booms brought, on average, a slowdown of more than one percentage point in GDP [208].

In the third, the correction reaches credit. A large lab fails to raise its next round and renegotiates its commitments. Oracle and the neoclouds face refinancing at a bad time: CoreWeave, for example, had $4.4 billion in debt maturing in the rest of 2026 and $6.2 billion in 2027 [93]. Lease and residual value guarantees start being called. Funds that mark data center bonds to market cut their prices, and semi-liquid private credit funds, where redemption requests already exceeded inflows in the first quarter of 2026, face more outflows [70]. This is the scenario in which wrong-way risk appears: whoever guarantees the debt loses value at the same time as the guarantee is called.

I do not assign probabilities to these scenarios. The argument of this article is more modest and more verifiable: the financing structure built in 2025 and 2026 makes the third scenario possible, in a way that was not possible when the cycle was paid for with cash and equity.

Who loses first

Given the capital structure described in the previous sections, losses would follow an order. First, the shareholders of the labs and the neoclouds, including the individual investors who came in through rounds open to retail. Then, the lenders to the neoclouds, whose loans have GPUs as collateral. Next, the suppliers that gave guarantees or bought stakes in customers, such as NVIDIA, AMD and Google, which lose revenue and investment value at the same time. After them, the holders of the bonds of data center vehicles, if the leases and residual value guarantees do not cover the debt, among them insurers and fixed-income funds. Last, the hyperscalers themselves, which remain obligated to pay the leases they signed. It is a likely order, not a guaranteed one: in 2007, commitments that seemed remote came back onto the balance sheet ahead of time, by contract or for reputational reasons.

What to watch

The indicators below are public and show, quarter by quarter, which scenario the cycle is in:

None of these signs, on its own, settles the question. Together, they show whether the bill is being paid by revenue or pushed down the road.

Conclusion

The question of this article was never whether artificial intelligence is real. The railways were real, and so was optical fiber, and both changed the economy after ruining a good part of those who financed them. The question is who pays the bill if the revenue takes longer than the debt.

The six markers in the section "How to recognize a bubble" give a partial answer. The real shock is present: investment by the four largest cloud companies went from $228 billion in 2024 to $376 billion in 2025 and may exceed $700 billion in 2026, and cloud and AI revenue is growing. So is credit financing the expansion: the hyperscalers accounted for more than 15% of U.S. investment-grade issuance in 2026 through May, and AI-related issuance had already passed, by September, twice the total for the whole of 2025. Innovation in liabilities is evident, from unconsolidated joint ventures to residual value guarantees, from loans with GPUs as collateral to platforms for turning chips into an investment asset, which BlackRock's chief executive compared to the beginning of the mortgage-backed securities market. Revenue that depends on its own financing has been measured: $1.4 trillion in commitments in the deals the IMF called circular, and 55.2% of the investment received by AI companies coming from other AI companies, according to the BIS. The last two markers appear only in part. The collateral ages fast on paper, with one chip generation a year and depreciation over six, but rental prices for older GPUs rose in 2026 because of the shortage. And prices are high, with the CAPE near 41 and stock market concentration at record levels, but they rest on profits that exist.

What 2008 teaches applies almost point by point. Off-balance-sheet commitments come back onto the balance sheet in a crisis, and today they add up to more than $1 trillion in leases not yet commenced, plus the guarantees from Meta, NVIDIA, AMD and Google. Triggers tied to market prices bring losses forward, and the neoclouds' loans are sized by the book value of the GPUs. Credit ratings may measure the guarantor, not the asset: Beignet's A+ bonds finance a campus whose sole tenant and guarantor is Meta. Guarantees that fall together with the risk fail when they are called, and the IMF already uses the term wrong-way risk for the AI ecosystem. And metrics watched by the market become targets for engineering: longer useful lives, a reclassification of leases that reduces reported capex without changing the investment, a new measure of capex net of customer prepayments.

The differences also need to be stated. The hyperscalers have cash and little leverage, unlike the banks of 2007; private credit does not depend on very short-term money as the commercial paper vehicles did; disclosure is better; and consolidation rules are stricter. There are not millions of indebted households at the end of the chain. If the cycle breaks, the most likely outcome is something between 2000 and 2008: a stock market correction, with credit losses concentrated in labs, neoclouds, suppliers that gave guarantees and the funds that bought their debt.

What is worrying is the direction. The cycle started out paid for with cash and equity and is shifting, fast, to debt, much of it off the balance sheets of those who use the assets. As of October 2026, the AI bill was scattered across footnotes, contingent guarantees and vehicles that no one has to consolidate. It is exactly the kind of bill that, in 2008, only showed up in full when it was already too late to split it.

References

1. [1]Bank of England, “Financial Stability Report – July 2026,” Bank of England, Jul. 7, 2026. [Online]. Available: https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026 ↩
2. [2]U.S. Securities and Exchange Commission, “Company facts: Microsoft Corporation (CIK 0000789019),” EDGAR XBRL API, Oct. 13, 2026. [Online]. Available: https://data.sec.gov/api/xbrl/companyfacts/CIK0000789019.json ↩
3. [3]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm ↩
4. [4]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Feb. 5, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm ↩
5. [5]Alphabet Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 23, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm ↩
6. [6]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Feb. 6, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm ↩
7. [7]Amazon.com, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 31, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm ↩
8. [8]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Jan. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm ↩
9. [9]Meta Platforms, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm ↩
10. [10]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2026,” U.S. Securities and Exchange Commission, Jun. 22, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm ↩
11. [11]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2023,” U.S. Securities and Exchange Commission, Jul. 27, 2023. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000095017023035122/msft-20230630.htm ↩
12. [12]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2025,” U.S. Securities and Exchange Commission, Jun. 18, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000095017025087926/orcl-20250531.htm ↩
13. [13]Microsoft Corporation, “Microsoft Fiscal Year 2026 Fourth Quarter Earnings Conference Call,” Microsoft Investor Relations, Jul. 29, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 ↩
14. [14]Alphabet Inc., “2026 Q2 Earnings Call (transcrição),” Alphabet Investor Relations, Jul. 22, 2026. [Online]. Available: https://abc.xyz/investor/events/event-details/2026/2026-Q2-Earnings-Call-2026-GgTAq7Is0z/default.aspx ↩
15. [15]A. Palmer, “Amazon hikes 2026 capex to $220 billion due to higher memory costs,” CNBC, Jul. 30, 2026. [Online]. Available: https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html ↩
16. [16]Meta Platforms, Inc., “Meta Reports Second Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Jul. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm ↩
17. [17]H. Son, “Moody's says 'unprecedented' AI spending threatens credit quality of Amazon, Meta, Alphabet and others,” CNBC, Jul. 24, 2026. [Online]. Available: https://www.cnbc.com/2026/07/24/moodys-ai-spending-credit-quality-amazon-meta-alphabet.html ↩
18. [18]Oracle Corporation, “Form 10-Q for the quarterly period ended August 31, 2026,” U.S. Securities and Exchange Commission, Sep. 11, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm ↩
19. [19]Alphabet Inc., “2025 Q4 Earnings Call (transcrição),” Alphabet Investor Relations, Feb. 4, 2026. [Online]. Available: https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx ↩
20. [20]Microsoft Corporation, “Microsoft Fiscal Year 2026 Third Quarter Earnings Conference Call,” Microsoft Investor Relations, Apr. 29, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3 ↩
21. [21]Meta Platforms, Inc., “Meta Reports First Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Apr. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028364/meta-03312026xexhibit991.htm ↩
22. [22]International Energy Agency, “Key Questions on Energy and AI,” executive summary, Apr. 16, 2026. [Online]. Available: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ↩
23. [23]U.S. Bureau of Economic Analysis, “National Income and Product Accounts: arquivos de dados anuais e trimestrais (NipaDataA.txt e NipaDataQ.txt),” version of Sep. 30, 2026. [Online]. Available: https://apps.bea.gov/national/Release/TXT/NipaDataA.txt ↩
24. [24]A. Odlyzko, “Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s,” University of Minnesota, preliminary version, Jan. 15, 2010. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf ↩
25. [25]Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Jul. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm ↩
26. [26]C. P. Kindleberger and R. Z. Aliber, “Financial Crisis: A Hardy Perennial,” ch. 1 in Manias, Panics, and Crashes: A History of Financial Crises, 5th ed. Hoboken: John Wiley & Sons, 2005 (publisher's excerpt). [Online]. Available: https://media.wiley.com/product_data/excerpt/46/04714671/0471467146.pdf ↩
27. [27]C. P. Kindleberger and R. Z. Aliber, “Anatomy of a Typical Crisis,” ch. 2 in Manias, Panics and Crashes: A History of Financial Crises, 5th ed. Basingstoke: Palgrave Macmillan, 2005 (publisher's preview). [Online]. Available: https://page-one.springer.com/pdf/preview/10.1057/9780230628045_2 ↩
28. [28]H. P. Minsky, “The Financial Instability Hypothesis,” Working Paper No. 74, The Jerome Levy Economics Institute of Bard College, May 1992. [Online]. Available: https://www.levyinstitute.org/pubs/wp74.pdf ↩
29. [29]C. Perez, “Great Surges of development and alternative forms of globalization,” Working Papers in Technology Governance and Economic Dynamics No. 15, The Other Canon Foundation and Tallinn University of Technology, Jul. 2007. [Online]. Available: https://hum.ttu.ee/wp/paper15.pdf ↩
30. [30]C. Perez, “The double bubble at the turn of the Century: Technological roots and structural implications,” CFAP/CERF Working Paper No. 31, University of Cambridge, May 10, 2009; published in Cambridge Journal of Economics, vol. 33, no. 4, pp. 779–805, 2009. [Online]. Available: https://cp.informaticaprojects.com/wp-content/downloads/publications/theoretical-framework/C.PEREZ_CJE_DoubleBubble_WP%20CERF%20No%2031.pdf ↩
31. [31]International Monetary Fund, “Chapter 1,” in Global Financial Stability Report: Global Financial Markets Confront the War in the Middle East and Amplification Risks, Apr. 14, 2026. [Online]. Available: https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf ↩
32. [32]R. J. Shiller, “Speculative Asset Prices,” Prize Lecture, Dec. 8, 2013. [Online]. Available: https://www.nobelprize.org/uploads/2018/06/shiller-lecture.pdf ↩
33. [33]R. J. Shiller, “U.S. Stock Markets 1871-Present and CAPE Ratio,” spreadsheet ie_data.xls, updated Oct. 5, 2026. [Online]. Available: https://shillerdata.com/ ↩
34. [34]A. Greenspan, “The Challenge of Central Banking in a Democratic Society,” speech at the American Enterprise Institute, Board of Governors of the Federal Reserve System, Dec. 5, 1996. [Online]. Available: https://www.federalreserve.gov/boarddocs/speeches/1996/19961205.htm ↩
35. [35]A. Odlyzko, “This time is different: An example of a giant, wildly speculative, and successful investment mania,” The B.E. Journal of Economic Analysis & Policy, vol. 10, no. 1, art. 60, 2010. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/mania01.pdf ↩
36. [36]A. Odlyzko, “The Railway Mania: Fraud, disappointed expectations, and the modern economy,” Journal of the Railway & Canal Historical Society, no. 215, pp. 2–12, Nov. 2012 (revised version of Aug. 14, 2012). [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/mania06.pdf ↩
37. [37]C. P. Kindleberger, “Bubbles in History,” in The New Palgrave Dictionary of Economics, 2nd ed. Basingstoke: Palgrave Macmillan, 2008 (publisher's preview). [Online]. Available: https://page-one.springer.com/pdf/preview/10.1057/978-1-349-95121-5_3073-1 ↩
38. [38]G. Richardson and T. Sablik, “Banking Panics of the Gilded Age,” Federal Reserve History, Federal Reserve System, Dec. 4, 2015. [Online]. Available: https://www.federalreservehistory.org/essays/banking-panics-of-the-gilded-age ↩
39. [39]K. Giesecke, F. A. Longstaff, S. Schaefer and I. Strebulaev, “Corporate Bond Default Risk: A 150-Year Perspective,” NBER Working Paper no. 15848, Mar. 2010 (© 2010 by Kay Giesecke, Francis A. Longstaff, Stephen Schaefer, and Ilya Strebulaev). [Online]. Available: https://www.nber.org/system/files/working_papers/w15848/w15848.pdf ↩
40. [40]National Bureau of Economic Research, “US Business Cycle Expansions and Contractions,” data updated Mar. 14, 2023. [Online]. Available: https://www.nber.org/research/data/us-business-cycle-expansions-and-contractions ↩
41. [41]S. Mixon, “The Crisis of 1873: Perspectives from Multiple Asset Classes,” The Journal of Economic History, vol. 68, no. 3, pp. 722–757, Sep. 2008. [Online]. Available: https://doi.org/10.1017/S0022050708000624 ↩
42. [42]E. A. Couper, J. P. Hejkal and A. L. Wolman, “Boom and Bust in Telecommunications,” Economic Quarterly, Federal Reserve Bank of Richmond, vol. 89, no. 4, fall 2003. [Online]. Available: https://www.richmondfed.org/-/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf ↩
43. [43]A. M. Odlyzko, “Internet traffic growth: Sources and implications,” in Optical Transmission Systems and Equipment for WDM Networking II, Proc. SPIE, vol. 5247, pp. 1–15, 2003. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/itcom.internet.growth.pdf ↩
44. [44]J. M. Kraushaar, “Fiber Deployment Update — End of Year 1998,” Industry Analysis Division, Common Carrier Bureau, Federal Communications Commission, Sep. 9, 1999, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20000817010609/http://www.fcc.gov:80/Bureaus/Common_Carrier/Reports/FCC-State_Link/Fiber/fiber98.pdf ↩
45. [45]M. Doms, “The Boom and Bust in Information Technology Investment,” FRBSF Economic Review, Federal Reserve Bank of San Francisco, pp. 19–34, 2004. [Online]. Available: https://www.frbsf.org/wp-content/uploads/sites/4/er19-34bk.pdf ↩
46. [46]NASDAQ OMX Group, “NASDAQ Composite Index [NASDAQCOM],” FRED, Federal Reserve Bank of St. Louis, daily series accessed Oct. 2026. [Online]. Available: https://fred.stlouisfed.org/series/NASDAQCOM ↩
47. [47]Bank for International Settlements, 72nd Annual Report, 1 April 2001–31 March 2002, Basileia, Jul. 8, 2002. [Online]. Available: https://www.bis.org/publications/aer-2002.pdf ↩
48. [48]Lucent Technologies Inc., “Form 10-K405 for the fiscal year ended September 30, 2001,” U.S. Securities and Exchange Commission, Dec. 28, 2001. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1006240/000095011701501896/a31752.txt ↩
49. [49]Nortel Networks Corporation, “Form 10-K for the fiscal year ended December 31, 2001,” U.S. Securities and Exchange Commission, Mar. 11, 2002. [Online]. Available: https://www.sec.gov/Archives/edgar/data/72911/000113031902000168/t06646e10-k.htm ↩
50. [50]Cisco Systems, Inc., “Form 10-K for the fiscal year ended July 28, 2001,” U.S. Securities and Exchange Commission, Sep. 24, 2001. [Online]. Available: https://www.sec.gov/Archives/edgar/data/858877/000109581101505065/f75710e10-k.txt ↩
51. [51]J. Frost, R. Kansal, K. Rishabh, V. Shreeti and L. S. Y. Zhang, “Circular relationships among AI firms,” BIS Bulletin, no. 137, Bank for International Settlements, Oct. 1, 2026. [Online]. Available: https://www.bis.org/publications/bulletin-137-circular-relationships-among-ai-firms.pdf ↩
52. [52]U.S. Securities and Exchange Commission, “SEC Charges Qwest Communications International Inc. with Multi-Faceted Accounting and Financial Reporting Fraud,” Litigation Release no. 18936, Oct. 21, 2004. [Online]. Available: https://www.sec.gov/litigation/litreleases/lr18936.htm ↩
53. [53]U.S. Securities and Exchange Commission, “In the Matter of Global Crossing Ltd., Thomas J. Casey, Dan J. Cohrs, and Joseph P. Perrone,” Exchange Act Release no. 51517, Apr. 11, 2005. [Online]. Available: https://www.sec.gov/files/litigation/admin/34-51517.pdf ↩
54. [54]U.S. Securities and Exchange Commission, “SEC Charges WorldCom with $3.8 Billion Fraud,” Litigation Release no. 17588, Jun. 27, 2002. [Online]. Available: https://www.sec.gov/litigation/litreleases/lr17588.htm ↩
55. [55]WorldCom, Inc., “Form 10-K for the fiscal year ended December 31, 2002,” U.S. Securities and Exchange Commission, Mar. 12, 2004. [Online]. Available: https://www.sec.gov/Archives/edgar/data/723527/000119312504039709/d10k.htm ↩
56. [56]W. C. Powers Jr., R. S. Troubh and H. S. Winokur Jr., “Report of Investigation by the Special Investigative Committee of the Board of Directors of Enron Corp.,” Feb. 1, 2002, Exhibit 99.2 to Enron Corp.'s Form 8-K of Feb. 7, 2002. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1024401/000090951802000089/big.txt ↩
57. [57]U.S. Securities and Exchange Commission, Office of the Chief Accountant, Office of Economic Analysis and Division of Corporation Finance, “Report and Recommendations Pursuant to Section 401(c) of the Sarbanes-Oxley Act of 2002 On Arrangements with Off-Balance Sheet Implications, Special Purpose Entities, and Transparency of Filings by Issuers,” Jun. 2005. [Online]. Available: https://www.sec.gov/news/studies/soxoffbalancerpt.pdf ↩
58. [58]Financial Accounting Standards Board, “Summary of Interpretation No. 46: Consolidation of Variable Interest Entities—an interpretation of ARB No. 51,” Jan. 2003, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20040611063520/http://www.fasb.org:80/st/summary/finsum46.shtml ↩
59. [59]Financial Accounting Standards Board, “Summary of Interpretation No. 46 (revised December 2003),” Dec. 2003, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20040611052814/http://www.fasb.org:80/st/summary/finsum46r.shtml ↩
60. [60]Financial Crisis Inquiry Commission, The Financial Crisis Inquiry Report: Final Report of the National Commission on the Causes of the Financial and Economic Crisis in the United States. Washington: U.S. Government Publishing Office, Jan. 2011. [Online]. Available: https://www.govinfo.gov/content/pkg/GPO-FCIC/pdf/GPO-FCIC.pdf ↩
61. [61]Board of Governors of the Federal Reserve System, “Asset-Backed Commercial Paper Outstanding [ABCOMP],” FRED, Federal Reserve Bank of St. Louis, weekly series accessed Oct. 13, 2026. [Online]. Available: https://fred.stlouisfed.org/series/ABCOMP ↩
62. [62]D. M. Covitz, N. Liang and G. A. Suarez, “The Evolution of a Financial Crisis: Panic in the Asset-Backed Commercial Paper Market,” Finance and Economics Discussion Series 2009-36, Federal Reserve Board, Aug. 18, 2009. [Online]. Available: https://www.federalreserve.gov/pubs/feds/2009/200936/200936pap.pdf ↩
63. [63]Citigroup Inc., “Citi Commits Support Facility for Citi-Advised SIVs,” press release (Exhibit 99.1 to Form 8-K), Dec. 13, 2007. [Online]. Available: https://www.sec.gov/Archives/edgar/data/831001/000114420407067635/v097312_ex99-1.htm ↩
64. [64]A. R. Valukas, “Report of Anton R. Valukas, Examiner,” vol. 1, In re Lehman Brothers Holdings Inc., Chapter 11 Case No. 08-13555, U.S. Bankruptcy Court, S.D.N.Y., Mar. 11, 2010, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20100331000019/http://lehmanreport.jenner.com:80/VOLUME%201.pdf ↩
65. [65]Basel Committee on Banking Supervision, “CRE50 – Counterparty credit risk definitions and terminology,” Basel Framework, version published Jul. 5, 2024. [Online]. Available: https://www.bis.org/basel_framework/chapter/CRE/50.htm ↩
66. [66]Basel Committee on Banking Supervision, “Strengthening the resilience of the banking sector,” documento consultivo, Dec. 2009. [Online]. Available: https://www.bis.org/publ/bcbs164.pdf ↩
67. [67]Financial Accounting Standards Board, “Accounting Standards Update No. 2009-17, Consolidations (Topic 810): Improvements to Financial Reporting by Enterprises Involved with Variable Interest Entities,” Dec. 2009. [Online]. Available: https://storage.fasb.org/ASU2009-17.pdf ↩
68. [68]Board of Governors of the Federal Reserve System, “H.8 Assets and Liabilities of Commercial Banks in the United States,” Apr. 9, 2010. [Online]. Available: https://www.federalreserve.gov/releases/h8/20100409/ ↩
69. [69]International Monetary Fund, “Chapter 2: The Rise and Risks of Private Credit,” in Global Financial Stability Report: The Last Mile: Financial Vulnerabilities and Risks, Apr. 16, 2024. [Online]. Available: https://www.imf.org/-/media/Files/Publications/GFSR/2024/April/English/ch2.ashx ↩
70. [70]Board of Governors of the Federal Reserve System, Financial Stability Report, May 2026. [Online]. Available: https://www.federalreserve.gov/publications/files/financial-stability-report-20260508.pdf ↩
71. [71]P. Abbassi, I. Aldasoro and S. Doerr, “Financing the digital economy: the role of private credit,” BIS Quarterly Review, Bank for International Settlements, Sep. 14, 2026. [Online]. Available: https://www.bis.org/publications/financing-digital-economy-role-private-credit.pdf ↩
72. [72]Financial Stability Board, “Report on Vulnerabilities in Private Credit,” May 6, 2026. [Online]. Available: https://www.fsb.org/uploads/P060526.pdf ↩
73. [73]NVIDIA, “OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems,” NVIDIA Newsroom, Sep. 22, 2025. [Online]. Available: https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems ↩
74. [74]NVIDIA Corporation, “Form 10-Q — quarterly report for the period ended October 26, 2025,” U.S. Securities and Exchange Commission (EDGAR), Nov. 19, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581025000230/nvda-20251026.htm ↩
75. [75]OpenAI, “Scaling AI for everyone,” OpenAI, Feb. 27, 2026. [Online]. Available: https://openai.com/index/scaling-ai-for-everyone/ ↩
76. [76]A. Capoot, “Nvidia CEO Huang says $30 billion OpenAI investment 'might be the last',” CNBC, Mar. 4, 2026. [Online]. Available: https://www.cnbc.com/2026/03/04/nvidia-huang-openai-investment.html ↩
77. [77]NVIDIA Corporation, “Form 10-Q for the quarterly period ended July 26, 2026,” U.S. Securities and Exchange Commission, Aug. 26, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm ↩
78. [78]NVIDIA Corporation, “Form 10-K — annual report for the fiscal year ended January 25, 2026,” U.S. Securities and Exchange Commission (EDGAR), Feb. 25, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm ↩
79. [79]Advanced Micro Devices, Inc., “Form 8-K” and Exhibit 99.1, “AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs,” U.S. Securities and Exchange Commission, Oct. 6, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000119312525230895/d28189d8k.htm ↩
80. [80]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended September 27, 2025,” U.S. Securities and Exchange Commission, Nov. 5, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248825000166/amd-20250927.htm ↩
81. [81]Advanced Micro Devices, Inc., “Form 8-K” (warrant issued to Meta Platforms), U.S. Securities and Exchange Commission, Feb. 24, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248826000045/amd-20260223.htm ↩
82. [82]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended June 27, 2026,” U.S. Securities and Exchange Commission, Aug. 5, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248826000123/amd-20260627.htm ↩
83. [83]Microsoft, “The next chapter of the Microsoft–OpenAI partnership,” Official Microsoft Blog, Oct. 28, 2025. [Online]. Available: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/ ↩
84. [84]Microsoft, “Microsoft Fiscal Year 2026 Second Quarter Earnings Conference Call,” Microsoft Investor Relations, Jan. 28, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q2 ↩
85. [85]OpenAI, “OpenAI and Amazon announce strategic partnership,” OpenAI, Feb. 27, 2026. [Online]. Available: https://openai.com/index/amazon-partnership/ ↩
86. [86]Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute,” Anthropic, Apr. 20, 2026. [Online]. Available: https://www.anthropic.com/news/anthropic-amazon-compute ↩
87. [87]Anthropic, “Anthropic raises $65B in Series H funding at $965B post-money valuation,” Anthropic, May 28, 2026. [Online]. Available: https://www.anthropic.com/news/series-h ↩
88. [88]Anthropic, “Microsoft, NVIDIA, and Anthropic announce strategic partnerships,” Anthropic, Nov. 18, 2025. [Online]. Available: https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships ↩
89. [89]Anthropic, “Expanding our use of Google Cloud TPUs and Services,” Anthropic, Oct. 23, 2025. [Online]. Available: https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services ↩
90. [90]CoreWeave, Inc., “CoreWeave Announces Agreement with OpenAI to Deliver AI Infrastructure,” CoreWeave Investor Relations, Mar. 10, 2025. [Online]. Available: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Announces-Agreement-with-OpenAI-to-Deliver-AI-Infrastructure/default.aspx ↩
91. [91]CoreWeave, Inc., “CoreWeave Expands Agreement with OpenAI by up to $6.5B,” CoreWeave Investor Relations, Sep. 25, 2025. [Online]. Available: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Expands-Agreement-with-OpenAI-by-up-to-6-5B/default.aspx ↩
92. [92]CoreWeave, Inc., “Current Report (Form 8-K) — Order Form with NVIDIA Corporation,” U.S. Securities and Exchange Commission (EDGAR), Sep. 15, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm ↩
93. [93]CoreWeave, Inc., “Quarterly Report (Form 10-Q) for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission (EDGAR), Aug. 12, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm ↩
94. [94]OpenAI, “OpenAI and SoftBank Group partner with SB Energy,” OpenAI, Jan. 9, 2026. [Online]. Available: https://openai.com/index/stargate-sb-energy-partnership/ ↩
95. [95]J. Bort, “Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments,” TechCrunch, Nov. 6, 2025. [Online]. Available: https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/ ↩
96. [96]A. Capoot and K. Rooney, “Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report,” CNBC, Oct. 8, 2026. [Online]. Available: https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html ↩
97. [97]International Monetary Fund, “Global Financial Stability Report, October 2026: Resilience, Risk, and Transformation,” IMF, Oct. 12, 2026. [Online]. Available: https://doi.org/10.5089/9798229055673.082 ↩
98. [98]A. Bao, “IMF warns bond yields could see a 'disorderly' spike as debt, inflation risks worsen,” CNBC, Oct. 13, 2026. [Online]. Available: https://www.cnbc.com/2026/10/13/treasury-yields-imf-debt-inflation-.html ↩
99. [99]OpenAI and Broadcom, “OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators,” OpenAI, Oct. 13, 2025. [Online]. Available: https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/ ↩
100. [100]Meta Platforms, “Meta Announces Joint Venture With Funds Managed by Blue Owl Capital to Develop Hyperion Data Center,” Meta Newsroom, Oct. 21, 2025. [Online]. Available: https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/ ↩
101. [101]C. Arroyo and L. Benitez, “Meta, Blue Owl Seal $30 Billion Private Capital Deal for AI,” Bloomberg, Oct. 16, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-16/blue-owl-seals-largest-private-capital-deal-for-meta-s-ai-growth ↩
102. [102]Blue Owl Real Estate Net Lease Trust, “Form 10-Q: quarterly report for the period ended September 30, 2025,” U.S. Securities and Exchange Commission (EDGAR), Nov. 7, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1944366/000194436625000126/osnl-20250930.htm ↩
103. [103]Meta Platforms, Inc., “Form 10-Q: quarterly report for the period ended March 31, 2026,” U.S. Securities and Exchange Commission (EDGAR), Apr. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm ↩
104. [104]Deloitte, “Roadmap: Consolidation — 7.1 Introduction (Determining the Primary Beneficiary),” DART – Deloitte Accounting Research Tool, Sep. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc810-10/roadmap-consolidation/chapter-7-determining-primary-beneficiary/7-1-introduction ↩
105. [105]J. Weil, “AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center,” The Wall Street Journal, Nov. 24, 2025. [Online]. Available: https://www.wsj.com/tech/meta-ai-data-center-finances-d3a6b464 ↩
106. [106]Deloitte, “Roadmap: Leases — 6.7 Amounts That It Is Probable That the Lessee Will Owe Under a Residual Value Guarantee,” DART – Deloitte Accounting Research Tool, Aug. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-6-lease-payments/6-7-amounts-that-it-is ↩
107. [107]Deloitte, “Roadmap: Leases — 8.4 Recognition and Measurement,” DART – Deloitte Accounting Research Tool, Aug. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-8-lessee-accounting/8-4-recognition-measurement ↩
108. [108]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q2_nport.zip,” SEC Division of Economic and Risk Analysis, Jul. 8, 2026. [Online]. Available: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q2_nport.zip ↩
109. [109]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q1_nport.zip,” SEC Division of Economic and Risk Analysis, Apr. 3, 2026. [Online]. Available: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q1_nport.zip ↩
110. [110]C. Arroyo and E. Ludlow, “xAI to Raise $20 Billion After Nvidia and Others Boost Round,” Bloomberg, Oct. 7, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-07/musk-s-xai-nears-20-billion-capital-raise-tied-to-nvidia-chips ↩
111. [111]C. Arroyo, “How AI Companies Are Keeping Debt Off Their Balance Sheets,” Bloomberg, Oct. 31, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-31/meta-xai-starting-trend-for-billions-in-off-balance-sheet-debt ↩
112. [112]Apollo Global Management, “Apollo Backs $5.4 Billion Valor and xAI Data Center Compute Infrastructure Transaction with $3.5 Billion Capital Solution,” Apollo (GlobeNewswire), Jan. 7, 2026. [Online]. Available: https://www.apollo.com/insights-news/pressreleases/2026/01/apollo-backs-5-4-billion-valor-and-xai-data-center-compute-infrastructure-transaction-with-3-5-billion-capital-solution-3214463 ↩
113. [113]NVIDIA Corporation, “Form 8-K (Items 1.01, 2.03 and 7.01),” U.S. Securities and Exchange Commission, Aug. 17, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm ↩
114. [114]A. Capoot and K. Rooney, “Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans,” CNBC, Jul. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html ↩
115. [115]NVIDIA, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” NVIDIA Newsroom, Aug. 17, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-guarantees-sb-energy-s-ports-pike-technology-campus-in-ohio-to-exclusively-host-nvidia-ai-compute ↩
116. [116]TeraWulf Inc., “TeraWulf Announces Fluidstack Expansion with 160 MW CB-5 Lease at Lake Mariner,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Aug. 18, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1083301/000110465925079463/tm2523651d3_ex99-1.htm ↩
117. [117]Cipher Mining Inc., “Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Sep. 25, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1819989/000095010325012168/dp234624_ex9901.htm ↩
118. [118]Hut 8 Corp., “Hut 8 Signs 15-Year, 245 MW AI Data Center Lease at River Bend Campus with Total Contract Value of $7.0 Billion,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Dec. 17, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1964789/000110465925122052/hut-20251217xex99d1.htm ↩
119. [119]NVIDIA, “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital,” NVIDIA Newsroom, Aug. 10, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital ↩
120. [120]K. Leswing, “Wall Street just endorsed Jensen Huang's ‘big concept’ for AI. What now?,” CNBC, Aug. 11, 2026. [Online]. Available: https://www.cnbc.com/2026/08/11/wall-street-endorsed-jensen-huangs-big-concept-for-ai-what-now.html ↩
121. [121]Bank for International Settlements, “Markets recalibrate amid shifting currents,” BIS Quarterly Review, Mar. 16, 2026. [Online]. Available: https://www.bis.org/publ/qtrpdf/r_qt2603a.htm ↩
122. [122]Bank for International Settlements, “I. Progress and peril,” BIS Annual Economic Report 2026, Jun. 28, 2026. [Online]. Available: https://www.bis.org/publications/aer-2026/progress-peril ↩
123. [123]D. Gautam, J. Kramer, H. Kraus and A. Yokoyama, “Background Analysis 1.1. Rising Leverage in the AI Ecosystem,” Global Financial Stability Report, October 2026 (IMF), Oct. 12, 2026. [Online]. Available: https://www.elibrary.imf.org/display/book/9798229055673/9798229055673.xml ↩
124. [124]CoreWeave, Inc., “Form S-1 Registration Statement,” U.S. Securities and Exchange Commission, Mar. 3, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm ↩
125. [125]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2024,” U.S. Securities and Exchange Commission, Feb. 7, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm ↩
126. [126]B. Caulfield, “‘Accelerate Everything,’ NVIDIA CEO Says Ahead of COMPUTEX,” NVIDIA Blog, Jun. 2, 2024. [Online]. Available: https://blogs.nvidia.com/blog/computex-2024-jensen-huang/ ↩
127. [127]NVIDIA, “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing,” press release, Mar. 18, 2024. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing ↩
128. [128]NVIDIA, “NVIDIA GB200 NVL72,” product page, version of Aug. 27, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/gb200-nvl72/ ↩
129. [129]NVIDIA, “NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer,” NVIDIA Technical Blog, Mar. 16, 2026. [Online]. Available: https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/ ↩
130. [130]S. Koparkar, “Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents,” NVIDIA Blog, Aug. 24, 2026. [Online]. Available: https://blogs.nvidia.com/blog/vera-rubin-nvl72-efficiency-ai-agents/ ↩
131. [131]NVIDIA, “NVIDIA H100 GPU,” product page, version of Jul. 1, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/h100/ ↩
132. [132]NVIDIA, “NVIDIA DGX B200,” product page, version of Apr. 29, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/dgx-b200/ ↩
133. [133]Llama Team, AI @ Meta, “The Llama 3 Herd of Models,” arXiv:2407.21783, Jul. 23, 2024. [Online]. Available: https://arxiv.org/abs/2407.21783 ↩
134. [134]M. Blake, M. Hsu, H. Petty and J. Huntington, “NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories,” NVIDIA Technical Blog, May 20, 2025. [Online]. Available: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/ ↩
135. [135]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2023,” U.S. Securities and Exchange Commission, Jan. 31, 2024. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm ↩
136. [136]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2022,” U.S. Securities and Exchange Commission, Jul. 28, 2022. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000156459022026876/msft-10k_20220630.htm ↩
137. [137]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2022,” U.S. Securities and Exchange Commission, Feb. 2, 2023. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000132680123000013/meta-20221231.htm ↩
138. [138]M. Burry (@michaeljburry), “Understating depreciation by extending useful life of assets artificially boosts earnings…,” X, Nov. 10, 2025. [Online]. Available: https://x.com/michaeljburry/status/1987918650104283372 ↩
139. [139]J. Sor, “Michael Burry's latest argument against hyperscalers like Meta and Oracle echoes one made by an investor who shorted Enron,” Business Insider, Nov. 11, 2025. [Online]. Available: https://www.businessinsider.com/michael-burry-ai-stocks-nvidia-meta-oracle-hyperscalers-jim-chanos-2025-11 ↩
140. [140]Y. Li, “Nvidia name-checks Michael Burry in secret memo pushing back on AI bubble allegations,” CNBC, Nov. 25, 2025. [Online]. Available: https://www.cnbc.com/2025/11/25/nvidia-pushes-back-on-charges-that-ai-investment-is-a-bubble.html ↩
141. [141]C. Yun, “Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances,” AWS News Blog, Jun. 5, 2025. [Online]. Available: https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/ ↩
142. [142]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2023-08-23),” AWS Pricing (Wayback capture), Aug. 25, 2023. [Online]. Available: https://web.archive.org/web/20230825130912/https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
143. [143]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2026-10-12),” AWS Pricing, Oct. 12, 2026. [Online]. Available: https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
144. [144]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS (Wayback capture), Jun. 9, 2025. [Online]. Available: https://web.archive.org/web/20250609034829/https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
145. [145]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS, Oct. 13, 2026. [Online]. Available: https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
146. [146]Lambda, “GPU cloud pricing,” Lambda (Wayback capture), Nov. 5, 2025. [Online]. Available: https://web.archive.org/web/20251105190524/https://lambda.ai/pricing ↩
147. [147]Lambda, “GPU cloud pricing: rent NVIDIA H100, H200, and B200,” Lambda (Wayback capture), Oct. 1, 2026. [Online]. Available: https://web.archive.org/web/20261001114850/https://lambda.ai/pricing ↩
148. [148]K. Leswing, “The question everyone in AI is asking: How long before a GPU depreciates?,” CNBC, Nov. 14, 2025. [Online]. Available: https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html ↩
149. [149]E. Lopatto, “Nvidia’s new financial strategy does not compute,” The Verge, Aug. 19, 2026. [Online]. Available: https://www.theverge.com/ai-artificial-intelligence/981668/nvidias-goldman-blackrock-gpu-compute-asset ↩
150. [150]CME Group, “CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks,” CME Group Press Release, Aug. 11, 2026. [Online]. Available: https://www.cmegroup.com/media-room/press-releases/2026/8/11/cme_group_and_silicondatatolaunchcomputefuturesonoctober5tounloc.html ↩
151. [151]S. Koparkar, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” NVIDIA Blog, Oct. 1, 2026. [Online]. Available: https://blogs.nvidia.com/blog/productive-durable-fungible-ai-factories/ ↩
152. [152]International Energy Agency, “Energy and AI,” executive summary, Apr. 10, 2025 (license CC BY 4.0). [Online]. Available: https://www.iea.org/reports/energy-and-ai/executive-summary ↩
153. [153]A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet and D. Sartor, “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, LBNL-2001637, Dec. 2024. [Online]. Available: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf ↩
154. [154]J. Rand, A. Cheyette, C. Talley, S. Zhang, W. Gorman, R. H. Wiser, J. Seel, S. Jeong and F. Kahrl, “Queued Up: 2026 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2025,” Lawrence Berkeley National Laboratory, Jun. 2026. [Online]. Available: https://emp.lbl.gov/queues ↩
155. [155]NVIDIA, “NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint With Broad Industry Support,” press release, Mar. 16, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-releases-vera-rubin-dsx-ai-factory-reference-design-and-omniverse-dsx-digital-twin-blueprint-with-broad-industry-support ↩
156. [156]GE Vernova Inc., “GE Vernova reports fourth quarter and full year 2025 financial results; increasing outlook with Prolec GE acquisition,” Exhibit 99 to Form 8-K, U.S. Securities and Exchange Commission, Jan. 28, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000012/gevpressrelease4q25.htm ↩
157. [157]GE Vernova Inc., “GE Vernova reports second quarter 2026 financial results and raises 2026 financial guidance,” Exhibit 99 to Form 8-K, U.S. Securities and Exchange Commission, Jul. 22, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000147/gevpressrelease2q26.htm ↩
158. [158]Constellation Energy, “Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to The Grid,” press release, Sep. 20, 2024. [Online]. Available: https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html ↩
159. [159]M. Terrell, “New nuclear clean energy agreement with Kairos Power,” The Keyword (Google), Oct. 14, 2024. [Online]. Available: https://blog.google/company-news/outreach-and-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/ ↩
160. [160]Constellation Energy, “Constellation, Meta Sign 20-Year Deal for Clean, Reliable Nuclear Energy in Illinois,” press release, Jun. 3, 2025. [Online]. Available: https://www.constellationenergy.com/news/2025/constellation-meta-sign-20-year-deal-for-clean-reliable-nuclear-energy-in-illinois.html ↩
161. [161]Talen Energy, “Talen Energy Expands Nuclear Energy Relationship with Amazon,” press release, Jun. 11, 2025. [Online]. Available: https://ir.talenenergy.com/news-releases/news-release-details/talen-energy-expands-nuclear-energy-relationship-amazon ↩
162. [162]Constellation Energy, “U.S. Government Backs Constellation's Plan to Launch Crane Clean Energy Center, Adding 835 MWs of New Baseload Power to the Grid,” press release, Nov. 18, 2025. [Online]. Available: https://www.constellationenergy.com/news/2025/11/us-government-backs-constellations-plan-to-launch-crane-clean-energy-center-adding-835-mws-of-new-baseload-power-to-the-grid.html ↩
163. [163]American Electric Power, “AEP Ohio, PUCO Staff, Ohio Consumers’ Counsel, Ohio Energy Group and Others Agree on How to Address Growing Data Center Power Needs,” Oct. 23, 2024. [Online]. Available: https://www.aep.com/news/stories/view/9829/aep-ohio-puco-staff-ohio-consumers-counsel-ohio-energy-group-and-others-agree-on-how-to-address-growing-data-center-power-needs/ ↩
164. [164]Ohio Power Company, “Data Center Tariff,” web page, accessed Oct. 13, 2026. [Online]. Available: https://www.aepohio.com/company/about/rates/data-center-tariff ↩
165. [165]American Electric Power Company, Inc., “Form 10-Q for the quarterly period ended June 30, 2025,” U.S. Securities and Exchange Commission, Jul. 30, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/4904/000000490425000141/aep-20250630.htm ↩
166. [166]Bank of England, “Financial Stability Report – December 2025,” Bank of England, Dec. 2, 2025. [Online]. Available: https://www.bankofengland.co.uk/financial-stability-report/2025/december-2025 ↩
167. [167]S. Friar, “A business that scales with the value of intelligence,” OpenAI, Jan. 18, 2026. [Online]. Available: https://openai.com/index/a-business-that-scales-with-the-value-of-intelligence/ ↩
168. [168]A. Capoot et al., “OpenAI closes record-breaking $122 billion funding round as anticipation builds for IPO,” CNBC, Mar. 31, 2026. [Online]. Available: https://www.cnbc.com/2026/03/31/openai-funding-round-ipo.html ↩
169. [169]OpenAI, “OpenAI raises $122 billion to accelerate the next phase of AI,” OpenAI, Mar. 31, 2026. [Online]. Available: https://openai.com/index/accelerating-the-next-phase-ai/ ↩
170. [170]S. Muppidi, “OpenAI Says Its Business Will Burn $115 Billion Through 2029,” The Information, Sep. 6, 2025. [Online]. Available: https://www.theinformation.com/articles/openai-says-business-will-burn-115-billion-2029 ↩
171. [171]S. Muppidi and S. Palazzolo, “OpenAI Boosts Revenue Forecasts, Predicts $111 Billion More Cash Burn Through 2030,” The Information, Feb. 20, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-boost-revenue-forecasts-predicts-112-billion-cash-burn-2030 ↩
172. [172]A. Gardizy and A. Efrati, “OpenAI CEO and CFO Diverge on IPO Timing,” The Information, Apr. 5, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-ceo-cfo-diverge-ipo-timing ↩
173. [173]E. Woo, “OpenAI Burned $3.7 Billion in First Three Months of 2026,” The Information, Jun. 16, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-burned-3-7-billion-first-three-months-2026 ↩
174. [174]OpenAI, “Confidential submission of draft S-1 to the SEC,” OpenAI, Jun. 8, 2026. [Online]. Available: https://openai.com/index/openai-submits-confidential-s-1/ ↩
175. [175]SoftBank Group Corp., “Follow-on Investments in OpenAI,” press release, Feb. 27, 2026. [Online]. Available: https://group.softbank/en/news/press/20260227 ↩
176. [176]SoftBank Group Corp., “Execution of Bridge Facility Agreement Primarily for the Follow-on Investments in OpenAI,” press release, Mar. 27, 2026. [Online]. Available: https://group.softbank/en/news/press/20260327 ↩
177. [177]SoftBank Group Corp., “Issuance of Foreign Currency-Denominated Senior Notes,” press release, Sep. 24, 2026. [Online]. Available: https://group.softbank/en/news/press/20260924 ↩
178. [178]CNBC, “Mega-IPOs could signal market top, say analysts as SpaceX and OpenAI prep record floats,” CNBC, May 22, 2026. [Online]. Available: https://www.cnbc.com/2026/05/22/ipo-flurry-top-market-analysts-ai-spacex-musk-altman.html ↩
179. [179]The Information, “How OpenAI’s Web of Business Relationships Could Complicate Its IPO,” The Information, Jun. 22, 2026. [Online]. Available: https://www.theinformation.com/articles/openais-light-balance-sheet-face-hard-look-ipo ↩
180. [180]CNBC, “OpenAI CFO Sarah Friar says company isn't seeking government backstop, clarifying prior comment,” CNBC, Nov. 6, 2025. [Online]. Available: https://www.cnbc.com/2025/11/06/openai-cfo-sarah-friar-says-company-is-not-seeking-government-backstop.html ↩
181. [181]A. Capoot and M. Sigalos, “Sam Altman says OpenAI will top $20 billion in annualized revenue this year, hundreds of billions by 2030,” CNBC, Nov. 6, 2025. [Online]. Available: https://www.cnbc.com/2025/11/06/sam-altman-says-openai-will-top-20-billion-annual-revenue-this-year.html ↩
182. [182]P. Singh et al., “An analyst worries OpenAI will miss its 'extravagant' spending commitments. These stocks would be the most vulnerable,” CNBC, Nov. 7, 2025. [Online]. Available: https://www.cnbc.com/2025/11/07/these-stocks-are-at-risk-if-openai-misses-its-extravagant-spending-plans.html ↩
183. [183]S. Subin, “OpenAI's revenue, growth estimates fall short as company races toward IPO: Report,” CNBC, Apr. 28, 2026. [Online]. Available: https://www.cnbc.com/2026/04/28/openais-revenue-growth-estimates-fall-short-report.html ↩
184. [184]Oracle Corporation, “Oracle Announces Fiscal Year 2026 First Quarter Financial Results,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Sep. 9, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312525199175/orcl-ex99_1.htm ↩
185. [185]U.S. Securities and Exchange Commission, “Company facts: Oracle Corporation (CIK 0001341439),” EDGAR XBRL API, Oct. 13, 2026. [Online]. Available: https://data.sec.gov/api/xbrl/companyfacts/CIK0001341439.json ↩
186. [186]Oracle Corporation, “Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Sep. 10, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-ex99_1.htm ↩
187. [187]Oracle Corporation, “Oracle Announces Record Q4 and FY 2026 Results Driven by Cloud Infrastructure & Cloud Applications,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Jun. 10, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526265848/orcl-ex99_1.htm ↩
188. [188]S. Mody, “Oracle's credit default swaps are plummeting as financing plan boosts investor confidence,” CNBC, Feb. 2, 2026. [Online]. Available: https://www.cnbc.com/2026/02/02/oracles-credit-default-swaps-are-plummeting-after-financing-plans.html ↩
189. [189]J. Novet, “Oracle beats on earnings, but stock drops on plans to raise another $20 billion,” CNBC, Jun. 10, 2026. [Online]. Available: https://www.cnbc.com/2026/06/10/oracle-orcl-q4-earnings-report-2026.html ↩
190. [190]J. Novet, “Oracle CEO Magouyrk: 'Of course' OpenAI can pay $60 billion per year for cloud infrastructure,” CNBC, Oct. 13, 2025. [Online]. Available: https://www.cnbc.com/2025/10/13/oracle-ceo-magouyrk-of-course-openai-can-pay-60-billion-per-year.html ↩
191. [191]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, Sep. 24, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312525215807/d49252dfwp.htm ↩
192. [192]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, Feb. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526033882/d44245dfwp.htm ↩
193. [193]Oracle Corporation, release of Feb. 1, 2026 on the 2026 funding plan, filed as Form FWP, U.S. Securities and Exchange Commission, Feb. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526032650/d36096dfwp.htm ↩
194. [194]J. Novet, “Oracle stock has worst week since 2001 dot-com bust as AI financing concerns escalate,” CNBC, Jun. 26, 2026. [Online]. Available: https://www.cnbc.com/2026/06/26/oracle-stock-ends-worst-week-since-2001-as-investors-dwell-on-finances.html ↩
195. [195]T. Burns, “What more expensive corporate debt could mean for the AI buildout,” CNBC, Jul. 26, 2026. [Online]. Available: https://www.cnbc.com/2026/07/26/what-more-expensive-corporate-debt-could-mean-for-the-ai-buildout.html ↩
196. [196]Reuters, “Oracle sued by bondholders over losses tied to AI buildout,” published by CNBC, Jan. 14, 2026. [Online]. Available: https://www.cnbc.com/2026/01/14/oracle-sued-by-bondholders-over-losses-tied-to-ai-buildout.html ↩
197. [197]S. Subin, “Oracle sends 'force majeure' notice about data center project — stock drops 3%,” CNBC, Sep. 24, 2026. [Online]. Available: https://www.cnbc.com/2026/09/24/oracle-data-center-force-majeure.html ↩
198. [198]CoreWeave, Inc., “CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing,” CoreWeave Investor Relations, Mar. 31, 2026. [Online]. Available: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx ↩
199. [199]CoreWeave, Inc., “CoreWeave Closes $2.6 Billion Loan Facility, Expanding Financing Flexibility for AI Infrastructure,” CoreWeave Investor Relations, Aug. 10, 2026. [Online]. Available: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-2-6-Billion-Loan-Facility-Expanding-Financing-Flexibility-for-AI-Infrastructure/default.aspx ↩
200. [200]CoreWeave, Inc., “Annual Report (Form 10-K) for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission (EDGAR), Mar. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm ↩
201. [201]Nebius Group N.V., “Unaudited Condensed Consolidated Financial Statements for the Three and Six Months Ended June 30, 2026,” Form 6-K, Exhibit 99.2 (SEC EDGAR), Aug. 12, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d2.htm ↩
202. [202]Nebius Group N.V., “Report of Foreign Private Issuer (Form 6-K) — Commercial Agreement with Microsoft,” U.S. Securities and Exchange Commission (EDGAR), Sep. 8, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm ↩
203. [203]K. Leswing and I. O'Brien, “Debt-hungry AI companies face increased risk as bond yields spike,” CNBC, Sep. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/09/27/debt-hungry-data-center-companies-increased-risk-bond-yields-spike.html ↩
204. [204]N. Lichtenberg, “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says,” Fortune, Oct. 7, 2025. [Online]. Available: https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist/ ↩
205. [205]Federal Reserve Bank of St. Louis, “ALFRED — Real GDP (GDPC1) and Real Private Fixed Investment: Information Processing Equipment and Software (A679RX1Q020SBEA), data vintage of Sep. 25, 2025,” ALFRED, Sep. 25, 2025. [Online]. Available: https://alfred.stlouisfed.org/series?seid=A679RX1Q020SBEA ↩
206. [206]U.S. Bureau of Economic Analysis, “NIPA Handbook: Concepts and Methods of the U.S. National Income and Product Accounts — Chapter 4: Estimating Methods,” BEA, Dec. 2023. [Online]. Available: https://www.bea.gov/resources/methodologies/nipa-handbook/pdf/chapter-04.pdf ↩
207. [207]International Monetary Fund, “Global Financial Stability Report, October 2025: Shifting Ground beneath the Calm,” IMF, Oct. 14, 2025. [Online]. Available: https://www.imf.org/en/Publications/GFSR/Issues/2025/10/14/global-financial-stability-report-october-2025 ↩
208. [208]I. Aldasoro, S. Doerr and D. Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin, no. 120, Jan. 7, 2026. [Online]. Available: https://www.bis.org/publications/bulletin-120-financing-ai-boom-cash-flows-debt ↩
209. [209]U.S. Census Bureau, “Monthly Construction Spending, August 2026 (Release CB26-158),” U.S. Census Bureau, Oct. 1, 2026. [Online]. Available: https://www.census.gov/construction/c30/pdf/release.pdf ↩
210. [210]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – October 2025,” Bank of England, Oct. 8, 2025. [Online]. Available: https://www.bankofengland.co.uk/financial-policy-committee-record/2025/october-2025 ↩
211. [211]Bank for International Settlements, “Volatility challenges risk-taking,” BIS Quarterly Review, Dec. 8, 2025. [Online]. Available: https://www.bis.org/publ/qtrpdf/r_qt2512a.htm ↩
212. [212]European Central Bank, “Financial Stability Review, May 2026,” ECB, May 27, 2026. [Online]. Available: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202605~50566915a7.en.html ↩
213. [213]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – September 2026,” Bank of England, Sep. 30, 2026. [Online]. Available: https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026 ↩
214. [214]Board of Governors of the Federal Reserve System, “Financial Stability Report – November 2025,” Federal Reserve Board, Nov. 7, 2025. [Online]. Available: https://www.federalreserve.gov/publications/files/financial-stability-report-20251107.pdf ↩
215. [215]European Central Bank, “Financial Stability Review, November 2025,” ECB, Nov. 26, 2025. [Online]. Available: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202511~263b5810d4.en.html ↩
216. [216]A. Bailey (Financial Stability Board), “FSB Chair’s letter to G20 Finance Ministers and Central Bank Governors: August 2026,” Financial Stability Board, Aug. 28, 2026. [Online]. Available: https://www.fsb.org/uploads/P310826.pdf ↩
217. [217]Bank for International Settlements, “Yields climb, yet risk appetite holds firm,” BIS Quarterly Review, Sep. 14, 2026. [Online]. Available: https://www.bis.org/publications/qr-202609/yields-climb-yet-risk-appetite-holds-firm ↩
218. [218]K. Leswing, “Apple ends day as world's most valuable company, passing Nvidia,” CNBC, Jul. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/07/27/apple-most-valuable-company-nvidia.html ↩
219. [219]Presidência da República, “Medida Provisória nº 1.318, de 17 de setembro de 2025,” Diário Oficial da União, Sep. 18, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/mpv/mpv1318.htm ↩
220. [220]G. Alckmin, F. Haddad and A. Silveira, “EMI nº 00010/2025 MDIC MF MME (Explanatory Memorandum of MP 1.318/2025),” Presidência da República, Sep. 16, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/Exm/Exm-1318-25.pdf ↩
221. [221]MDIC, “MP cria o Redata, que estimula datacenters e impulsiona economia digital no Brasil,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, Sep. 17, 2025. [Online]. Available: https://web.archive.org/web/20250929123902/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/setembro/mp-cria-o-redata-que-estimula-datacenters-e-impulsiona-economia-digital-no-brasil ↩
222. [222]Congresso Nacional, “Ato Declaratório do Presidente da Mesa do Congresso Nacional nº 11, de 2026,” Diário Oficial da União, Feb. 27, 2026. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/congresso/adc-11-mpv1.318.htm ↩
223. [223]W. Máximo, “Governo vai buscar diálogo para reativar incentivo para data centers,” Agência Brasil, Feb. 25, 2026. [Online]. Available: https://web.archive.org/web/20260226002349/https://agenciabrasil.ebc.com.br/economia/noticia/2026-02/governo-vai-buscar-dialogo-para-reativar-incentivo-para-data-centers ↩
224. [224]Câmara dos Deputados, “PL 278/2026: legislative record,” Câmara dos Deputados, Sep. 28, 2026. [Online]. Available: https://www.camara.leg.br/proposicoesWeb/fichadetramitacao?idProposicao=2600838 ↩
225. [225]Senado Federal, “PL 278/2026 (Redata),” Senado Federal, Sep. 28, 2026. [Online]. Available: https://www25.senado.leg.br/web/atividade/materias/-/materia/172786 ↩
226. [226]Presidência da República, “Lei nº 15.504, de 15 de setembro de 2026,” Diário Oficial da União (Extra Edition), Sep. 15, 2026. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/lei/l15504.htm ↩
227. [227]Presidência da República, “Lei Complementar nº 237, de 15 de setembro de 2026,” Diário Oficial da União (Extra Edition), Sep. 15, 2026. [Online]. Available: https://www.in.gov.br/web/dou/-/lei-complementar-n-237-de-15-de-setembro-de-2026-731891987 ↩
228. [228]P. Peduzzi, “Sancionado regime especial para incentivar instalação de datacenters,” Agência Brasil, Sep. 15, 2026. [Online]. Available: https://agenciabrasil.ebc.com.br/economia/noticia/2026-09/sancionado-regime-especial-para-incentivar-instalacao-de-datacenters ↩
229. [229]Imprensa Nacional, “DOU search: ‘Redata’ (Sep. 15 to Oct. 13, 2026),” Diário Oficial da União, accessed Oct. 13, 2026. [Online]. Available: https://www.in.gov.br/consulta/-/buscar/dou?q=Redata&s=todos&exactDate=personalizado&publishFrom=15-09-2026&publishTo=13-10-2026 ↩
230. [230]Microsoft News Center Brasil, “Microsoft anuncia investimento de R$ 14,7 bilhões ao longo de três anos em infraestrutura de nuvem e IA e oferece treinamento em IA em larga escala para capacitar 5 milhões de pessoas no Brasil,” Microsoft, Sep. 26, 2024. [Online]. Available: https://news.microsoft.com/pt-br/microsoft-anuncia-investimento-de-r-147-bilhoes-ao-longo-de-tres-anos-em-infraestrutura-de-nuvem-e-ia-e-oferece-treinamento-em-ia-em-larga-escala-para-capacitar-5-milhoes-de-pessoas-no-brasil/ ↩
231. [231]About Amazon Brasil, “AWS investirá mais de R$10 bilhões para expandir infraestrutura no Brasil,” Amazon, Sep. 11, 2024. [Online]. Available: https://www.aboutamazon.com.br/noticias/aws/aws-investira-mais-de-r-10-bilhoes-para-expandir-infraestrutura-no-brasil ↩
232. [232]MDIC, “CZPE aprova 12 projetos para R$ 585 bilhões em novos investimentos, incluindo datacenter, metais e amônia verde,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, Nov. 3, 2025. [Online]. Available: https://web.archive.org/web/20251104021618/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/outubro/czpe-aprova-12-projetos-para-r-585-bilhoes-em-novos-investimentos-incluindo-datacenter-metais-e-amonia-verde ↩
233. [233]Governo do Estado do Ceará, “Com R$ 200 bilhões de investimento, Data Center do TikTok no Ceará será o maior do Brasil,” Portal do Governo do Ceará, Dec. 3, 2025. [Online]. Available: https://web.archive.org/web/20251203170929/https://www.ceara.gov.br/2025/12/03/com-r-200-bilhoes-de-investimento-data-center-do-tiktok-no-ceara-sera-o-maior-do-brasil/ ↩
234. [234]Governo do Estado do Ceará (Ascom ZPE), “Com novo data center, ZPE Ceará consolida o Estado como hub tecnológico do País,” Portal do Governo do Ceará, Feb. 5, 2026. [Online]. Available: https://web.archive.org/web/20260205195108/https://www.ceara.gov.br/2026/02/05/com-novo-data-center-zpe-ceara-consolida-o-estado-como-hub-tecnologico-do-pais/ ↩
235. [235]Scala Data Centers, “Com investimento inicial de R$ 3 bilhões, governo do RS e Scala Data Centers assinam acordo para o maior projeto de infraestrutura digital do estado,” Scala Data Centers, Sep. 11, 2024. [Online]. Available: https://scaladatacenters.com/com-investimento-inicial-de-r-3-bilhoes-governo-do-rs-e-scala-data-centers-assinam-acordo-para-o-maior-projeto-de-infraestrutura-digital-do-estado/ ↩
236. [236]Governo do Estado do Rio Grande do Sul, “Ministério de Minas e Energia autoriza reserva energética para cidade de data centers no Rio Grande do Sul,” Portal do Estado do RS, May 14, 2025. [Online]. Available: https://estado.rs.gov.br/ministerio-de-minas-e-energia-autoriza-reserva-energetica-para-cidade-de-data-centers-no-rio-grande-do-sul ↩
237. [237]MME/SNTEP, “Portaria SNTEP/MME nº 2.942, de 9 de maio de 2025,” Diário Oficial da União, May 13, 2025. [Online]. Available: https://www.in.gov.br/web/dou/-/portaria-sntep/mme-n-2.942-de-9-de-maio-de-2025-628902215 ↩
238. [238]Elea Data Centers, “Elea assina MoU com Tapestry para assegurar capacidade energética de 1,5 GW da Rio AI City,” Elea Data Centers, Nov. 6, 2025. [Online]. Available: https://eleadatacenters.com/2025/11/06/elea-assina-mou-com-tapestry-para-assegurar-capacidade-energetica-de-15-gw-da-rio-ai-city/ ↩
239. [239]Elea Data Centers, “Elea anuncia a Rio AI City, projeto com capacidade de até 3,2GW em energia verde para impulsionar a Inteligência Artificial no mundo,” Elea Data Centers, May 8, 2025. [Online]. Available: https://eleadatacenters.com/2025/05/08/elea-anuncia-a-rio-ai-city/ ↩
240. [240]Elea Data Centers, “Elea Data Centers é adquirida pela I Squared Capital e inicia nova fase de expansão na América Latina,” Elea Data Centers, Apr. 28, 2026. [Online]. Available: https://eleadatacenters.com/2026/04/28/elea-data-centers-e-adquirida-pela-i-squared-capital-e-inicia-nova-fase-de-expansao-na-america-latina/ ↩
241. [241]Equinix, “Equinix fortalece liderança na América Latina com inauguração de novo data center SP6 em São Paulo,” Equinix Newsroom, Apr. 16, 2026. [Online]. Available: https://newsroom.equinix.com/2026-04-16-Equinix-fortalece-lideranca-na-America-Latina-com-inauguracao-de-novo-data-center-SP6-em-Sao-Paulo ↩
242. [242]EPE (Superintendência de Estudos Econômicos e Energéticos), “A Expansão da Carga de Data Centers no mercado elétrico brasileiro (III COPAM),” Empresa de Pesquisa Energética, Sep. 10, 2025. [Online]. Available: https://web.archive.org/web/20251004104216/https://www.epe.gov.br/sites-pt/areas-de-atuacao/energia-eletrica/WorkshopCOPAM/%5bEPE%5d%20Datacenter_AllexYukizaki%20III%20COPAM_09-2025.pdf ↩
243. [243]Imprensa Nacional, “DOU search: ‘alternativa de acesso à Rede Básica’ (Jan. 1, 2024 to Oct. 13, 2026),” Diário Oficial da União, Oct. 13, 2026. [Online]. Available: https://www.in.gov.br/consulta/-/buscar/dou?q=%22alternativa%20de%20acesso%20%C3%A0%20Rede%20B%C3%A1sica%22&s=todos&exactDate=personalizado&publishFrom=01-01-2024&publishTo=13-10-2026 ↩
244. [244]ANEEL, “Resolução Normativa ANEEL nº 1.122, de 20 de maio de 2025,” Diário Oficial da União, May 26, 2025. [Online]. Available: https://www.in.gov.br/web/dou/-/resolucao-normativa-aneel-n-1.122-de-20-de-maio-de-2025-631621948 ↩
245. [245]EPE, “Balanço Energético Nacional 2026: Relatório Síntese, ano base 2025,” Empresa de Pesquisa Energética, Aug. 3, 2026. [Online]. Available: https://www.epe.gov.br/pt/publicacoes-dados-abertos/publicacoes/balanco-energetico-nacional-2026 ↩
246. [246]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Eólicas (open data),” Operador Nacional do Sistema Elétrico, Oct. 12, 2026. [Online]. Available: https://dados.ons.org.br/dataset/restricao_coff_eolica_usi ↩
247. [247]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Fotovoltaicas (open data),” Operador Nacional do Sistema Elétrico, Oct. 12, 2026. [Online]. Available: https://dados.ons.org.br/dataset/restricao_coff_fotovoltaica ↩
248. [248]Presidência da República, “Lei nº 15.269, de 24 de novembro de 2025 (conversion of MP 1.304/2025),” Diário Oficial da União, Nov. 24, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/lei/l15269.htm ↩
249. [249]Casa dos Ventos, “Casa dos Ventos bate recorde com captação de US$ 1,1 bi nos EUA,” Casa dos Ventos, Jun. 18, 2026. [Online]. Available: https://casadosventos.com.br/noticias/captacao-nos-estados-unidos ↩
250. [250]B3, “Relatório anual de BDRs 2025,” B3, Dec. 8, 2025. [Online]. Available: https://www.b3.com.br/data/files/5D/D3/83/92/6430B9105B12E5A9AC094EA8/Relatorio%20Anual%20BDRs%202025.pdf ↩
251. [251]CVM, “Resolução CVM nº 175, de 23 de dezembro de 2022: Normative Annex I (consolidated text),” Comissão de Valores Mobiliários, Mar. 5, 2025. [Online]. Available: https://conteudo.cvm.gov.br/export/sites/cvm/legislacao/resolucoes/anexos/100/resol175consolid_Anexo01.pdf ↩
252. [252]Conselho Monetário Nacional, “Resolução CMN nº 4.994, de 24 de março de 2022 (as amended by CMN Resolution nº 5.202/2025),” Banco Central do Brasil, Mar. 24, 2022. [Online]. Available: https://www.bcb.gov.br/estabilidadefinanceira/exibenormativo?tipo=Resolu%C3%A7%C3%A3o%20CMN&numero=4994 ↩
253. [253]Banco Central do Brasil, “Relatório de Estabilidade Financeira, vol. 25, no. 1,” Banco Central do Brasil, May 22, 2026. [Online]. Available: https://www.bcb.gov.br/content/publicacoes/ref/202605/RELESTAB202605-refPub.pdf ↩
254. [254]Alphabet Inc., “Alphabet Announces Second Quarter 2026 Results,” Alphabet Investor Relations, Jul. 22, 2026. [Online]. Available: https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf ↩
255. [255]OpenAI, “DevDay 2026 Recap,” OpenAI, Sep. 29, 2026. [Online]. Available: https://openai.com/index/devday-2026-recap/ ↩
256. [256]E. Brynjolfsson, D. Li and L. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889–942, Feb. 4, 2025. [Online]. Available: https://doi.org/10.1093/qje/qjae044 ↩
257. [257]A. Bick, A. Blandin and D. Deming, “The Impact of Generative AI on Work Productivity,” Federal Reserve Bank of St. Louis — On the Economy, Feb. 27, 2025. [Online]. Available: https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity ↩
258. [258]U.S. Census Bureau, “Business Trends and Outlook Survey (BTOS): National estimates (AI use, collection ending Oct. 4, 2026),” U.S. Census Bureau, Oct. 8, 2026. [Online]. Available: https://www.census.gov/hfp/btos/downloads/National.xlsx ↩
259. [259]DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437, Dec. 27, 2024. [Online]. Available: https://arxiv.org/abs/2412.19437 ↩
260. [260]S. Subin, “Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history,” CNBC, Jan. 27, 2025. [Online]. Available: https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html ↩
261. [261]S. Meredith, “Nvidia becomes first company to reach $5 trillion valuation, fueled by AI boom,” CNBC, Oct. 29, 2025. [Online]. Available: https://www.cnbc.com/2025/10/29/nvidia-on-track-to-hit-historic-5-trillion-valuation-amid-ai-rally.html ↩

摘要

2024 至 2026 年間，美國最大的幾家科技公司開始投資 AI 資料中心，其速度被英格蘭銀行形容為「unprecedented historically」[1]。Microsoft、Alphabet、Amazon 與 Meta 的資本支出（capex）從 2024 年的 2,280億美元增加到 2025 年的 3,760億美元，而對 2026 年的預測合計最高達 7,450億美元。本文探討這筆投資是如何支付的，以及如果營收遲遲不來，風險由誰承擔。

本文的論點是：即使背後有真實的技術，AI 週期仍具備信貸泡沫的各項典型徵兆。供應商投資向自己採購的客戶：國際貨幣基金（IMF）統計，自 2023 年以來共有 14 項這類協議，涉及 1.4兆美元的未來營收承諾。部分債務留在表外。Meta 興建了一座 270億美元的園區，資金來自 273億美元的債券，發行者是一家合資企業的多數股東，而 Meta 並未將這家合資企業納入合併報表；產業中最大的幾家公司，已簽署但尚未開始的租約合計超過 1兆美元。擔保品老化得很快：NVIDIA 每年推出新產品，這些公司卻把伺服器分五或六年折舊。風險則集中在少數幾家公司，例如 OpenAI、Oracle 與 CoreWeave，並擴散到基金、保險公司與個人投資人。

本文把歷史當作機制的證據，而不是預測：英國的鐵路狂熱與美國 1873 年的恐慌、2000 年的光纖、Enron，以及 2008 年的危機。本文也說明哪些事實可能推翻這個論點，例如 AI 實驗室以外的雲端需求持續成長、資產負債表依然穩健，以及 2026 年舊款 GPU 的租金走高；最後提出幾種情境與一份值得追蹤的指標清單。其中一節專門討論巴西。

賭注有多大

2024 年，Microsoft、Alphabet、Amazon 與 Meta 合計花了約 2,280億美元的資本支出，也就是現金流量表上記載的建築物、機器與設備購置。2025 年是 3,760億美元，增加了 65%。光是 2026 年上半年就達 2,950億美元，接近前一整年的 80%。這些數字是我依各公司向美國證券交易委員會（SEC）申報的資料計算的，其中 Microsoft 的數字依季報換算為曆年 [2], [3], [4], [5], [6], [7], [8], [9]。再加上 Oracle（其資本支出在截至 2026年5月的會計年度增加了一倍以上），這五家公司每年的支出規模超過 4,300億美元 [10]。

下表列出各公司的資本支出（單位為十億美元），以及對 2026 年的最新預測。Microsoft 與 Oracle 的會計年度與曆年不一致，各家預測的定義也不同：Microsoft 的預測包含融資租賃，Meta 的則包含這類租賃所支付的本金 [11], [12], [13], [14], [15], [16]。

Microsoft、Alphabet、Amazon 與 Meta 對 2026 年的預測加總後，介於 7,200億美元至 7,450億美元之間，幾乎是 2025 年的兩倍。信用評等機構 Moody's 在 CNBC 引述的一份 2026年7月報告中預估，六家公司（上述四家，加上 Oracle 與 CoreWeave）2026 年的資本支出將達 7,850億美元，2027 年約 1兆美元 [17]。光是 Oracle 一家，在 2026年6月至8月這一季就花了 285億美元，比它整個 2025 會計年度還多 [18]。

錢買了什麼

大部分支出流向壽命短的設備。Alphabet 表示，2025 年的投資配置是「approximately 60% of that investment in servers and 40% in data centers and networking equipment」[19]。Microsoft 談到 2026年4月至6月這一季時說：「Roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs」[13]。正是這些設備，會以「折舊的物理」一節所描述的速度貶值。

部分增幅來自價格，而不是數量。2026年4月，Microsoft 預估全年資本支出約 1,900億美元，其中包括「approximately $25 billion from the impact of higher component pricing」[20]。Meta 在同月上調預測，並解釋這次修正「reflects our expectations for higher component pricing this year」[21]。Amazon 在 7月把自己的預測調高到 2,200億美元，執行長 Andy Jassy 將增幅歸因於記憶體價格 [15]。國際能源總署（IEA）指出，高頻寬記憶體（也就是 AI 晶片所使用的 HBM）出現短缺，預計將「persist through at least the end of 2027」[22]。在投資週期中，掌握稀缺零組件的一方往往能拿走利潤，而 2026 年的稀缺零組件就是記憶體。

這些公司以需求為這筆帳辯護。Microsoft 表示自己「confident in the return on these investments given higher demand signals and increasing product usage」[20]。Jassy 說得更進一步：「But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027 too,」以及「In fact, the demand we already have for 2028 is striking.」AWS 的待履行合約（backlog），也就是已簽署但尚未交付的合約，達到 4,960億美元 [15]。這些數字是反駁本文論點最有力的論據，「什麼可能推翻這個論點」一節會再談到。但 2025 與 2026 年簽下的最大幾筆雲端合約，另一方都是仍在燒錢、仰賴新一輪融資的 AI 實驗室，如「循環」一節所示。

相對於經濟的規模

依我的計算，四大公司 2025 年的資本支出約相當於美國名目 GDP 的 1.2%；依 2026 年的預測，則會落在 2.2% 至 2.3% 之間 [23]。這些錢並非全都花在美國境內，因為有些資料中心位於海外，而且很大一部分設備是進口的，如總體經濟那一節所示。即便如此，這個量級仍然驚人。根據 IEA：「Capital expenditure of just five technology companies is now larger than global investment in oil and natural gas production.」[22]。以占經濟的比例而言，這個週期仍遠不及英國的鐵路狂熱：1847 年，鐵路投資達到 GDP 的 7.3% [24]。當前週期的不同之處在於集中度：四或五家公司就占了大部分支出，而最大的幾家科技公司（包括投資最多的這四家）合計占 S&P 500 的三分之一以上，如總體經濟那一節所示。

現金已不足以支應一切

直到不久前，大型科技公司還能用自己產生的現金支付投資，而且綽綽有餘。這個餘裕縮小了。下表比較營運現金流與資本支出（單位為十億美元），依我的計算，根據的是向 SEC 申報的數字：

這些公司自己也記錄了這個轉折。Amazon 在 2026年7月表示：「Free cash flow decreased to an outflow of $7.6 billion for the trailing twelve months, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment, net of proceeds from sales and incentives. This increase primarily reflects investments in artificial intelligence.」[25]。Alphabet 在同月則說：「We had negative free cash flow of $5.9 billion in the second quarter, driven by our investments in CapEx.」[14]。

當資本支出超過現金時，差額就得來自債務、發行股票或長期租賃合約。據 CNBC 報導，Moody's 如此總結這個轉變：「Previously, these companies relied on asset-light structures centered on software, intellectual property, and scalable cloud services that required modest capital investment,」以及「The transition from asset-light to asset-heavy models requires unprecedented levels of investment and capital raising.」該機構估計，這六家公司的直接債務約 4,600億美元，租賃承諾為 1.2兆美元，其中超過 8,200億美元屬於尚未開始的合約；它並表示把這些承諾視為等同債務 [17]。「表外」一節會說明這些數字從何而來。

這些本身都不足以證明泡沫存在。賺錢的公司大舉投資一項新技術，正是技術革命時期理應出現的情況。關於泡沫的文獻要我們看的是另一件事：投資是怎麼支付的，以及如果營收沒有到來，風險由誰承擔。這是下一節的主題。

如何辨認泡沫

「泡沫」已經成了吵架用語。認為 AI 太貴的人拿這個詞來罵人；不同意的人則回應說，這項技術是真的。兩者可能同時成立。金融危機領域經典著作的作者 Charles Kindleberger 與 Robert Aliber，對這些詞彙的區分更加謹慎：「The term mania describes the frenzied pattern of purchases, often an increase in prices accompanied by an increase in trading volumes; individuals are eager to buy before the prices increase further. The term bubble suggests that when the prices stop increasing, they are likely—indeed almost certain—to decline.」[26]。同一本書中還有一個但書，值得在閱讀全文時時放在眼前：「Virtually every mania is associated with a robust economic expansion, but only a few economic expansions are associated with a mania.」[26]。

他們描述的劇本從一個真實的衝擊開始。Kindleberger 與 Aliber 這樣概述 Hyman Minsky 的模型：「Minsky argued that the events that lead to a crisis start with a ‘displacement’, some exogenous, outside shock to the macroeconomic system. If the shock was sufficiently large and pervasive, the economic outlook and the anticipated profit opportunities would improve in at least one important sector of the economy.」[27]。接著是狂熱期，此時「investment spending surges because credit is plentiful」，錢看起來像是免費的（「money always seems free in manias」）。然後是漲勢暫停、借錢買進的人被迫賣出，有時還會出現恐慌 [26]。這本書的核心論點並非心理學上的。它是關於信貸的論點：「the cycle of manias and panics results from the pro-cyclical changes in the supply of credit」[26]。

Minsky 為這段歷程提供了一套可以用來解讀資產負債表的詞彙。他把借錢的人分成三種類型。避險型（hedge）單位「can fulfill all of their contractual payment obligations by their cash flows」。投機型單位用營收支付利息，但必須展延本金：「Such units need to "roll over" their liabilities」。至於龐氏（Ponzi）單位，營運現金流連利息都不夠付，它們「can sell assets or borrow」才能撐下去 [28]。金融不穩定假說的第二定理，是這裡最值得注意的一句話：「over periods of prolonged prosperity, the economy transits from financial relations that make for a stable system to financial relations that make for an unstable system」[28]。

套用到 AI 週期，這套分類能把辯論中常被混為一談的公司區分開來。以自身營運現金支付投資的大型平台，依 Minsky 的標準屬於避險型單位，即使它們也發債。專攻 GPU 的雲端公司以貸款購買晶片，再靠未來的合約與再融資來償還，比較接近投機型。而一家採購承諾高達自身營收數倍、必須仰賴新一輪募資才能履行承諾的公司，就進入了 Minsky 所說的龐氏範疇。這是一個技術性用語，並不是指控任何人詐欺。它描述的是一種只有在資金持續流入時才撐得住的結構。「循環」與「最弱的環節」兩節會說明每家公司落在哪個位置。

Minsky 也指出是誰把體系從一種狀態推向另一種狀態：「bankers (using the term generically for all intermediaries in finance), whether they be brokers or dealers, are merchants of debt who strive to innovate in the assets they acquire and the liabilities they market.」[28]。2007 年，創新出現在證券化工具上。在 AI 週期中，創新則出現在興建資料中心的特殊目的實體（SPV）、以 GPU 為擔保的貸款，以及買進這些債務的私募信貸基金。形式改變了，角色依然相同。

經濟學家 Carlota Perez 補上了理解科技泡沫所缺的那一塊。在她的解讀中，每一次技術革命都有一段為期 20 至 30 年的導入期，由金融資本主導，並以一場大泡沫告終。經過一個轉折點之後，便進入由生產資本主導的展開期 [29]。泡沫不是意外：「A major technology bubble (MTB) is not an accidental event. It regularly occurs midway along the process of assimilation of each technological revolution.」[30]。

Perez 理論中最令人不安的一點是：泡沫有其功能。「It is the switch to short-term gains during the bubble that attracts the necessary capital to be poured into the infrastructural networks of each revolution.」[30]。鐵路與光纖都是用期待快速回報的資金過度興建的，後來才由其他人使用。她用兩句話描述高峰期的氛圍：「Soon there is more capital wanting a piece of the action than projects looking for funds.」以及「The illusion is that there are high profits to be had with very low risk.」[30]。她也提出辨認重大科技泡沫的兩項判準：「The two defining characteristics of these major technology bubbles are: (a) their concentration on the new technologies –especially the new infrastructural networks– and (b) their decoupling from the real economy.」[30]。

Perez 把這類泡沫與 Minsky 所描述的信貸泡沫區分開來：「They are the result of opportunity pull rather than of easy credit push. But they are indeed bubbles.」[30]。這個區分是本文的主軸。AI 週期一開始是一場機會型泡沫，資金主要來自現金與股票。從 2025 年起，它開始帶有信貸泡沫的印記。IMF 指出，自 2025年1月以來，超大規模雲端業者（hyperscalers）已發行超過 1,000億美元的債券，「supplemented by leveraged loans and intercorporate arrangements that are, at times, part of circular financing structures」[31]。正是這種組合，讓 2008 年的經驗具有參考價值。

畢生研究泡沫的 Robert Shiller 偏好行為面的定義：「A situation in which news of price increases spurs investor enthusiasm which spreads by psychological contagion from person to person, in the process amplifying stories that might justify the price increase and bringing in a larger and larger class of investors, who, despite doubts about the real value of the investment, are drawn to it partly through envy of others’ successes and partly through a gambler’s excitement.」[32]。以他命名的指標 CAPE，是把 S&P 指數的價格除以十年平均實質盈餘。根據 Shiller 在 2026年10月5日更新的資料序列，CAPE 在 1999年12月達到 44.20 的歷史紀錄，2007年5月為 27.55。2026年8月為 40.99，是 2000 年以來的最高讀數；2026年10月的初步數值為 40.71。1881 年以來的平均值是 17.80 [33]。

及早辨認出泡沫，並不等於能猜中它何時破裂。1996年12月5日，聯準會主席 Alan Greenspan 公開提問：「But how do we know when irrational exuberance has unduly escalated asset values, which then become subject to unexpected and prolonged contractions as they have in Japan over the past decade?」[34]。CAPE 當月為 27.72，之後又上升了三年，在 1999年12月達到 44.20 [33]。

Shiller 本人也提醒，這個指標無法標出時間點，而且泡沫不一定會一次破裂：「Though the abrupt ends of stock market booms in 1929, 2000 and 2007 might seem consonant with such a metaphor, these booms were reflated again before long」[32]。研究過鐵路狂熱的 Andrew Odlyzko，記錄了一個推翻這條規律的案例：1830 年代的英國鐵路狂熱，既是投機性的，也讓投資人獲利。他的結論可以為本文踩煞車：「The example of the railway mania of the 1830s serves as a useful antidote to claims that bubbles are easy to detect or that all large and quick jumps in asset valuations are irrational.」[35]。

接下來，本文使用六個標誌，全都取自這些文獻，也都可以用公開資料驗證：

沒有任何一個標誌單獨就能證明泡沫存在。本文的論點是：2026 年，這六個標誌同時出現；而其中三個（信貸、負債端的創新與循環營收），在接下來要談的幾段歷史中，正是區分市場修正與金融危機的關鍵。

鐵路：基礎建設留下，投資人買單

最常被拿來與 AI 比較的是 2000 年的網際網路，但最有用的比較對象或許更古老。在 1840 年代的英國，鐵路是將改變一切的技術，而它也確實改變了一切。每條路線都需要國會核准，因此留下了這股狂熱的逐年紀錄。1843 年核准了 91 英里，1844 年 805 英里，1845 年 2,700 英里，1846 年則創下 4,538 英里的紀錄。1847 年為 1,354 英里；1850 年只有 8 英里 [24]。

在所有數據中，投資規模最能與 2026 年對照。根據 Odlyzko，鐵路支出「peaked in 1847, at a level of about £44 million (in contrast to about £50 million for the national government budget)」。依他的計算，「railway investments of £44 million in 1847 came to 7.3% of a GDP of £600 million, which is comparable to $1.1 trillion for the U.S. today」，而文中的「today」指的是 2010 年。在同一篇論文的表格中，1847 年全國投資的 8,800萬英鎊裡，鐵路占了 4,390萬英鎊，接近全英國投資的一半 [24]。作者最有力的比較是與軍事預算相比：「At its height, individual capitalists, in pursuit of private profit, were plowing more than twice as much into the construction of a public infrastructure as their nation was spending on the military.」[24]。

股價下跌時，投資並沒有停止。Odlyzko 解釋了原因：「Even in 1850, when investors felt crushed by the stock market decline, investments were high, higher than they had ever been before 1845. The reason was not that any shareholders were enthusiastic about the spending. Rather, so much money had been spent that it was felt better to go on and complete the lines and earn small returns than abandon them completely and get nothing.」[24]。這是沉沒成本的邏輯，而它對 AI 很重要。蓋到一半的資料中心園區一文不值，因此在信心消失之後，資本支出往往還會維持在高檔一段時間。

對出錢的人來說，結果很糟。「The best known index of railway shares prices declined from 167.9 in the summer of 1845 to 60.5 at the end of 1849, a drop of 64%.」[36]。Economist 在 1855 年總結了這個矛盾，Odlyzko 轉引了這句話：「[m]echanically or scientifically, the railways, with all their multiplied conveniences and contrivances, are an honour to our age and country: commercially, they are great failures.」[24]。持有其中一條鐵路股票的 Charlotte Brontë 在 1849 年寫道（同樣引自 Odlyzko）：「The original price of Shares in this Railway was £50. At one time they rose to 120; and for some years gave a dividend of 10 per cent; they are now down at 20, and it is doubtful whether any dividend will be declared this half-year.」[24]。

一般的解釋把責任歸咎於詐欺者，例如粉飾帳目的 George Hudson。Odlyzko 不同意：「accounting fraud was not the fundamental cause of the investment disaster of the Railway Mania. Distorted financials helped mask for a while the disappointing returns that new railways produced for investors, but were not the reason for the lack of profits. The main problem was the headlong expansion during the Mania. It tripled British railway mileage, and the new lines cost more than expected and brought revenues lower than expected.」[36]。每英里的年營收從 1845 年的 3,470英鎊降到 1850 年的 2,080英鎊 [36]。這個教訓今天依然適用：基礎建設泡沫的問題通常是算術問題，而不是道德問題。產能過多、造價過高，而需求雖然在成長，卻比當初承諾的慢。

Kindleberger 描述了投資人的組成如何隨著週期改變：「In early railway investment, for example, initial investors were persons doing business along the rights of way who sought benefits from the railroad for their other concerns. They were followed by a second group of investors interested in the profits the railroad would earn, and by a third group, made up of speculators who, seeing the rise in the railroad’s shares, borrowed money or paid for the initial instalments with no intention of completing the purchase, to make a profit on resale.」[37]。在 AI 領域，第一群是為自家產品建置運算容量的公司。第二群是建置容量以出售或出租給第三方的業者。第三群則出現在資金開始湧入的時候，而湧入的原因是價格上漲，不是現金流足以支撐。

Perez 記下了鐵路狂熱中一個令人想起 Minsky 龐氏型態的細節：「During the railway mania in the 1840s, though, companies without profits did pay dividends out of the new capital they attracted in order to keep previous investors in and new investors coming.」[30]。當新資金被用來支付先來的人，這個結構就得仰賴新資金持續流入。

Odlyzko 明確地把維多利亞時代的狂熱與 2000 年的狂熱連結起來。在他看來，網路泡沫中兩個最大的實體投資災難，是美國的長途光纖與歐洲的 3G 頻譜拍賣，「each of which cost on the order of $100 billion to €100 billion」[24]。在這兩個案例中，「there was hard quantitative evidence that once those went into service, there would not be enough demand to provide the expected profits.」[24]。而他用來形容這兩段插曲的話，也適用於第三段：「society decided to adopt the new offerings at its own speed, slower than that expected by inventors, promoters, and investors, and financial ruin followed in both cases.」[24]。

也有反面的觀點，而且出自同一位作者。1830 年代的鐵路狂熱規模較小、時間較早，投資人賺到了錢。Odlyzko 也把以股權融資的基礎建設與債務危機區分開來：「One could argue that the infrastructure construction funded by equity is radically different from the much more frequent, far more dramatic and attention-grabbing financial crises that are associated with large accumulations of debt, typically for consumption.」[35]。這個區分決定了損害的大小。當帳單由股東承擔時，損失雖然痛苦，卻不會像銀行危機那樣擴散。當帳單落在高槓桿的債權人身上，損失就會擴散。

1873：當鐵路靠舉債支付

在大西洋彼岸，故事有不同的結局，因為資金的性質不同。用聯準會歷史學者的話來說，美國 1873 年的危機「arose from investments in railroads. Railroads had expanded rapidly in the nineteenth century, and investors in many early projects had earned high returns. As the Gilded Age progressed, investment in railroads continued, but new projects outpaced demand for new capacity, and returns on railroad investments declined.」[38]。當維也納股市的崩盤促使歐洲投資人拋售美國鐵路債券時，這些公司便沒有錢營運，也無法為債務再融資。對 Northern Pacific 曝險甚深的 Jay Cooke & Co. 銀行在 1873年9月18日倒閉。兩天後，紐約證券交易所史上首次休市，十天沒有交易，全美至少有一百家銀行倒閉 [38]。

與英國鐵路狂熱的差別在於負債端。「Railroads were built mainly with borrowed money」，Kay Giesecke、Francis Longstaff、Stephen Schaefer 與 Ilya Strebulaev 在一項涵蓋 150 年公司債違約的研究中寫道。1873 至 1875 年間，違約總額達「36 percent of the par value of the entire corporate bond market」；大蕭條時期最糟的三年則是 12.88% [39]。隨後的經濟收縮持續了 65 個月，從 1873年10月到 1879年3月，是 NBER 年表中最長的一次 [40]。作者們自己也把它與近年的週期連結起來：19 世紀以債務融資的技術熱潮之後，都接著出現危機，「closely resembling patterns in the telecommunications and structured-credit-products cycles of the past decade.」[39]。

這裡同樣有反方觀點。經濟史學家 Scott Mixon 指出，1873 年之前，股票、選擇權與債券就已經在為風險定價。投資人「were not caught up in an irrational, speculative mania」，而危機是從槓桿最高的鐵路公司開始的 [41]。並不是沒有警告。問題出在債務集中於最弱的環節，而這正是「最弱的環節」一節要回答的問題。

比較各次衰退持續的時間，就能概括本文的論點。根據 NBER，主要靠股權融資的網路泡沫破裂之後，隨之而來的衰退持續了 8 個月，從 2001年3月到11月。2007–2009 年那次是信貸危機，持續了 18 個月。1873 年那次是鐵路債券危機，持續了 65 個月 [40]。損害的大小，與其說取決於技術，不如說取決於誰手上握著債務。

鐵路為 AI 留下兩個教訓。基礎建設會比泡沫活得更久：狂熱時期的路線都完工了，在股價崩跌之後仍持續使用；今天資料中心的建築、變電站與輸電線，也終將為某些人所用。但一條鐵路可以用上數十年，而作為這筆投資理由的晶片只能用幾年，如折舊那一節所示。此外，直到 2024 年，AI 週期主要都還是以現金與股票支付，就像英國的幾次鐵路狂熱；自 2025 年起，它正轉向債務，就像最後以 1873 年恐慌收場的美國鐵路擴張。

電信，1996–2002：閒置光纖與來回營收

如果說鐵路展示了基礎建設泡沫的算術，那麼 1990 年代末的電信業展示的就是會計。那是最近一次有網路技術吸引到這種規模資本的週期，本文討論的機制幾乎都在當時出現過：以容量衡量需求、供應商為客戶提供融資、競爭對手之間的交叉銷售，以及一項讓供給倍增、速度快過市場使用能力的技術進步。

投資規模大，速度也快。里奇蒙聯邦準備銀行的經濟學家計算，從 1996 年第一季到 2000 年第四季，「investment in communications equipment grew from approximately $62 billion per year to over $135 billion per year in constant 1996 dollars」，平均每年成長近 18%。下滑也同樣迅速：2001 年第四季，這個數字跌到 930億美元以下，「only 69 percent of the same figure one year earlier」；通訊設備占民間投資的比重，也從 2000 年的近 7% 降到 2002 年底的 4.8% [42]。

支撐投資的迷思

支撐這波資本支出的說法，是一項統計數字。Andrew Odlyzko 這樣描述它：「The most popular and extremely misleading myths of the dot-com and telecom bubbles was that “Internet traffic doubles every 100 days” (or 3 months, or 4 months).」[43]。他說，真實的數字小得多，但仍然非常驚人：流量「close to doubling each year since 1997」。這個迷思的主要來源有名有姓：「At its base, most of the support for the myth of astronomical growth rates in Internet traffic was coming from WorldCom.」[43]。在 2000年3月的一場簡報中，WorldCom 總裁 Bernie Ebbers 表示，由於需求的關係，公司 2003 年的資本支出「would have to exceed $100 billion」[43]。

對 2026 年最有參考價值的細節，是這個迷思背後的混淆：「most of the WorldCom claims about astronomical growth rates of their network were about network capacity, not traffic.」[43]。裝置容量由投資的人決定。流量由使用的人決定。當前者被當成後者來呈現，市場就失去了溫度計。「循環」一節會說明，AI 的需求數字有很大一部分（例如簽約的吉瓦數與待履行合約）衡量的是容量承諾，而不是使用量。

光纖太多，每一條還載得更多

FCC 最後一次光纖年度調查（資料截至 1998 年底）就已經呈現出這樣的格局。老牌業者使用了自有光纖中相當大的比例：AT&T 是 50%，Sprint 是 85%。從零開始建設整張網路的新業者，則幾乎沒有點亮自己的光纖。1998 年，Qwest 已點亮的光纖比例是 2%，Williams 1%，GST 2%，NEON 5%。光是那一年，長途電信業者鋪設的光纖就成長了超過 30%，達到超過 360萬光纖英里 [44]。FCC 的報告用一句話解釋了過剩背後的邏輯：「Although up-front costs for fiber deployment are high, a significant portion of the total investment can be deferred until actual demand materializes.」[44]。里奇蒙的經濟學家補充，1996 年的四大業者擁有美國 72% 的長途光纖，到了 1999 年只剩 30%，同時每年新鋪設的光纖增加到原來的四倍以上 [42]。

舊金山聯邦準備銀行的 Mark Doms 用另一種方式衡量這場競賽：「in 1996 there were 143,000 route miles of fiber in the country, but by 2000 this had increased to 320,000.」1990 年，有四家公司擁有全國性的光纖網路；2000 年則有十家。「Chicago and Minneapolis are each served by twenty long-haul firms」，而 1990 年約為五家 [45]。

而且每一條光纖能承載的量也大增。「In 1996, a strand of fiber transmitted data at approximately 2.5 gigabits per second (Gbps). By 2000, the capacity of the same fiber could reach 100 Gbps.」這要歸功於分波多工技術；單一光纖的容量「approximately every nine months between 1996 and 2000」就倍增一次，比摩爾定律還快 [42]。作者列出造成過剩的三個原因：新產業典型的過熱、每條光纖的容量成長超出預期，以及「perhaps most importantly」，需求「had not grown as fast as many had forecast: the pessimists turned out to be right.」[42]。

這是與 AI 最直接的技術類比。當效率提升得比需求快，為預期需求而建的實體容量就會過剩，即使需求仍在成長。在 AI 領域，效率來自每年推出的新晶片，以及用更少運算就能得到相同結果的模型。折舊那一節會用數字討論這一點。

付出的代價

Nasdaq 的電信類股指數從 1997年4月的 198，漲到 2000年3月10日的 1,230.06，到當年年底下跌了 62%。2003年5月16日時為 136。光是 S&P 的電信公司，在 2000 至 2002 年間就損失了約 7,000億美元的市值，「more than 3.5 percent of the entire value of U.S. corporate equities at the stock market peak in 2000」[42]。Nasdaq Composite 指數在同樣的 2000年3月10日收在 5,048.62，2002年10月9日收在 1,114.11，跌幅 77.9%。直到 2015年4月23日，它才再度收在 2000 年的高點之上，那已是十五年多以後 [46]。

國際清算銀行（BIS）在 2002 年解釋了銀行為何能相對毫髮無傷：「a considerable amount of the financing for the most risky elements of the late 1990s boom was obtained from outside the banking system. This is perhaps best illustrated by the fact that firms in the technology and telecommunications sectors relied heavily on vendor financing, venture capital and the equity and bond markets for their financial needs.」[47]。這與今天出現的融資組合相同，只是換了名字：供應商信貸、創業投資、股票與債券。

借錢給客戶買自家產品的供應商

設備製造商 2001 年的財報，記錄了將在 2025 年重現的機制。Lucent 在年報中解釋：「Some of our customers worldwide are requiring their suppliers to arrange or provide long-term financing for them as a condition of obtaining or bidding on infrastructure projects. These projects may require financing in amounts ranging from modest sums to more than a billion dollars.」[48]。截至 2000年9月30日，Lucent 對客戶的融資承諾總額為 81億美元。在 2001 會計年度，該公司為呆帳與客戶融資提列了 22億美元的準備，其中約 13億美元涉及三個專案，包括在 2001年4月聲請破產保護的 Winstar [48]。在風險因素中，有一句話適用於今天任何投資自家客戶的供應商：「We are vulnerable to downturns in our customers' businesses as a result of our customer financing arrangements.」[48]。

Nortel 講的也是同一個故事。「As a result of the competitive environment during 1999 and 2000, we provided significant amounts of medium-term and long-term customer financing」[49]。客戶融資總額從 2000 年底的 51.68億美元降到 2001 年底的 20.75億美元，因為「our ability to place customer financing with third-party lenders was significantly reduced during 2001」。Nortel 2001 年的淨損為 273億美元 [49]。Cisco 則把信貸當成銷售的武器：「We believe customer financing is a competitive factor in obtaining business, particularly in supplying customers involved in significant infrastructure projects.」它提供的貸款可以包括「additional funds for soft costs associated with network installation and integration of our products and for working capital purposes」[50]。

2026年10月，BIS 的經濟學家正是用這段插曲來說明 AI 循環關係的風險：「upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors' equipment [...]. This meant that part of the equipment vendors' reported sales was being funded by the vendors themselves.」[51]。

來回營收

在某些案例中，循環超出了融資的範圍，變成捏造的營收。2004 年，SEC 指控 Qwest 在 1999 至 2002 年間以詐欺方式認列了超過 38億美元的營收。其中一部分來自出售光纖使用權（IRU），而在需求下滑之後，則來自交換交易：「When the demand for IRUs declined, Qwest engaged in IRU "swaps" whereby Qwest bought IRUs from other companies in exchange for agreements from those companies to buy IRUs from Qwest.」[52]。據 SEC 表示，這種手法成了例行公事，員工甚至把這種做法比作成癮，並把非經常性銷售稱為「heroin」。Qwest 在不承認也不否認指控的情況下達成和解，並支付 2.5億美元罰款 [52]。

Global Crossing 是最具教學意義的案例，因為 SEC 量化了其中的循環。2001 年上半年，該公司的容量銷售開始仰賴同時向同一批業者進行的採購：「GX’s ability to sell capacity often became linked to, and in some cases dependent on, GX’s simultaneous purchase of a nearly equal dollar amount of capacity.」2001 年第二季，在 5.67億美元的 IRU 銷售中，這類「reciprocal transactions」占了 5.16億美元，也就是 91%；若沒有這些交易，公司的「Recurring Adjusted EBITDA」，「would have gone from the reported $472 million to a negative $43 million.」[53]。Global Crossing 於 2002年1月28日聲請破產保護。這是一宗資訊揭露案，以和解結案，SEC 並未指控詐欺 [53]。

流量迷思的源頭 WorldCom，最後成了最單純的案子。2002年6月，SEC 指控它犯下「more than $3.8 billion」的會計詐欺，手法是把支付給其他電信業者的網路使用費列為投資，而非費用 [54]。該公司於 2002年7月21日聲請破產保護 [55]。

反方觀點

並不是每個人都認為電信業出現過泡沫。里奇蒙的經濟學家在 2003 年撰文，稱這個週期「in large part—a rational response to the changing fundamentals of technology and regulatory environment」，其中伴隨著在不確定環境中的重大預測失誤：「our explanation for the telecommunications boom and bust does not involve fraud, irrationality, or a bubble. To be sure, as the bust became apparent, fraud did occur. But it is not clear that fraud played an important role in the boom and the early stages of the bust.」[42]。這個論點值得認真看待，也適用於 AI：需求預測可能出錯，而完全不涉及詐欺。詐欺是後來才出現的，出現在數字必須兜得攏的時候。

Doms 補充了一個值得留給資料中心參考的技術論點：「the marginal cost of fiber is small relative to the cost of laying a new fiber route; in other words, it would be much cheaper to lay additional strands of fiber today while trenches are open than to dig up the trenches again several years from now and lay some additional fiber lines. So, although the utilization rate is low, that does not imply that firms were irrational in the amount of fiber they installed.」即便如此，他也承認 1999 與 2000 年的技術投資遠高於模型的預測，因為「expectations were too high, especially in two sectors of the economy, telecommunications services and the dot-com sector.」[45]。

電信業留下了四個值得在 2026 年尋找的訊號，而且沒有一個需要假設有人心懷不軌。以簽約容量衡量的需求。由賣方自己提供融資的銷售。互為客戶與供應商的公司之間的交叉合約。以及一項讓供給倍增、速度快過使用量的技術。接下來幾節會說明，這四個訊號都出現在 AI 週期中，而且規模更大。

Enron：一樁醜聞催生的規則

Enron 出現在本文中，是基於兩個與詐欺本身無關的理由。第一個是促使該公司把債務移出資產負債表的經濟動機，因為這與任何必須先大量投資、之後才有收入的公司的動機相同。第二個是，今天決定一座資料中心是否出現在大型科技公司資產負債表上的會計規則，正是從這個案子的調查中誕生的。

Enron 董事會特別委員會的報告，也就是一般所稱的 Powers 報告，不加修飾地描述了這個問題：「Much of this growth involved large initial capital investments that were not expected to generate significant earnings or cash flow in the short term. While Enron believed these investments would be beneficial over a period of time, they placed immediate pressure on Enron's balance sheet. Enron already had a substantial debt load. Funding the new investments by issuing additional debt was unattractive because cash flow in the early years would be insufficient to service that debt and would place pressure on Enron's credit ratings. Maintaining Enron's credit ratings at investment grade was vital to the conduct of its energy trading business.」[56]。

解決辦法是把外部合夥人引進獨立的實體，這些實體可以用自己的名義借款，「although in many cases a guaranty or other form of credit support was required from Enron」。報告也解釋了管理階層為何偏好這種形式：「Enron management preferred the latter treatment--known as "off-balance-sheet"--because it would enable Enron to present itself more attractively as measured by the ratios favored by Wall Street analysts and rating agencies.」[56]。龐大的投資、遙遠的回報，再加上一個需要保護的信用評等：這與一座 270億美元資料中心園區的動機完全相同；「表外」一節會說明，今天如何在規則之內解決這個問題。

3% 規則

2003 年以前，表內與表外的界線幾乎是一道算術題。Powers 報告總結道，公司可以把一個特殊目的實體（SPE）視為獨立實體，條件是「(1) an owner independent of the company must make a substantive equity investment of at least 3% of the SPE's assets, and that 3% must remain at risk throughout the transaction; and (2) the independent owner must exercise control of the SPE.」[56]。SEC 幕僚後來描述了實際效果：只要把實體的表決控制權交給一個出資 3% 的第三方，「a sponsor could generally avoid consolidation, despite the fact that the activities of the SPE could not be substantively changed by the "controlling" investor.」[57]。

Enron 有些結構甚至不符合這條規則；另一些則是形式上符合，實質上不符合。其中最有名的「Raptor」工具，是用來防範 Enron 自身投資跌價的保護機制，資本卻來自 Enron 的股票。Powers 報告指出了其中的缺陷：「If the value of the investments fell at the same time as the value of Enron stock fell, the SPEs would be unable to meet their obligations and the "hedges" would fail. This is precisely what happened in late 2000 and early 2001.」[56]。2000 年下半年，來自 Raptor 的「獲利」占 Enron 公布之稅前盈餘的 80% 以上 [56]。一項恰好在需要時失去價值的擔保，有個專業名稱：錯向風險（wrong-way risk）；這個名稱會在 2008 年那一節以及談 AI 循環的那一節再次出現。

繞一圈回來的營收

Enron 的機制中，與本文主題最相近的不是 Raptor，而是「prepays」。這是三方交易，參與者包括一家 Enron 關係企業、一家銀行，以及由該銀行設立的一個工具，款項在其中繞了完整的一圈。SEC 幕僚描述了結果：「The circular nature of delivery and payments with respect to the commodities had the effect of eliminating any material risk or any potential gain with respect to the changes in the price of the underlying commodity. Each party's apparent assumption of price risk was illusory. With the elimination of price risk, "prepays" were effectively debt.」[57]。Enron 把這筆錢記為營運現金，把義務記為風險管理負債，而不是債務：「these prepay transactions allowed Enron to hide debt, lower key financial ratios followed by analysts, and provide the illusion of cash flow from operations.」[57]。

同一段文字也提供了本文用來解讀 AI 協議的分析工具，也就是所謂的會計單位問題：「The economics of a transaction may look quite different depending on how broadly or narrowly one defines the boundaries of the transaction. That is, a particular contract may appear to have certain economic characteristics when viewed in isolation—and may be given a certain accounting treatment that corresponds to those economic characteristics—but if understood as a piece of a larger agreed-upon transaction may actually have quite different economics, and be properly accorded different accounting treatment.」[57]。換句話說：要理解一項循環協議，必須看整個迴圈，而不是單獨看每一段。這是一種方法，而不是指控。後文討論的 AI 協議都是公開的，也沒有任何一項被指控涉及詐欺。

隱藏了多少

被排除在數字之外的規模，是這個案子中最令人心驚的數據。SEC 幕僚引述破產法院指派的檢查人指出：「For 2000, barring the use of these techniques, Enron's reported debt would have been $22.1 billion rather than $10.2 billion.」不僅如此：2001年11月19日，Enron 公布資產負債表上的債務約為 130億美元，但同一天卻告訴銀行，債務約為 380億美元 [57]。Powers 報告提到，其中很大一部分其實出現在財報附註中，只是寫得毫無用處：「these disclosures were obtuse, did not communicate the essence of the transactions completely or clearly, and failed to convey the substance of what was going on between Enron and the partnerships.」[56]。可取得的資訊與可理解的資訊，是兩回事。

回應：變動利益實體

制定美國會計準則的機構 FASB 在 2003年1月以第 46 號解釋作出回應，並於同年12月修訂 [57]。其核心概念是：光靠表決權控制，不足以決定由誰合併一個實體：「The voting interest approach is not effective in identifying controlling financial interests in entities that are not controllable through voting interests or in which the equity investors do not bear the residual economic risks.」[58]。權益資本不足以在沒有第三方支持下自行融資的實體，從此被稱為變動利益實體（VIE），而負責合併這些實體的是「primary beneficiary」：「the party that absorbs a majority of the entity's expected losses, receives a majority of its expected residual returns, or both」[58]。

這些規則並沒有消滅這類結構。SEC 幕僚早在 2005 年就觀察到，「many arrangements with potential VIEs were restructured such that the entity either would not be considered a VIE or such that no party would be required to consolidate the VIE.」[57]。而且有一個例外在改革後存活下來，也就是合格的證券化工具（QSPE）：「Transferors to qualifying special-purpose entities [...] do not consolidate those entities.」[59]。2007 年的槓桿，有很大一部分就是從這扇門通過的。

同一份 SEC 報告還衡量了另一個灰色地帶：租賃。在「全有或全無」的規則下，合約被設計成恰好落在必須於資產負債表上認列的門檻之下；SEC 幕僚估計，美國企業未出現在資產負債表上、只出現在附註中的營業租賃未來義務，約為 1.25兆美元 [57]。這個診斷後來廣為人知：「lease structuring to meet various accounting, tax, and other goals, has become an industry unto itself in the last 30 years.」[57]。

Enron 為解讀 AI 資料中心留下了三個教訓，全都關於結構，沒有一個關於詐欺。把投資移出資產負債表的動機，仍然是保護財務比率與信用評等。決定哪些項目列入資產負債表的規則仰賴判斷（誰擁有權力、誰吸收損失、合約期限多長），因此可以透過設計來繞過。而價值會隨所要保障的風險一起下跌的擔保，恰恰會在被動用時失效。

2008：離開資產負債表的，危機時會回來

負責調查 2008 年危機的美國國會委員會，也就是金融危機調查委員會（Financial Crisis Inquiry Commission，FCIC），在結論的開頭寫了一句話，應該會讓任何聲稱危機無法預測的人感到不自在：「We conclude this financial crisis was avoidable. The crisis was the result of human action and inaction, not of Mother Nature or computer models gone haywire.」[60]。本文把 2008 年當作嚴格意義上的證據。目的不是要說 AI 會造成同樣規模的危機，而是要說明哪些機制把一個產業的損失變成整個體系的損失，再到當前的週期裡尋找這些機制。

第一個機制是看不見的槓桿。根據 FCIC，五大投資銀行的營運狀況是：「extraordinarily thin capital. By one measure, their leverage ratios were as high as 40 to 1, meaning for every $40 in assets, there was only $1 in capital to cover losses. Less than a 3% drop in asset values could wipe out a firm.」[60]。而且其中一部分槓桿是看不見的：「the leverage was often hidden—in derivatives positions, in off-balance-sheet entities, and through "window dressing" of financial reports available to the investing public.」[60]。委員會用一句話總結了這個體系：「We had a 21st-century financial system with 19th-century safeguards.」[60]。

從 Enron 到商業本票

FCIC 把 Enron 之後的改革和 2007 年的槓桿直接連在一起。銀行把放款放進表外的資產擔保商業本票（ABCP）計畫，因此不必再為這些放款保留資本。Enron 事件之後，FASB 讓這條出路變得困難，市場隨即陷入停滯，銀行監理機關於是做出回應：2004 年，它們允許這些資產不計入資本計算，只對提供給這些計畫的流動性額度收取「up to 0.8%」的資本計提。「Growth in this market resumed.」[60]。這個教訓令人不安：會計規則收緊了，審慎監理規則卻放寬了，而這個結構繼續成長。

根據聯準會的資料，ABCP 市場在截至 2007年8月8日的那一週達到 1.2257兆美元。到了那年年底，規模是 7,745億美元 [61]。聯準會自己的經濟學家把當時發生的事形容為一場沒有銀行的銀行擠兌：「more than 100 programs (one-third of all ABCP programs) were in a run within weeks of the onset of the turmoil and the odds of subsequently leaving the run state were very low.」[62]。結構型投資工具（SIV）是這類計畫的一種變形，主要投資 AAA 評等的證券；這類工具共有 36 個，資產將近 4,000億美元，槓桿倍數為「an average of just under 14-to-1」。其中一個名為 Sigma 的 SIV，投資人損失超過 95%。「As of fall 2010, not a single SIV remained in its original form.」[60]。

表外的東西回到表內

對本文來說，2008 年的核心重點在於：危機來臨時，不在資產負債表上的承諾會怎麼樣。Citigroup 曾經出售「liquidity puts」，也就是承諾在沒有其他人願意購買時，買下擔保債務憑證（CDO）發行的商業本票，涉及金額 250億美元。2004 年以前，這些承諾完全不需要計提資本；之後則是 0.16%。FCIC 敘述，2007 年發生的事件將會「catapult the entire $25 billion in commercial paper straight onto the bank's balance sheet, requiring it to come up with $25 billion in cash as well as more capital to satisfy bank regulators.」[60]。

該銀行管理的 SIV 也遇到同樣的事，只是原因是壓力，而不是合約。2007年12月13日，Citi 宣布為這些工具提供一條支援額度，並表示：「As a result of this commitment, Citi will consolidate the SIVs' assets and liabilities onto its balance sheet under applicable accounting rules.」這些 SIV 的資產已經從 8月的 870億美元降到 490億美元 [63]。2008年11月，這家銀行把最後的 174億美元也納入了資產負債表 [60]。紐約聯邦準備銀行的 William Dudley 向 FCIC 承認，很少人了解「the backstops that banks were providing SIV conduits either explicitly or implicitly.」[60]。在銀行監理上，這後來被稱為介入風險（step-in risk）：發起機構沒有義務拯救這個工具，但還是會出手，因為不救的話，聲譽和業務要付出更高的代價。

虧損之前的追繳保證金

AIG 的案例呈現了另一個機制：在虧損出現之前就先到來的損失。這家保險公司的金融商品子公司，透過信用違約交換（credit default swaps），為 790億美元的 CDO 出售了違約保護 [60]。合約規定，如果這些證券的市值下跌，或 AIG 的信用評等遭到調降，就必須提供現金擔保品：「collateral calls could be triggered even if there were no actual cash losses in, for example, the super-senior tranches of CDOs upon which the protection had been written.」[60]。2008年9月12日，擔保品追繳金額合計 234億美元，AIG 已經付出 189億美元。政府最後承諾投入「more than $180 billion」，以免 AIG 倒閉的衝擊擴散開來 [60]。當擔保品按市值計價、合約又附有信評觸發條款時，流動性危機會在任何實際損失發生之前就到來。在 AI 這一輪週期中，最接近的對照是以 GPU 為擔保品的貸款，這類貸款的規模取決於晶片的帳面價值，如「最弱的環節」一節所示。

品質標章

在 FCIC 看來，信用評等機構是「essential cogs in the wheel of financial destruction」。2000 年至 2007 年間，Moody's 給了將近 4.5萬檔房貸相關證券最高評等，而「83% of the mortgage securities rated triple-A that year ultimately were downgraded」（那一年是 2006 年）[60]。問題不只出在方法上。那些證券有許多之所以評等很高，是因為結構把另一方的信用品質轉移到它們身上。「表外」一節會說明，史上為資料中心發行的最大一筆債券之所以評等很高，也是出於同樣的原因：評等反映的是保證，而不是建築物。

季末粉飾

Lehman Brothers 補上了這份清單的最後一個機制。根據破產案審查人 Anton Valukas，這家銀行利用內部稱為「Repo 105」的交易，在 2008 年第一季和第二季季末，把約 500億美元的資產暫時移出資產負債表。靠著這些交易，第二季季末揭露的淨槓桿比率是 12.1，而不是 13.9 [64]。Lehman 自己的員工把這種做法稱為「accounting gimmick」[64]。報告沒有判定這些交易在技術上是否違反會計準則；重點在於，一個市場密切關注的比率成了財務操作的目標。

錯向風險

串起所有這些案例的概念是錯向風險（wrong-way risk）。巴塞爾委員會這樣定義特定錯向風險：「Specific wrong-way risk arises when the exposure to a particular counterparty is positively correlated with the probability of default of the counterparty due to the nature of the transactions with the counterparty.」[65]。在危機後的檢討中，委員會指出，信用保險公司（monolines）是信用品質與曝險同向變動的交易對手，而這些交易「contributed to the losses during the crisis」。委員會舉的教科書範例，是交易對手以自家股票作為擔保品 [66]，這正是 Enron 那些 Raptor 工具的缺陷。2026年10月，IMF 把同樣的概念套用到 AI 的循環交易上，如下一節所示。

2009 年的改革

FASB 在 2009年6月以兩項準則回應：FAS 166 和 FAS 167。這兩項準則取消了合格證券化工具的例外，並把合併報表的量化測試改為質化測試。從此以後，一家公司如果具備以下條件，就必須合併一個變動利益實體（VIE）：「a. The power to direct the activities of a VIE that most significantly impact the VIE's economic performance b. The obligation to absorb losses of the VIE that could potentially be significant to the VIE or the right to receive benefits from the VIE that could potentially be significant to the VIE.」[67]。既有的結構也沒有倖免，而且效果有數據可查：在截至 2010年3月31日的那一週，美國商業銀行已經把原本在表外的工具的資產與負債共 3,778億美元納入資產負債表，其中大部分是信用卡證券化 [68]。今天的資料中心合資企業之所以留在表外，依據的正是這項規則，也就是「權力」測試與「承擔損失的義務」測試。

當年的影子銀行與今天的私募信貸

2008 年時，槓桿存在於後來被稱為影子銀行體系的地方。今天，銀行以外的信貸中擴張最快的是私募信貸基金。IMF 估計這個產業在 2023 年的規模約 2.1兆美元，並列出了它的脆弱之處：「relatively fragile borrowers, increased exposure of pensions and insurers to the asset class, a growing share of semiliquid investment vehicles, multiple layers of leverage, stale valuations, and unclear interconnections between participants.」[69]。2026年5月，聯準會記錄到美國私募信貸放款達 1.4兆美元，以及一件前所未有的事：2026 年第一季，半流動性信貸基金（永續型 BDC）獲准的贖回申請超過了資金流入，「the first time this has occurred since these vehicles were created」[70]。同一份報告也指出，最大型壽險公司的槓桿「well into the upper quartile of its historical distribution」[70]。

有兩座橋把這個市場和 AI 連在一起。第一座是資金的去向：BIS 的經濟學家計算出，科技公司在私募信貸直接放款中的占比「doubled to over 40% between 2020 and 2025」[71]。第二座是誰買下了風險。金融穩定委員會（FSB）在 2026年5月警告，私人評等（有時來自名氣不大的評等機構）正被用來吸引保險公司，因為保險公司要靠評等來符合資本規定：「The growing use of private ratings, sometimes from lesser-known providers, to facilitate investment by rating-reliant investors (including insurers) also warrants monitoring.」[72]。FSB 並總結：「Private credit remains untested to a prolonged economic downturn and so warrants close attention.」[72]。

類比失靈之處

兩者之間有重要的差異，略過不提，就等於用樂觀派對待 AI 的方式來對待 2008 年。2008 年時，風險在於靠存款和極短期債務融資的銀行，資產則是數百萬個家庭的房貸。今天的私募信貸主要以長期資本融資，IMF 自己在 2024 年也寫道：「Private credit loans are funded largely with long-term capital, mitigating maturity transformation risks. The use of leverage appears modest, as do liquidity and interconnectedness risks.」[69]。2026年4月，IMF 仍然認為超大規模雲端業者「thus far have financed investments robustly without materially weakening their financial positions」[31]。FCIC 內部也有不同意見書：部分委員不接受多數派的敘事，而由 Keith Hennessey、Douglas Holtz-Eakin 和 Bill Thomas 署名的不同意見書中，有一節的標題是「The dangerous imprecision of the term 'shadow banking'」[60]。

所以，2008 年的教訓不是歷史會以同樣的規模重演。它提供的是五個值得尋找的機制。表外承諾在危機時會回到表內，原因可能是合約，也可能是聲譽。與市價和信用評等連動的觸發條款，會讓損失提前到來。評等衡量的可能是另一家公司的保證，而不是資產本身。與風險一起下跌的擔保，在被動用時會失靈。市場緊盯的比率，則會成為財務操作的目標。接下來幾節會在 AI 的融資中逐一尋找這些機制。

循環：誰投資、誰賣、誰買

電信那一節說明了這個機制：Lucent、Nortel 和 Cisco 借錢給電信業者，讓它們購買自己的設備，設備商的營收隨著自己出資撐起的需求而成長。2025 年和 2026 年，這個機制以更大的規模、更多樣的形式回來了。晶片和雲端供應商入股 AI 實驗室，這些實驗室再和同一批供應商簽下採購承諾。其中有給買方的股票獎勵、租金保證、買下剩餘算力的承諾，以及隨產品交付而撥付的融資。

這個通用的設計可以畫成一個迴圈：

晶片或雲端供應商
   |  1. 投資實驗室，或為實驗室的債務提供保證
   v
AI 實驗室
   |  2. 與供應商簽訂採購承諾
   v
供應商的待履行合約（RPO）
   |  3. 未來營收與市值上升
   v
供應商募集更多資金，再次投資
   |
   +--> 回到步驟 1

這些交易

下表整理了有文件記載的主要交易。金額是各公司宣布或申報的數字，其中許多是「最高」金額、以交付為條件，或尚未定案：

資料來源見以下各段，其中會逐一詳述每個案例。

NVIDIA 與 OpenAI

2025年9月，NVIDIA 與 OpenAI 宣布簽署意向書，涵蓋至少 10 GW 的 NVIDIA 系統：「To support the partnership, NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed.」[73]。供應商的資金會隨著客戶安裝供應商的產品而逐步撥付。在下一份季報中，NVIDIA 提醒：「There is no assurance that we will enter into definitive agreements with respect to the OpenAI opportunity」[74]。這筆 1,000億美元的交易最後沒有談成。2026年2月，NVIDIA 在 OpenAI 1,100億美元的募資輪中投入 300億美元，「at a $730B pre-money valuation」，OpenAI 則取得了「3GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems」[75]。3月，Jensen Huang 表示，那筆 1,000億美元的機會大概已經「not in the cards」，因為 OpenAI 準備公開上市 [76]。

NVIDIA 投資的不只 OpenAI。2026年7月，NVIDIA 揭露「equity investments of $99 billion and equity investment commitments of $25 billion」，而它持有的未上市公司股權，帳面價值在一年內從 38億美元增加到 479億美元 [77]。在年報中，NVIDIA 自己描述了其中一部分投資：「These investments include AI model makers that purchase our products directly or through CSPs.」[78]。這些持股的未實現利得會計入獲利。客戶的估值一上升，賣東西給它們的那一方，獲利也會跟著上升。

同一季，NVIDIA 推出了它所謂的「a new business model」：AI 雲端業者購買它的基礎設施，NVIDIA 則承諾向它們租用算力，而依照合約，「the AI clouds can unilaterally stop providing to us and sell to third-party customers at more advantageous rates」。這些承諾合計 360億美元 [77]。實際上，供應商賣出設備，同時為買方保證一個需求下限。「表外」一節會說明下一步：針對 OpenAI 租約、最高 1,050億美元的保證。

AMD 與 OpenAI

AMD 選擇了另一種形式。2025年10月，AMD 給予 OpenAI 以每股 0.01美元認購最多 1.6億股 AMD 股票的權利，這些股票分批解鎖，條件是 OpenAI 購買 AMD GPU（最多 6 GW），以及 AMD 股價達到一系列目標，而這些目標「escalate to $600 per share for the final tranche」[79]。依我的計算，1.6億股約占 2025年10月底 AMD 流通在外股數的 9.8% [80]。AMD 財務長預估，這項合作會帶來「tens of billions of dollars in revenue for AMD」[79]。在會計上，這筆股票獎勵會變成營收的減項：「The grant date fair value of the warrant will be recognized as a reduction to revenue when revenue is recognized.」[80]。2026年2月，AMD 與 Meta 複製了這個模式，規模相同，目標也相同 [81]。AMD 還開始為一家合作夥伴的資料中心租約提供保證，最大曝險為 41億美元；公司表示，這類保證「may be issued in exchange for warrants」[82]。

Microsoft 與 OpenAI

Microsoft 同時是 OpenAI 的股東、供應商，也是 OpenAI 營收分成的受益者。2025年10月重組之後，它的持股估值約 1,350億美元，「roughly 27 percent」，而 OpenAI「has contracted to purchase an incremental $250B of Azure services」[83]。2026年1月，Microsoft 財務長表示「Approximately 45% of our commercial RPO balance is from OpenAI」[84]；依我的計算，約 2,800億美元。在截至 2026年6月的會計年度，Microsoft 認列了來自 OpenAI 的營收 241億美元，對 OpenAI 有 60億美元的應收帳款，相較之下，它從一開始累計投資了 119億美元。當年度的獲利還包括 OpenAI 持股帶來的 65億美元淨利得，公司表示這些利得「primarily relate to the dilution gain from the OpenAI Recapitalization」[3]。在更高的估值下被稀釋，為這位股東帶來了會計上的獲利。

Amazon、OpenAI 與 Anthropic

在 Amazon 這邊，資金在同樣兩家公司之間雙向流動。2026年2月，Amazon 承諾投資 OpenAI 500億美元，OpenAI 則把原本 380億美元的 AWS 合約擴大「by $100 billion over 8 years」，使用約 2 GW 的 Trainium 晶片算力 [85]。對 Anthropic，Amazon 在 2023 年至 2025 年間投資了 80億美元的可轉換債券，2026 年第二季又投資了 50億美元的股權，並提供最高 200億美元的融資額度 [7]。另一方面，Anthropic 宣布：「We are committing more than $100 billion over the next ten years to AWS technologies, securing up to 5GW of new capacity to train and run Claude.」[86]。

Amazon 的資產負債表顯示了估值的效果。2026 年第二季，由於 Anthropic 新的募資輪，Amazon 對其持有的 Anthropic 股票認列了「upward adjustments of approximately $50.5 billion」[7]。依我的計算，這大約是 AWS 同一季營業利益（166億美元）的三倍。Anthropic 則在 2026年5月以 9,650億美元的估值募得 650億美元，並表示其年化營收已超過 470億美元 [87]。成長是真的；投資方的利得，目前仍是帳面上的。

Anthropic、Google 與 Alphabet 年報裡的警語

Anthropic 和多家供應商重複同樣的設計。2025年11月，Anthropic 宣布它「has committed to purchase $30 billion of Azure compute capacity」，同時「NVIDIA and Microsoft are committing to invest up to $10 billion and up to $5 billion respectively in Anthropic」[88]。Google 是另一家大供應商。2025年10月，Anthropic 宣布將使用最多一百萬顆 Google TPU，這項擴張「worth tens of billions of dollars」[89]。Alphabet 的年報對這類交易的風險提出了最明確的警語，但沒有點名任何公司：「Our financial instruments also include debt and equity investments in companies with which we also entered into commercial arrangements at or near the same time.」接著又說：「if our investment in such companies becomes impaired, we may need to re-evaluate the accounting for the commercial arrangement, including reducing any remaining performance obligations.」[4]。也就是說，如果投資減損，待履行合約也可能縮水。兩者會一起下跌。

CoreWeave 與 SB Energy

同樣的設計也出現在規模較小的公司之間。2025年3月，OpenAI 與 CoreWeave 簽下第一份合約時，獲得了價值 3.5億美元的 CoreWeave 股票 [90]；到了 9月，雙方的合約總額達到約 224億美元 [91]。為 CoreWeave 供應全部 GPU 的 NVIDIA，買進了 20億美元的 CoreWeave 股票，並承諾到 2032 年為止，買下找不到客戶的剩餘算力 [92], [93]。2026年1月，OpenAI 和 SoftBank 向 SoftBank 集團旗下的資料中心開發商 SB Energy 投資 10億美元；OpenAI 與它簽下 1.2 GW 的租約，而 SB Energy「will become a major customer of OpenAI」[94]。在單一一筆交易中，OpenAI 投資了自己的房東，這個房東又成為它的客戶，而且隸屬於它最大股東之一的集團。

待履行合約值多少

這些交易出現在供應商的剩餘履約義務中。2026 年年中，Microsoft 的商業待履行合約為 6,780億美元 [3]，Oracle 為 6,640億美元 [18]，Google Cloud 為 5,140億美元 [14]，AWS 約 4,960億美元 [7]，CoreWeave 則為 1,037億美元 [93]。依我的計算，總額超過 2.4兆美元。另一端在很大程度上是同樣那幾家實驗室：光是 OpenAI 就提到八年內約 1.4兆美元的承諾，而它在 2026年9月底的年化營收約為 500億美元 [95], [96]。

BIS 的經濟學家有系統地衡量了這種循環性。2021 年至 2025 年間，「55.2% of incoming investments in AI firms came from other AI firms」，而 AI 公司之間的投資金額中，有 46.4% 同時涉及投資方與被投資方之間的供貨關係。他們指出的後果是：「circular relationships make reported demand partly endogenous to firms’ own financing decisions. For example, when a supplier finances a customer, part of the supplier’s revenue growth reflects its own capital investment, rather than organic final demand.」和電信業的類比也是他們自己提出的：「The parallel with the telecommunications boom of the late 1990s is instructive: upstream equipment vendors such as Lucent and Nortel financed network operators so that the operators could buy the vendors’ equipment」[51]。

IMF 在 2026年10月的報告中，為這個循環標上了數字：「Fourteen publicly disclosed transactions since 2023 account for approximately $300 billion of aggregate equity investments led by hyperscalers in financially weaker AI firms, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments through services or chip orders」。這些債務中約有 3,200億美元落在被投資公司身上，「including debt backed by investor guarantees, contingent financing, and other off-balance-sheet debt」。IMF 也點出了這個風險的名稱：「distress at key investee firms could present significant "wrong-way risk" to the AI ecosystem: Equity and debt investments in the firms depreciate in value precisely when future revenue streams tied to them are impaired.」[97]。IMF 的 Jason Wu 在記者會上總結：「If there are questions about the monetization in AI, the equity investments will deplete in value, the debt may be difficult to service, and the revenue commitments in the future may not materialize」，他稱之為「triple whammy」[98]。4月時，IMF 就已估計，2025年9月至12月間，循環中這些公司的股價上漲約 12 個百分點，其中 7 個百分點來自它們之間的相互強化 [31]。

並非每筆交易都是循環的

這個論點也有其限度。OpenAI 與 Broadcom 的交易涉及 10 GW 的自有加速器，並沒有供應商投資客戶的成分；雙方簽的是「term sheet」，也沒有公布金額 [99]。許多承諾以交付為條件，有些從未成為正式合約，例如 NVIDIA 的 1,000億美元。這降低了實際的風險，但也顯示頭條數字高估了確定的需求。此外，Microsoft 表示，在 2026年4月至6月這一季，其待履行合約的成長全部來自非前沿實驗室的客戶，如「什麼可能推翻這個論點」一節所示。

這個循環不是詐欺，每一筆交易也都有商業邏輯：實驗室需要有保障的算力，供應商則需要客戶來使用它們興建的算力。問題在於這些交易合起來造成的效果。它用供應商自己出資撐起的需求，灌大了供應商的營收和待履行合約。它把實驗室在私募市場的估值，變成投資方帳面上的獲利。它還把風險集中在少數幾個仍在燒錢的客戶身上。如果其中一家出事，其他公司會同時在三個方面受損：持股、債權，以及承諾的營收。

表外：Hyperion、xAI 與資料中心的財務工程

Enron 和 2008 年那兩節說明了表外工具如何運作，以及危機來臨時它們如何回到表內。2025 年和 2026 年，同一套工具箱又被拿來為 AI 融資：特殊目的實體（SPV）、由公司持有少數股權的合資企業、殘值保證，以及可續約的短期租約。與 Enron 的差別很重要：一切都已揭露、經過查核，而且就目前所知都合乎規則。相似之處也一樣重要：債務不在使用該資產的公司的資產負債表上，而實際上支撐還款的正是這家公司。

Hyperion：Meta 表外的一座 270億美元資料中心

2025年10月21日，Meta 宣布與 Blue Owl Capital 管理的基金成立合資企業，在路易斯安那州興建 Hyperion 園區。Blue Owl 的基金持有 80%，Meta 持有 20%，總開發成本估計約 270億美元。Meta 以「a four-year initial term with options to extend」的合約承租整個園區，並向合資企業提供「a residual value guarantee for the first 16 years of operations」[100]。根據年報，Meta 還獲得一筆 25.5億美元的一次性分配；新聞稿說的則是約 30億美元 [8]。

資金來自一檔債券。據 Bloomberg 報導，Morgan Stanley 安排了「over $27 billion of debt and about $2.5 billion of equity into a special purpose vehicle」，主要買家是 PIMCO；而根據一份評等報告，這批債券獲得 S&P 的 A+ 評等。Bloomberg 這樣概括這個設計：「Under the SPV structure, Meta is not borrowing the capital itself, the financing entity is.」[101]。發行人出現在 Blue Owl 旗下一檔基金向 SEC 提交的文件中：持有合資企業 80% 股權的 Beignet Investor LLC，發行了「$27,293,849 in senior secured notes with an interest rate of 6.581% and a maturity date of May 30, 2049」（金額以千美元為單位）[102]。發行後隔天早上，這批債券的報價就已經達到面額的 105% [101]。

Meta 的年報解釋了為什麼這一切都不在它的資產負債表上。租約從 2029 年開始，初始承諾約 123.1億美元。殘值保證的總上限約 280億美元，會隨時間遞減，而且「As of December 31, 2025, RVG payments are not probable and therefore, no liability has been recorded.」關於合資企業：「We do not have the power to direct the activities that most significantly impact the Venture's economic performance. Therefore, we are not the primary beneficiary and do not consolidate the variable interest entity (VIE). Our maximum exposure to loss related to the Venture was $45.95 billion as of December 31, 2025」[8]。2026年6月，最大曝險為 460.3億美元 [9]。

這個論據寫在 2026 年第一季的季報裡。根據 Meta 的說法，對合資企業績效影響最大的決策是「decisions pertaining to remarketing the data center campus, including but not limited to, negotiations with future lease tenants and individual property sales」，而這些決策不歸 Meta 決定 [103]。這是會計規則的一種可能解讀：規則要求同時具備主導最重要活動的權力，以及對重大損失的曝險，才需要合併 [104]。但是，只有在 Meta 退出時，把園區重新推向市場的決策才有意義，而每四年決定是否退出的，正是 Meta。Meta 也負責興建和管理這座園區。查核人員察覺到其中的微妙之處：Ernst & Young 把合資企業的合併問題列為關鍵查核事項，並寫道「Auditing the Company's determination of the primary beneficiary of the VIE was especially challenging」[8]。在 Wall Street Journal 上，專欄作家 Jonathan Weil 在標題中說得更直接：「AI Meets Aggressive Accounting at Meta's Gigantic New Data Center」，副標題則是：「Favorable treatment off the balance sheet hinges on some convenient assumptions.」[105]。

關於保證的規則有助於理解這個設計。根據美國的租賃會計準則，租賃負債只包含公司很可能要支付的殘值保證金額。Deloitte 指出，正因如此，經濟效益有很大一部分落在殘值保證上的合約，「may continue to yield favorable accounting results (e.g., reduced leverage) under ASC 842」[106]。而尚未開始的租約，要到開始日才會進入承租人的資產負債表 [107]。

依我的計算，Meta 揭露的最大曝險 460億美元，相當於它 2025 年底全部長期債務（587億美元）的 78% [8]。而且這個模式正在被複製。2026年7月，Meta 為第二個合資案簽署了獨家協議，地點在德州艾爾帕索，Meta 同樣持股 20%，並表示將提供「residual value guarantees with a maximum aggregate exposure of approximately $13 billion」[9]。

誰承擔了風險

在私募債券中，要知道誰是持有人並不容易。其中一部分會出現在美國註冊基金向 SEC 提交的報告，也就是 N-PORT 表格中。依我對這些資料的分析，以幾乎全為 2026年3月31日的持倉計算，有 317 檔基金持有 94.4億美元的 Beignet 債券，約占發行額的 35%。名稱中帶有 PIMCO 的申報機構旗下基金合計持有 65.7億美元，光是 PIMCO Income Fund 就持有 41.1億美元，約占發行額的 15%。專供保險商品（例如變額年金）投資的基金持有約 4.73億美元，另有五檔封閉式基金以這些債券作為擔保品借款 [108]。基金申報的平均價格，從 2025年12月面額的 105.7%，降至 2026年3月的 102.9% [109], [108]。發行額的另外三分之二，掌握在這些資料看不到的人手中，例如保險公司的自有投資組合、退休基金和私募基金。

股權部分也流到了個人投資人手中。持有 Beignet Investor 63% 股權、承諾出資 15.3億美元的那檔 Blue Owl 基金，是一檔不在交易所掛牌交易的不動產基金，以私募方式賣給「accredited investors」，也就是資產雄厚的個人和機構 [102]。單一公司單一園區的風險，被分散到固定收益基金、ETF、退休金計畫和富裕投資人手中，就像 2006 年的房貸被分散到數以千計的投資組合中一樣。

xAI：把 GPU 放進專屬的融資工具

Elon Musk 的 xAI 用了另一個版本。據 Bloomberg 報導，2025年10月，xAI 安排了一筆 200億美元的募資，其中包括一個特殊目的實體，「that will buy Nvidia processors and rent them to xAI」，股本約 75億美元，債務最高 125億美元；NVIDIA 將在股本部分投資最多 20億美元，xAI 則會租用這些晶片五年 [110]。在一篇解說文章中，Bloomberg 總結：「And xAI's exposure? Just a 5-year lease; nothing else.」[111]。2026年1月，Apollo 主導了 35億美元的融資，讓 Valor Equity Partners 旗下的一個工具買下 54億美元的基礎設施（包括 GB200 GPU），再租給 xAI 的一家子公司。Apollo 表示「NVIDIA invested in VCI as an anchor Limited Partner」[112]。晶片供應商是這個工具的合夥人，而這個工具買進晶片，是為了租給客戶。

Bloomberg 同一篇文章借 UBS 策略師 Matthew Mish 之口，拿現在和過去做比較：「During the dot-com era, most of the growth was financed with equity not debt. So when it burst, the impact in the economy was manageable,」以及「Now, there's a rapid growth of capex in AI firms driven by debt and it's starting to be kept off balance sheet.」[111]。

賣方提供的保證

這個設計最新的形式，是由供應商提供保證。2026年8月，NVIDIA 在向 SEC 申報的表格中、處理直接財務義務與表外義務的項目下，揭露了最高 1,050億美元的殘值保證，標的是 SB Energy 出租給一家 OpenAI 關係企業的租約，涵蓋俄亥俄州一座園區約 4.25 GW 的容量。如果 OpenAI 無力償債或停止付款，NVIDIA 要補足「any shortfall between the guaranteed minimum value of a lease (as set forth in the Agreements) and amounts recovered through a replacement lease or sale」。這些保證會在幾種情況下終止，其中之一是 OpenAI 取得「a satisfactory credit rating」[113]。據 CNBC 報導，雙方 7月開始談判時，談的是最高 2,500億美元的保證，可讓 OpenAI「on the strength of Nvidia's credit」為一座 10 GW 的園區舉債 [114]。作為交換，這座園區只會使用 NVIDIA 的基礎設施 [77]。也有一些緩和因素：這些保證分階段生效，第一階段預計在 2029 會計年度生效，並在 20 年的租期內逐步遞減，而且 OpenAI 承諾會償還 NVIDIA [77]。Jensen Huang 這樣描述目標：「We are securing long-lived infrastructure for NVIDIA compute so OpenAI can deploy the most productive AI factories that can be upgraded repeatedly with each new generation delivering more intelligence and better economics.」[115]。建築物可以用 20 年，晶片每一代都會更換，而為租約提供保證的，正是賣晶片的那一方。

Google 也對改建成資料中心的前比特幣礦商這麼做。它為新興雲端業者（neocloud）Fluidstack 的租約提供保證，讓這些礦商取得融資：TeraWulf 一案約 32億美元，附帶的認股權證可讓 Google 的持股達到約 14%；Cipher Mining 一案為 14億美元；至於 Hut 8 一案，則是一項「covers the lease payments and related pass-through obligations」的保證，涵蓋一份為期 15 年、金額 70億美元的合約，並由銀行提供最高達專案成本 85% 的融資 [116], [117], [118]。據 Bloomberg 報導，Google 把這些保證列為信用衍生性商品 [111]。

2026年8月，NVIDIA 把這個構想推向另一個規模。它與 Apollo、BlackRock、Blackstone、Brookfield、Goldman Sachs 和 KKR 簽署備忘錄，要建立能夠「mobilize over $500 billion of third-party capital」的融資平台，並把自家晶片定位為投資資產：「NVIDIA compute is an investable asset」[119]。據 CNBC 報導，NVIDIA 將可選擇為每筆貸款的 25% 提供保證。BlackRock 執行長 Larry Fink 做了本文一直在做的比較：「This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s」。KKR 負責數位基礎設施的主管則補充：「you can securitize it or effectively divide that risk and sell it to investors who want to participate anywhere in that stack」[120]。這些協議都是合作備忘錄，並沒有簽訂正式合約。

資產負債表上看不到的

這一輪週期中最大的幾家公司，已簽署但尚未開始的租約加總超過 1兆美元。這些租約都不在資產負債表上，因為會計規則要到租約開始時才認列：

資料來源是各公司的季報和年報 [3], [18], [9], [7], [5], [93]。依我的計算，總額約 1.15兆美元，是 Microsoft、Alphabet、Amazon、Meta 和 Oracle 在 2026 年年中合計金融債務（約 4,800億美元）的兩倍以上。這些是名目金額，要在最長 20 年的期間內支付，部分還取決於合約條件；它們不是債務。但它們是固定承諾，而 Moody's 表示，它把這些承諾視同債務 [17]。

投資與租賃之間的界線也在移動。2026年7月，Microsoft 把資料中心的估計耐用年限從 15 年延長為 25 年，並說明了影響：「more of our future datacenter leases will shift from finance leases to operating leases as a result of this update. Finance leases are included in capital expenditures while operating leases are not.」這樣一來，當年度資本支出的預估從約 1,900億美元降為約 1,750億美元，而投資計畫「unchanged」[13]。投資金額沒有變；市場關注的那個數字變小了。

監理機關為這個現象取了名字。2026年3月，BIS 寫道：「These arrangements amount to "shadow borrowing": obligations that are economically akin to debt but largely reside outside corporate balance sheets. By channelling sizeable private credit into AI infrastructure, these structures strengthen links between hyperscalers and non-bank investors such as private credit vehicles and insurers.」[121]。在 6月的年度報告中，BIS 補充說，資料中心的興建愈來愈常外包給那些再回租給超大規模雲端業者的業者，「on long-dated contracts with embedded exit clauses. The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times.」[122]。2026年10月，IMF 指出，未來融資需求的一個來源是「the eventual realization of sizable off-balance-sheet commitments」[123]。IMF 也提醒，與資料中心相關的證券化自 2018 年以來共 460億美元，預計在 2028 年達到 1,500億美元 [31]。

這些都不違法，而且今天的會計規範比 2001 年或 2007 年嚴格得多。2008 年留下的問題是另一個：資產貶值時，這些承諾會怎麼樣。Meta 的殘值保證，只有在 Meta 不續租、而且園區價值低於上限時才需要支付，也就是資料中心的需求下滑時。NVIDIA 的保證，只有在 OpenAI 不付款時才會被動用。這兩個觸發條件都取決於同一個事件，也就是 AI 的修正，而這個事件會同時壓低保證方的股票價值和現金流。債務並沒有消失。它換了地址，而且仍然綁在同一個風險上。

折舊的物理：GPU 不是不動產

每筆債務背後都有擔保品，而擔保品決定了出問題時債權人能收回多少。在 AI 資料中心裡，擔保品分成三層，壽命差異很大。建築物和土地可以用幾十年：Alphabet 把資料中心和辦公室「over periods of seven to 40 years」提列折舊 [4]。電力和冷卻設備大約可用十年：CoreWeave 把「Data center equipment」分 8 到 12 年折舊，Amazon 則把支撐其物流中心和資料中心的「heavy equipment」耐用年限從 10 年延長為 13 年 [124], [125], [6]。而伺服器，也就是大部分資金所在之處，只有幾年的價值。本節要問的是：究竟是幾年。

供應商的節奏

2024年6月，NVIDIA 執行長 Jensen Huang 宣布了新品推出的節奏：「Our company has a one-year rhythm. Our basic philosophy is very simple: build the entire data center scale, disaggregate and sell to you parts on a one-year rhythm, and push everything to technology limits」[126]。在那之前，世代之間的跨越要兩年：2024年3月 Blackwell 的發表新聞稿，把前一代架構 Hopper 描述為「launched two years ago」[127]。照這個宣布的節奏，今天買一台伺服器、分六年折舊的人，在供應商推出五或六次年度新品之後，這台伺服器仍然會留在他的資產負債表上。

供應商自己就說明了每一代舊產品落後多少。Blackwell 發表時，NVIDIA 表示新平台能以「at up to 25x less cost and energy consumption than its predecessor」執行一兆參數的模型 [127]。在 GB200 NVL72 系統的產品頁上，與 H100 的比較是：語言模型推論「30x」，能源效率「25x」[128]。2026年3月，下一代 Vera Rubin 號稱能做到「one-tenth the token cost relative to NVIDIA Blackwell」[129]。2026年8月，NVIDIA 又引用顧問公司 SemiAnalysis 的測試，指出 Vera Rubin NVL72 能提供「30x higher throughput per megawatt and 45x lower token costs than NVIDIA GB300 NVL72」[130]。這些是賣方的數字，是在賣方選定的條件下測得的，閱讀時也應該記住這一點。但重要的是它傳達的訊號：賣方正在告訴市場，前一年買的設備生產出來的 token，比新設備貴得多。

NVIDIA 網站上的 H100 頁面甚至為這個差距標上了價格。根據該頁面引用的 SemiAnalysis 2026年4月資料，H100 的推論成本約為每百萬個 token 0.09美元，B200 則約為 0.02美元：「roughly 4.5x cheaper than H100」[131]。這項比較在兩顆晶片上使用了不同的軟體和不同的每位使用者速度，因此並不精確。同一個網站也記錄了另一面：Blackwell 本身的每 token 成本在發表後降到原來的五分之一，靠的只是軟體，「driven by continuous software optimization」[132]。GPU 的持有者能從軟體最佳化中獲益，但當下一代產品推出、所有人的 token 價格都下跌時，就會蒙受損失。

經濟機制很簡單。當供給跟得上需求時，GPU 每小時的租金，往往會被拉向以現有最有效率的設備生產 token 的成本。舊 GPU 還能運作，但它只能賺到 token 市價與自身成本之間的差額，而自身成本包括它多耗的電力。所以 GPU 的經濟耐用年限，和它的實體壽命不是同一回事。

實體耗損

當機器在極限下運轉時，實體壽命也不長。Meta 發表了少數詳細記錄大規模訓練故障的報告之一。在 Llama 3 405B 為期 54 天的預訓練期間，在一個最多有 1.6萬顆 H100 GPU、「each running at 700W TDP with 80GB HBM3」的叢集上，共發生 466 次中斷，其中 419 次是非預期的：「Approximately 78% of the unexpected interruptions are attributed to confirmed hardware issues, such as GPU or host component failures, or suspected hardware-related issues like silent data corruption and unplanned individual host maintenance events.」[133]。在原因統計表中，148 次中斷來自故障的 GPU，72 次來自 HBM3 記憶體故障 [133]。自動化機制應付得來：有效訓練時間「higher than 90%」。但一批以這種速度損耗元件的機器，老化速度比資料庫伺服器更快。

建築物也會老化

建築物可以用幾十年、只需更換晶片的想法，碰上了功率的問題。一顆 H100 的功耗最高 700 W；一台配備八顆 Blackwell GPU 的 DGX B200 伺服器則是「~14.3 kW max」[131], [132]。GB200 NVL72 機櫃已經採用液冷 [128]。NVIDIA 的工程師在 2025 年寫道，目前機櫃內使用的配電方式，在每櫃超過 200 kW 時「begins to hit physical limits」，並宣布推出 800 伏特直流電架構，「to support 1 MW IT racks starting in 2027」。在現行標準下，單一 1 MW 機櫃「requires up to 200 kg of copper busbar」[134]。國際能源總署（IEA）估計，AI 伺服器的功率密度在 2020 年至 2025 年間「increased by 11 times」，到 2027 年還會再增加為四倍 [22]。為某一代晶片設計的建築物，可能需要在電力和冷卻上大規模施工，才能容納下一代。

會計走向了相反的方向

就在晶片的節奏加快之際，大公司的會計卻反其道而行：延長了伺服器的耐用年限。耐用年限愈長，成本就分攤到愈多年，每一年的獲利也就愈高。這些變更都寫在年報裡：

理由各不相同。Microsoft 提到「investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology」[136]。Meta 在 2022 年的說法是「expected longer refresh cycles in our data centers」[137]。Amazon 是例外，也是最能說明問題的一家。2025 年，它縮短了部分伺服器的耐用年限，並說明了原因：「The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.」[6]。2024 年底，Amazon 已經決定提前汰除部分伺服器，並認列了「approximately $920 million of accelerated depreciation and related charges」[125]。Microsoft 在 2026 年的年報中，開始把伺服器與網路設備的耐用年限描述為「two to six years」[3]。

涉及的金額隨著投資一起增加。Meta 的伺服器與網路設備折舊，2025 年為 133.6億美元，2023 年則為 73.2億美元；這些資產的帳面總額在一年內從 684億美元增加到 980億美元 [8]。IMF 為整個產業做了計算。根據十家超大規模雲端業者揭露的折舊推算，全部固定資產（包括建築物）的隱含耐用年限約為七年，但 GPU 和先進晶片「could face obsolescence within two years」。IMF 也測試了較短的耐用年限：「Under a scenario of three-year useful life of capital and assuming hyperscalers fulfill these additional financing needs by issuing debt, their debt levels could rise from the present $800 billion to more than $1 trillion.」根據 IMF 的模型，這些公司的信用違約交換（CDS）利差會上升約 60 個基點，「a significant increase considering that these spreads are currently in the range of 20 to 160 basis points」[31]。IMF 把過時視為經營風險，認為它對金融穩定沒有一階效果。但對以這些資產為擔保放款的人來說，這是核心問題。

Burry、NVIDIA 與租金價格

2025年11月，投資人 Michael Burry 把這場討論帶到了市場的中心。他寫道「Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era」[138]，又寫道「Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment. Yet this is exactly what all the hyperscalers have done」[139]。據 Business Insider 報導，他估計 Meta、Google、Oracle、Microsoft 和 Amazon 在 2026 年至 2028 年間會少計約 1,760億美元的折舊，並估計到了 2028 年，Oracle 的獲利會被高估約 26%，Meta 則約 20%，「without elaborating on his calculations」。曾放空 Enron 的投資人 Jim Chanos 也持相同看法：「the chips basically depreciate over two years」[139]。

「詐欺」這個詞是 Burry 用的，必須小心解讀。耐用年限的變更都寫在年報裡，並附有以美元計的影響，如上表所示；爭論的是估計，而不是隱匿。NVIDIA 以一份給分析師的備忘錄回應。據 CNBC 報導，NVIDIA 主張客戶把 GPU 分四到六年折舊，是「based on real-world longevity and utilization patterns」，又說 2020 年推出的 A100「continue to run at high utilization rates and retain meaningful economic value well beyond the two to three years claimed by critics」。Burry 反駁：「I am not claiming Nvidia is Enron. It is clearly Cisco.」[140]。

到目前為止，租賃市場證明雙方各對了一部分。到 2025 年為止，價格是下跌的。那年 6月，AWS 把搭載 NVIDIA GPU 的執行個體價格調降最多 45% [141]；在 AWS 的公開價目表上，一台配備八顆 H100 的機器每小時價格，從 2023年8月的 98.32美元降到 2026年10月的 55.04美元 [142], [143]。不過，到了 2026 年，預留算力的租金上漲了。在 AWS 的容量區塊方案中，H100 的價格從 2025年6月每 GPU 小時 3.93美元，漲到 2026年10月13日的 5.97美元，漲幅 52% [144], [145]。在新興雲端業者 Lambda，H100 的隨需價格從 2025 年的每小時 2.99美元，漲到 2026年4月起的 3.99美元；租用數週的叢集價格，則從 2025年11月的 2.40美元，漲到 2026年10月的 5.54美元至 6.16美元 [146], [147]。據 CNBC 報導，CoreWeave 有一批 2022 年的 H100 在合約到期後空出來，重新出租時的價格是「at 95% of their original price」[148]。

最常被引用的解釋是短缺。The Verge 報導指出「The price to rent old chips has been rising, and Silicon Data projects that it will continue rising through 2028」，該網站採訪的一位分析師則把漲價歸因於「a major shortage of inference chips」[149]。IEA 指出的 HBM 記憶體短缺，也指向同一個方向 [22]。2026年8月，衍生性商品交易所 CME 宣布推出以 H100 和 B200 租金指數為標的的期貨合約，並以「sharp swings in compute prices」為理由 [150]。在 2026年10月一篇為 GPU 耐用性辯護的文章中，NVIDIA 引用第三方根據轉售價格所做的估計：H100 系統的耐用年限為五到六年，GB300 NVL72 則為九到十年 [151]。

這是反駁快速過時論最有力的論點，必須認真看待。但是靠短缺撐起來的價格，並不等於長期價值。NVIDIA 同一篇文章寫道：「Power is the binding constraint on an AI factory. This makes tokens per second per megawatt the number that governs earning capacity.」文章還說，與 GB300 相比，Vera Rubin 提供「over 30x higher throughput per megawatt」[151]。在電力有限的情況下，把舊機櫃換成新機櫃，同樣一百萬瓦電力的產出就會成倍增加。等晶片短缺過去，汰換舊機器的壓力就會回來，Burry 的問題也會跟著回來。

為什麼這對放款人很重要

下表只是算術：以直線法折舊、在伺服器上投資 100美元為例。它顯示依所選的耐用年限不同，三年後還有多少成本留在資產負債表上：

如果耐用年限是六年，一顆 GPU 用滿三年時，仍有一半的成本留在資產負債表上，而且至少已經有兩代更新的架構在市面上販售。如果這顆三年 GPU 的市價遠低於原價的一半，差額就是還沒有出現在損益中的損失。對以 GPU 作為貸款擔保品的人來說，這份擔保品比看起來的要少。以出租 GPU 為業的 CoreWeave，在自己的公開說明書中描述了這項風險：「Any changes to the significant assumptions underlying our estimates or to the estimates of our components’ useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and future prospects.」[124]。

另一方也有很好的論點。不再適合訓練前沿模型的 GPU，可以再花好幾年執行推論或要求較低的工作。軟體最佳化能提高同一顆晶片的產出，NVIDIA 自己的資料就顯示了這一點。Alphabet 也提醒，耐用年限的選擇依據的是「historical asset performance, expected technology advancements, and our future infrastructure deployment plans」[4]：這不是隨便定的數字。結論並不是這些公司算錯了。而是支撐 AI 相當大一部分債務的擔保品，經濟價值流失的速度比建築物快，比輸電線路快，而且在好幾種結構中，比債務本身還快。這是泡沫表格中的第五個標誌，而賣晶片的人和為晶片記帳的人都證實了它。

能源：可能變成擱淺資產的瓶頸

AI 資料中心首先是一個電力連接點。晶片每年都在換代，但為它們供電的變電所、輸電線路和發電廠，設計壽命是 20、30 或 40 年。這和上一節是同一種期限錯配，只不過這次落在科技公司的資產負債表之外。為資料中心興建電力基礎設施的一方，要承擔需求在工程回本之前就消失的風險。一旦發生這種情況，剩下的就叫做擱淺資產：一項仍然矗立在那裡、卻產生不了當初據以興建之營收的投資。

AI 消耗多少電力

國際能源總署估計，資料中心在 2024 年消耗了 415 TWh，「around 1.5% of the world’s electricity consumption」。美國占其中的 45%，中國占 25%，歐洲占 15%。在該機構的基準情境中，用電量「is set to more than double to around 945 TWh by 2030」，略高於日本目前的總用電量。在美國，到 2030 年為止的用電需求成長，將近一半來自資料中心；到了這個十年的尾聲，美國「is set to consume more electricity for data centres than for the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined.」[152]。2026年4月的更新證實了這條軌跡：資料中心的用電量在 2025 年成長了 17%，專供 AI 使用的資料中心，用電量則成長了 50% [22]。

就美國而言，勞倫斯柏克萊國家實驗室（LBNL）計算出，資料中心在 2023 年消耗了 176 TWh，占全國用電量的 4.4%。對 2028 年的預測介於 325 到 580 TWh 之間，相當於預估用電量的 6.7% 到 12.0% [153]。高情境幾乎是低情境的兩倍。為這個市場興建發電廠或輸電線路的人，必須在這個區間裡選定一個數字，然後與它共處數十年。同一份報告假設 AI 伺服器和傳統伺服器一樣耐用，在 2023 年約為 5.7 年：「For AI accelerated and nonaccelerated servers, the same lifetime assumptions are used.」[153]。就連能源預測，也帶著上一節所質疑的耐用年限假設。

併網排隊

限制不只在於發電，也在於連上電網。國際能源總署估計，如果不採取措施疏通電網，「around 20% of planned data centre projects could be at risk of delays」。在先進經濟體，一條新輸電線路需要四到八年，而且「wait times for critical grid components such as transformers and cables have doubled in the past three years」[152]。國際貨幣基金則提到，併網時程「now stretching from two to seven years」[31]。

美國的併網排隊反映出電網的步調。根據 LBNL，「As of the end of 2025, over 2,060 gigawatts (GW) of total generation and storage capacity were actively seeking connection to the grid.」以 2025 年完工的專案來看，從提出申請到開始運轉，時間的中位數超過五年。而且「Only 13% of capacity that submitted interconnection requests from 2000-2020 had reached commercial operations by the end of 2025; 75% of that capacity had been withdrawn」[154]。這條隊伍排的是發電廠和電池，不是資料中心，但它顯示了併網申請的下場：大多數從未真正動工。NVIDIA 自己也在 2026年3月承認了這個問題：「Energy is now the biggest bottleneck for AI infrastructure buildouts, with over $300 billion in equipment backlogs and more than 200 gigawatts of projects waiting in U.S. interconnection queues.」[155]。

渦輪機與反應爐

電網沒有空間，許多專案便轉向自行發電，而且幾乎都是燃氣發電。GE Vernova 是最大的渦輪機製造商之一，它表示，其燃氣設備的訂單積壓加上產能預留，在 2025 年第四季從 62 GW 增加到 83 GW，在 2026 年第二季從 100 GW 增加到 116 GW，並預期年底至少達到 125 GW。該公司的生產在 2026 年第三季達到每年 20 GW 的速度，預計 2028 年提高到 24 GW，目標是 2030 年達到 30 GW [156], [157]。依我的計算，這 116 GW 已簽約或已預留的設備，相當於目前將近六年的產量。國際能源總署記錄到「A 70% surge in gas turbine orders in 2025」，並估算資料中心廠區內的燃氣發電必須比需求超額配置 30% 到 70%，才能確保供電 [22]。

另一項押注是核能，其合約期限遠長於任何晶片的壽命：

Crane 的重啟還獲得美國能源部 10億美元的貸款 [162]，等於把這一輪週期的部分風險放上了政府的資產負債表。對賣電的一方來說，20 年的合約是支撐這筆投資的保障。對買電的一方來說，這是一筆長達二十年的固定成本，為的是替五或六年就折舊完畢的設備供電。

需求消失時誰來買單

有些美國配電公司已經針對這項風險要求擔保。在俄亥俄州，AEP 於 2024 年與州監管機構（PUCO）的幕僚、消費者權益代表機構及其他各方達成協議。這項協議「requires large new data center customers to pay for a minimum of 85% of the energy they say they need each month – even if they use less – to cover the cost of infrastructure needed to bring electricity to those facilities.」[163]。監管機構在 2025年7月核准了這套費率，最低用電量的級距以契約容量的 85% 為上限。初始合約涵蓋最長四年的爬坡期，再加上八年。信用評等不高的客戶，必須為合約中全部最低付款金額的 50% 提供擔保。客戶要等到爬坡期結束後的第五年過後才能退出，而且要支付 36 個月的最低費用。此外，承諾書「requires customers to reimburse Ohio Power 100% of the buildout costs if the customer cancels — or delays the project by more than 12 months — prior to the target energization date.」[164]。在德州，根據 AEP 的說法，規範大型負載併網的州法 SB 6，目的是「minimizing the potential for stranded infrastructure costs」[165]。

這是任何審慎債權人的邏輯：配電公司無法確認承諾的需求是否真的會出現，所以要求做出承諾的一方提供擔保。風險並沒有消失。它轉移到資料中心的所有者身上，而所有者可能是一家特殊目的實體，就像「表外」一節談到的那些。如果配電公司收不到錢，成本就會回到其他用戶的電費帳單上。國際能源總署描述了這項風險：資料中心「often oversize their grid connections initially」，而資料中心建得快、能源投資跟得慢，兩者的落差「could raise electricity prices in some places if unmitigated.」[22]。

該機構也把能源的步調和信貸的步調連在一起：「Data centre investments have grown too large to be funded from company balance sheets alone, and large amounts of funding from capital markets will be critical for their buildout. As a result, the pace of data centre growth, and the resulting increase in energy consumption, will be sensitive to market sentiment, including expectations for returns on investment in data centres and AI deployment, as well as to broader macroeconomic and financing conditions.」[22]。換句話說，讓渦輪機和輸電線路顯得值得興建的電力需求，取決於信貸市場的情緒，而這種情緒可能在幾週內就改變。

電網端的物理

還有一個不會出現在試算表上的工程問題。訓練叢集的耗電並不穩定。Meta 指出，在訓練 Llama 3 時，數萬顆 GPU 可能同時增加或減少耗電，例如在等待檢查點時，而這「can result in instant fluctuations of power consumption across the data center on the order of tens of megawatts, stretching the limits of the power grid.」[133]。與此同時，國際能源總署提醒，AI 資料中心「is 10 times more capital-intensive than an aluminium smelter, which means curtailing its operations to provide flexibility to the grid is very costly.」[152]。資料中心必須幾乎不停機地運轉，才能償付投入的資本，而電網則必須吸收其他產業不會產生的波動。

效率是一把雙面刃。根據國際能源總署，「Software and hardware advances have resulted in the energy use per AI task dropping by at least an order of magnitude annually in recent years.」[22]。每項任務都變得更便宜，這可能讓使用量成倍增加。但同樣的效率也會減少完成一定工作量所需的電力，而且如上一節所述，會壓低舊機器的價值。

能源是 AI 週期中最像鐵路和光纖的部分：壽命很長的資產，在需求出現之前就已建成，即使投資人賠錢，也依然存在。2002 年，閒置光纖變成了接下來十年的廉價頻寬。閒置的變電所和渦輪機同樣會找到別的用途。問題在於這段期間由誰買單：如果設有最低用電量的費率發揮作用，是資料中心的所有者；如果所有者倒閉，是它的債權人；如果規則失靈，則是其他電力用戶。

最弱的環節

信用危機很少從最強的機構開始。2007 年，最先倒下的是那些依賴短期融資、擔保品又不斷貶值的投資工具和基金 [60]。在這一輪 AI 週期中，主要的超大規模雲端業者資產負債表都很穩健：根據英格蘭銀行，「the AI hyperscalers all have a credit rating of AA- or higher」[166]。風險集中在三類公司：簽下遠超過自身營收之承諾的 AI 實驗室、靠舉債轉型為運算容量供應商的傳統企業，以及新興雲端業者，它們出租 GPU，而這些 GPU 是用以 GPU 本身為擔保的債務買來的。資料最齊全的例子是 OpenAI、Oracle 和 CoreWeave。

OpenAI：上兆的承諾，數百億的營收

OpenAI 的成長是真實的。財務長 Sarah Friar 在 2026年1月寫道，年化營收依序為「$2B ARR in 2023, $6B in 2024, and $20B+ in 2025」，與運算容量同步成長，後者在同一期間從 0.2 GW 增加到約 1.9 GW [167]。年化營收反映的是年底的營收速度，而不是全年營收：根據 CNBC，2025 年的營收為 131億美元。2026年3月，該公司表示自己正「generating $2B in revenue per month」[168], [169]。使用量也是真實的：2026年2月，ChatGPT 擁有「more than 900M weekly active users」，以及超過 5,000萬名訂閱用戶 [75]。依我的計算，每週使用者中約有 5.5% 付費。

問題在於承諾的規模。2025年11月，Sam Altman 寫道：「We are looking at commitments of about $1.4 trillion over the next 8 years」[95]。依我的計算，這相當於平均每年 1,750億美元。根據 CNBC，2026年9月底，OpenAI 向投資人報告的年化營收約為 500億美元，不到這個數字的三分之一；前一個月流傳的 680億美元，則計入了來自合作夥伴的總額營收 [96]。OpenAI 不公布經過審計的財務報表。外界對它現金狀況的了解，來自寄給股東、再由 The Information 網站報導的文件：該公司曾預計到 2029 年將燒掉 1,150億美元 [170]，之後又預估「$111 Billion More Cash Burn Through 2030」[171]，而到了 2026年4月，預期它將燒掉「more than $200 billion before it starts generating cash」[172]。2026 年第一季，該公司燒掉了 37億美元，「more than half its $5.7 billion in revenue」[173]。

資金來自規模愈來愈大的募資輪。OpenAI 在 2025年3月以 3,000億美元的估值募得 400億美元，一年後又以 8,520億美元的估值完成 1,220億美元的募資。它首次透過銀行，開放個人投資者參與這一輪募資，「raising over $3 billion from individual investors」，並宣布將被納入 ARK Invest 旗下的 ETF [169]。2026年6月，它向美國證券交易委員會秘密遞交了首次公開發行的申請 [174]。股東的部分資金也來自舉債。SoftBank 總共投資了 646億美元，取得 OpenAI 約 13% 的股權；它為 2026 年那筆 300億美元的注資付款時，先是動用銀行的過橋貸款，最後一期則靠發行約 111億美元的債券支應，這批債券的評等為 BB+，低於投資級，年利率最高達 9.75% [175], [176], [177]。CNBC 在 5 月訪問的策略師，已經把 SpaceX、OpenAI 和 Anthropic 這一連串排隊中的超大型上市案，解讀為市場見頂的訊號。其中一位做了比較：「Back in 1999, we saw the same kind of thing where people were just rushing to get these IPOs out.」[178]。

OpenAI 的資產負債表很輕。根據看過該公司財務報表的 The Information，截至 2026年3月31日，它帳上「zero debt」[179]，約 47億美元的銀行信用額度也分文未動 [169]。債務是存在的，只是記在別人的資產負債表上：在 Oracle 和開發商簽下的資料中心租約裡，在新興雲端業者的合約裡，也在 NVIDIA 的擔保裡，「表外」一節對此有詳細說明。NVIDIA 自己也解釋了為什麼需要為客戶提供擔保：「AI model makers have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity to secure the AI infrastructure necessary to grow.」[77]。這正是供應商融資（vendor financing）的定義：供應商提供市場不願給客戶的信用。

2025年11月，財務長一度談到由聯邦政府為晶片融資提供「backstop」，當天就改口：「I used the word 'backstop' and it muddied the point」[180]。Altman 表示，OpenAI 的資料中心既沒有、也不想要政府擔保，並寫道：「But we of course could be wrong, and the market—not the government—will deal with it if we are.」[181]。D.A. Davidson 的分析師 Gil Luria 由此得出結論：「If sanity prevails and this federal guarantee doesn't materialize, it could become clear that OpenAI cannot live up to its extravagant commitments. In which case we would view the companies propped up by OpenAI promises as the most vulnerable,」並點名 Broadcom、AMD、Oracle、CoreWeave 和 Oklo [182]。根據 CNBC，Wall Street Journal 在 2026年4月報導，OpenAI 沒有達成自己設定的營收和用戶目標，財務長也擔心公司支付未來運算合約的能力。該公司稱這篇報導「ridiculous」，但 Oracle 和晶片製造商的股價當天都下跌了 [183]。2026年10月8日，也就是 CNBC 證實 500億美元年化營收的那一天，NVIDIA 股價下跌 3%，Oracle 下跌將近 6%，CoreWeave 下跌將近 8% [96]。傳導管道已經清晰可見。

Oracle：從軟體公司變成 AI 的債權人

沒有一家大公司的樣貌改變得這麼劇烈。在截至 2025年5月的會計年度結束時，Oracle 的剩餘履約義務（英文縮寫為 RPO）為 1,378億美元。三個月後，這個數字變成 4,550億美元，「up 359%」。時任執行長 Safra Catz 表示：「It was an astonishing quarter」[184]。下表列出隨之而來的變化，單位為十億美元：

這些數字來自該公司向美國證券交易委員會提交的報告與公告；2025年3月至5月那一季的數字，是以整個會計年度減去前九個月得出 [185], [12], [10], [186], [18]。用 Oracle 自己的話來說，2026 會計年度的自由現金流為負 237億美元 [187]。而在 6,640億美元的 RPO 中，只有約 13% 預計會在接下來 12 個月內轉為營收 [18]。

Oracle 沒有揭露 RPO 中各客戶分別占多少，但在年報中提出警示：「The economic returns on these investments are dependent on customer demand and the ability of our key customers to meet their contractual obligations.」[10]。分析師算過這筆帳：根據 D.A. Davidson，2025年11月的 5,230億美元 RPO 中，至少有 3,000億美元來自 OpenAI [188]；根據 Bank of America，2026年6月時有超過一半的 RPO 來自 OpenAI [189]。被問到 OpenAI 是否付得起每年 600億美元的雲端費用時，共同執行長 Clay Magouyrk 回答「of course」[190]。

為了支付工程費用，Oracle 轉向市場籌資。2025年9月，它分六個系列發行了 180億美元的債券，期限最長的一檔於 2065 年到期，票面利率 6.10% [191]。2026年2月，它又分八個系列發行了 250億美元，期限最長的一檔至 2066 年到期，票面利率 6.85%，另外還發行了 50億美元的可轉換特別股 [192]。2026年6月至8月，它在市場上出售了 200億美元的股票 [186]。信用評等落在 Baa2 和 BBB，展望皆為負向；CNBC 談到 Moody's 的評等時，形容這是「just two notches above junk status」[17]。根據該公司自己的條款說明書，兩次發行之間，相對於美國公債的溢酬上升了：十年期債券從 1.05 個百分點升至 1.45 個百分點，40 年期債券從 1.37 個百分點升至 1.95 個百分點 [191], [192]。

在表外，承諾的規模更大。已簽署但尚未開始的租約，「substantially all related to data center arrangements」，租期 15 到 19 年，金額從 2025年5月的 434億美元增加到 2026年8月的 2,880億美元 [12], [18]。這是該公司全部金融債務的兩倍多。Oracle 還為它的一家出租人所借的貸款提供了最高 33億美元的擔保 [10]。

部分投資開始由客戶支付。在 2026年6月至8月這一季，Oracle 收到 114億美元的預付款，「that included a significant financing component」，也就是內含利息 [18]。該公司隨後創造了一項不在會計準則規範內的新指標「net cash outlay for capital expenditures」，從資本支出中扣除這些預付款，以及與設備製造商之間被歸類為融資的部分現金流。依照這項指標，2026 會計年度的投資為 477億美元，而依會計帳列的資本支出為 557億美元；下一季則是 180億美元對 285億美元 [186]。Oracle 沒有說明是誰預付了這筆錢。在它自己列出的雲端客戶中，有 AMD、Meta、NVIDIA、OpenAI、TikTok 和 xAI [193]，而 OpenAI 要靠新一輪的投資才付得出自己的帳單。

市場有了反應。2026年6月，Oracle 股價自 2025年9月的高點累計下跌約 55%，並出現了「the steepest weekly drop since a 20% plunge in August 2001, during the depths of the dot-com bust」[194]。7 月，資產管理公司 Apollo 的 Torsten Slok 對 CNBC 表示：「CDS for Oracle is at the same level it was at in 2008.」[195]。2025年9月那批債券的買家對該公司提起訴訟，指控它沒有揭露自己將需要多得多的債務，這項指控尚未經法院判決 [196]。2026年9月，Oracle 向新墨西哥州 Project Jupiter 的開發商發出不可抗力通知，以免承擔更高的成本；根據 CNBC 引述的 Financial Times，與這座資料中心相關的 180億美元債務，交易價格已經「at stressed levels」[197]。

CoreWeave 與新興雲端業者：以 GPU 作擔保

CoreWeave 向 AI 實驗室和超大規模雲端業者出租 NVIDIA GPU 的運算容量。2026年6月30日，它有 351億美元的債務，對應 50億美元的股東權益，約為七倍；此外還有 163億美元的租賃負債，以及列在表外、355億美元已簽署但尚未開始的租約 [93]。2026 年第二季，營收為 26億美元，比一年前增加 112%，但淨利息支出達 6.4億美元，約為營收的四分之一，虧損則為 6.26億美元。上半年，該公司投資了 141億美元，營運產生的現金為 37億美元 [93]。

把 CoreWeave 和折舊那一節連在一起的細節，就在貸款合約裡。信用額度的計算方式是「based upon the depreciable cost of computing equipment」，其中一條額度則是「based upon the depreciable cost of graphics processing unit servers」[93]。也就是說，公司能借多少錢，取決於 GPU 的帳面價值，而帳面價值又取決於所選擇的耐用年限。有效利率介於每年 7% 到 15%，部分債務放在特殊目的子公司裡，這些子公司在破產時與母公司隔離，並以 182億美元的長期資產作為擔保 [93]。

2026 年，這類信貸成了一種資產類別。3 月，一條 85億美元的信用額度獲得 Moody's 的 A3 評等，CoreWeave 稱之為「the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract」[198]。8 月，另一條額度明白呈現出期限錯配：「the DDTL 5.5 Facility carries an approximate five-year maturity while its underlying customer contracts average approximately three years in length. By financing these shorter-date commitments, lenders are signalling confidence in the long-term value of NVIDIA GPUs running on CoreWeave’s cloud platform and a willingness to underwrite renewal risk.」[199]。貸款方以三年期的合約為依據，放出五年期的貸款，押注合約結束時 GPU 仍然會有客戶。

營收很集中。Microsoft 約占 2025 年營收的 67%；2026 年第二季，前三大客戶分別占 36%、26% 和 10% [200], [93]。該公司警告，如果有大客戶違約，它可能會留下閒置的運算容量，還要承擔設備、租約和融資的費用，並補充說：「We do not currently maintain credit insurance to insure against customer credit risk.」[93]。另一端則是 NVIDIA：它供應 CoreWeave 所有的 GPU，在 2026年1月買進該公司 20億美元的股票，並在 2025年9月承諾，到 2032 年為止買下所有找不到客戶的剩餘運算容量，初始合約金額為 63億美元 [92], [93]。「循環」一節已經談過這份合約。

Nebius 是另一家在美國上市的新興雲端業者，呈現出同樣的結構。2026 年上半年，它投資了 81億美元，營收為 9.8億美元；營運現金流之所以為正，全靠客戶預付的 44億美元。1 月，它把伺服器和網路設備的耐用年限從四年延長到五年 [201]。它與 Microsoft 簽訂的 174億美元合約，要等到 Nebius 確認「that it has secured any additional financing required in respect of the capital expenditures under the Agreement」才會生效 [202]。合約取決於融資，而融資又取決於合約。

資金成本上升了。根據 CNBC，2026年9月底，美國十年期公債殖利率接近 5.17%，放款人也表示自己變得更挑剔：「Instead of a roster of 50 neoclouds, there's probably 20 that the market's truly interested in.」[203]。CoreWeave 估計，利率每上升一個百分點，其浮動利率債務的利息支出每季約增加 3,000萬美元 [93]。Mizuho 的分析師在 7 月就已經談到「neoclouds generating negative free cash flow, widening credit spreads and capital raise concerns」[195]。

這三個環節彼此綁在一起。OpenAI 的承諾是 Oracle 和 CoreWeave 待履行合約中的重要部分；Oracle 和 CoreWeave 的債務以這些合約為支撐；NVIDIA 則為這條鏈的部分環節提供擔保。如果 OpenAI 的營收令人失望，壓力會先落到 Oracle 和 CoreWeave，接著是它們的債權人，最後是 NVIDIA 的擔保。這就是 2008 年那一節所描述的錯向風險（wrong-way risk）：保障恰恰在需要動用時失去價值。

總體經濟與官方警示

2000 年的泡沫破裂拖垮了股市，並帶來一場為期八個月的衰退。2008 年的那一次則讓銀行倒閉，並帶來二戰以來最長、長達 18 個月的衰退 [40]。差別在於傳導管道：一次是股票，一次是信貸。本節衡量 AI 在美國經濟中的分量，並呈現中央銀行與國際組織對於修正可能經由哪個管道傳導的說法。

成長有多少來自 AI

2025年10月，Fortune 雜誌刊出了哈佛大學經濟學家 Jason Furman 的一項計算：「Technically, as Furman notes, investment in information-processing equipment and software was only 4% of U.S. GDP for the first half of 2025, yet it also accounted for fully 92% of GDP growth over that period.」Furman 本人也提醒，如果沒有 AI 熱潮，利率和電價大概會比較低，這會為其他產業帶來一些成長 [204]。我用當時的資料重算了一遍，結果吻合：這 92% 是從 GDP 的連鎖美元數值中減去科技投資的連鎖美元數值而得 [205]。問題出在方法。負責計算美國 GDP 的美國經濟分析局（BEA）提醒，對於價格快速下跌的產品，例如電腦，「the chained-dollar levels (as distinct from chain-weighted indexes and percent changes) will overstate their relative importance to GDP growth.」[206]。

依我的計算，根據 BEA 本身的成長貢獻表（2026年9月的修訂版），資訊處理設備與軟體占 2025 年上半年成長的約 56%、全年成長的 30%（2.3 個百分點中的 0.70 個百分點），以及 2026 年上半年成長的約 41% [23]。這些都是毛額數字。很大一部分設備是進口的，而 GDP 會加計投資，但扣除進口。美國的電腦、周邊設備與零件進口，從 2024 年的 2,170億美元增加到 2025 年的 3,620億美元，並在 2026 年第二季達到年化 6,010億美元的水準 [23]。各項官方估計因方法不同而差異很大：

表中各項來源依序為 [204], [207], [166], [208], [23] 和 [31]。以毛額計算，科技投資約占 2025 年上半年美國經濟成長的一半到三分之二。扣除進口後，影響小得多。各項估計之間的差距本身就說明了一些事：AI 在美國經濟中的分量取決於方法，而只提出單一數字的人，其實是在挑選其中一個。

影響不只停留在投資。2026年8月，美國資料中心營建支出達到年化 849.5億美元（初步數據），比一年前高出 73%，同期整體營建支出則下滑 1.7% [209]。而股價上漲也牽動消費：英格蘭銀行的幕僚估計，美股的漲幅貢獻了「around 10% of the increase in US real consumption since the start of 2024, with AI-related equities the main driver during this period.」[166]。

國際清算銀行是中央銀行的銀行，該行的經濟學家把這一輪週期放進歷史脈絡中檢視。根據他們的計算，IT 設備工廠和資料中心的支出在 2025 年年中約相當於美國 GDP 的 1%，「similar in size to the US shale boom of the mid-2010s and half as large as the rise in IT investment during the dot-com boom of the 1990s.」他們還補充：「The end of previous investment booms was associated with a slowdown in GDP growth of more than 1 pp on average (Graph 4.B). Notably, the largest contraction was after the US dot-com boom, even though the boom was small relative to GDP.」[208]。單憑週期的規模，看不出它可能造成多大的傷害。

集中度

第二個管道是股市。2025年10月，英格蘭銀行金融政策委員會指出，S&P 500 前五大公司合計約占指數的 30%，這個比重「higher than at any point in the past 50 years」，也因此「any AI-led price adjustment would have a high level of pass-through into the returns for investors exposed to the aggregate index.」[210]。國際貨幣基金在同一個月計算，七大科技公司占指數的 33%，而一項以 Herfindahl-Hirschman 指數為基礎的集中度指標「substantially higher than during the dot-com bubble」[207]。2026年7月，英格蘭銀行估計，與 AI 相關的公司已約占 S&P 500 的一半，2022 年時則是四分之一，而且這個比重可能隨著 OpenAI 和 Anthropic 的股票發行而上升 [1]。

國際清算銀行套用了一種用來偵測泡沫的統計檢定，它衡量價格是否已進入爆炸性的軌跡。2025年12月，S&P 500 和黃金都已「entered explosive territory」，而且這是「the only time in at least the last 50 years in which gold and equities have entered this territory simultaneously」。國際清算銀行自己也提醒這項檢定的局限：「it provides no information on when bubbles may burst.」[211]。

從股權到債務

最讓中央銀行擔心的不是股價，而是付錢方式的改變。歐洲中央銀行在 2026年5月寫道，AI 企業與基礎設施「have started to rely increasingly on credit financing」，並提醒一個歷史規律：「Strong growth in either equity prices or business debt in isolation has not been associated with financial crises in the past. That said, 15% of those years that showed particularly strong growth in both equity prices and business debt have been followed by a financial crisis within two years」。隨文附上的圖表標題十分直接：「Greater debt financing for AI is concerning for financial stability」[212]。同一份報告也附帶說明，就整體而言，美國和歐元區企業債務的成長仍然偏低。

英格蘭銀行 2026年7月的數字顯示了速度之快。五大超大規模雲端業者在 2025 年底只占美國投資級公司債的 3%，但在 2026 年截至 5 月初的發行量中占了超過 15%。年初以來，與 AI 相關的發行人占美國高收益債發行量（不含再融資）的 41%；而根據經濟合作暨發展組織（OECD），私募信貸在 AI 融資中的占比從 2024 年的 9% 升至 2025 年的 34% [1]。9 月，該行的委員會引用 Morgan Stanley 的估計：2026 年截至當月初，與 AI 相關的債務發行約 4,500億美元，是 2025 年全年的兩倍多 [213]。在聯準會對市場參與者所做的調查中，AI 在風險項目中被提及的比例從 2025 年的 30% 升至 2026 年的 50%，受訪者並指出，這些投資「are increasingly funded by debt, creating leverage in the system」[214], [70]。

下表整理了官方警示的演變：

國際清算銀行 2026年6月的年度報告，做出了本文所提出的歷史連結。1830 年代的運河熱、1840 年代的鐵路熱、1920 年代的電氣化熱，以及 1990 年代的網際網路熱，「all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify. These episodes ended with an eventual reversal in investment, inducing economy-wide recessions.」至於當前的週期：「The scale and pace of the current AI investment boom accompanied by expectations of large productivity payoffs bear resemblance to these precedents, highlighting potential downside risks in the near term.」[122]。

2026年7月：一次預演

市場已經有過一次預演。根據英格蘭銀行，「Equity valuations for AI companies fell sharply in July. The scale of the adjustment was amplified by an unwinding of stretched positions and associated deleveraging activity.」使用槓桿的投資人蒙受了重大損失，但「there was no spillover to core markets.」[213]。國際清算銀行記錄了「a highly leveraged AI-focused hedge fund」的倒閉、超大規模雲端業者的本益比下跌將近 15%、半導體製造商的本益比下跌 40%，以及這些公司的 CDS 利差上升，而 S&P 500 其他公司的利差則保持穩定 [217]。7月27日，Apple 超越 NVIDIA，成為全球市值最高的公司；根據 CNBC，投資人「rewarded its reluctance to spend heavily on capital expenditures for AI, preferring to rent capacity instead of building its own.」[218]。

這次事件同時說明了兩件事。市場已經在向投資更多、舉債更多的公司索取溢酬。而且，只要這一輪週期的大部分仍是用現金和股票支付，股市重挫就還不會演變成信用危機。

監管機構沒有說的事

這些文件沒有一份預測會發生危機。英格蘭銀行在 2025年12月寫道：「The predominantly cash and equity-financed nature of AI development to date limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term.」[166]。國際貨幣基金在 2026年4月認為「financial stability risks remain contained」[31]。歐洲中央銀行則指出，與 2000 年不同，超大規模雲端業者「combine high profit margins, strong earnings growth, little debt and diversified underlying businesses beyond AI」[215]。

它們一致指出的是機制：從以現金和股票融資，轉向以債務融資，無論在表內還是表外。正是這種轉變，可能把一場像 2000 年那樣的股市修正，變成一場像 2008 年那樣的信用問題。英格蘭銀行模擬了一個與 AI 相關的假設性修正情境：美股在六季內下跌 45%，信用利差上升 350 個基點，這將使英國 GDP 減少 2.2 個百分點 [1]。該行在 9 月還提醒，成長預測和公共財政也都取決於 AI：「A reassessment of those expectations could therefore affect not only AI-related asset valuations but also sovereign debt markets.」[213]。

那巴西呢？

巴西並不在本文所描述的循環中心。沒有任何巴西企業銷售 GPU、訓練前沿模型，或為了興建資料中心而發行數百億美元的債券。但巴西透過三個管道與這個週期相連：公共政策，押注於吸引資料中心；能源，這是巴西最主要的訴求，也是投資中最持久的部分；以及投資人，他們對美國股票和全球金融情勢都有曝險。

Redata：遲來的獎勵措施

聯邦層級的主要押注是資料中心服務特別稅制，簡稱 Redata。2025年9月17日的第 1,318 號臨時措施，暫停對獲得資格的資料中心在購買科技設備時課徵 PIS/Cofins、IPI 和進口稅。作為交換，企業必須把享有優惠所建置容量的至少 10% 提供給國內市場，全部電力需求都要由潔淨或再生能源滿足，用水效率（WUE）不得超過每 kWh 0.05 公升，並將產品價值的 2% 投入研究與發展 [219]。其理由說明書指出，「約 60% 的國內數位工作負載」是在國外處理的，而在巴西營運的成本平均高出 30%，「主要是由於對資通訊科技（TIC）設備的課稅」。巴西聯邦稅務局估計，稅收損失在 2026 年為 52.0億雷亞爾，2027 年為 10.0億雷亞爾，2028 年為 10.5億雷亞爾 [220]。然而，最常被引用的數字並不在任何技術文件裡。在簽署儀式上，副總統 Geraldo Alckmin 說：「預期巴西可在 10 年間吸引 2兆雷亞爾的投資」[221]。這個數字沒有公布任何計算方法。

這項臨時措施未經表決，於 2026年2月25日失效 [222]。根據 Agência Brasil 的報導，參議院議長 Davi Alcolumbre 決定不在期限內將其付諸表決 [223]。其內容改以法案形式重新提出，眾議院於 2月24日通過，參議院則直到 9月1日才通過 [224], [225]，並以 2026年9月15日的第 15,504 號法律簽署公布。新法把「潔淨或再生能源」改為「再生或低排放能源」，同時保留了一個關鍵細節：由於稅制改革將過渡到 CBS 和 IBS，PIS/Cofins 和 IPI 的優惠「效力至 2026年12月31日為止」[226]。同一天，一項補充法在財政規則中開了例外，讓這項優惠得以給予 [227]。簽署公布時，Agência Brasil 報導指出，部分規定「仍有待施行細則，以界定企業的資格認定程序、技術標準，以及實際運用租稅優惠所需的其他條件」[228]。截至 2026年10月13日，在巴西聯邦官方公報上搜尋，仍找不到施行細則的法令 [229]。如果法令遲遲不出，2026 年主要租稅優惠的適用期間將會很短。

宣布與官方文件

為巴西資料中心宣布的金額非常龐大，但一讀官方文件，性質就變了：

ByteDance 的案例最有啟發性。出口加工區委員會核准時，這項計畫「預計在 2035 年前投資共 1,080億雷亞爾於高科技設備」，而且「在 2035 年之後的十年（36-46），公司預計再投資 1,350億雷亞爾於技術升級」[232]。標題上的數字幾乎全是設備，以及設備在二十年間的汰換，也就是折舊那一節討論的那種壽命短的資產。第一個具體里程碑則小得多：「預計到 2027 年第三季，至少有 40 MW 全面投入運轉。」[234]。

在南大河州，州政府表示，礦業暨能源部已發布「保證 Eldorado do Sul 資料中心城取得 5 吉瓦（GW）電力併網的決議」[236]。公布的部令根本沒有提到任何容量。部令認定該併網方案「符合整體最低成本的標準」，以「經 ONS 確認的系統可用性」作為接入條件，並規定若在 2033年12月31日前沒有簽訂輸電系統使用合約，部令即撤銷 [237]。Elea 曾說 Rio AI City 有 1.5 GW 的再生能源「將立即交付」[238]，但公司自己描述的具體階段合計約 200 MW [239]。2026年4月，Elea 出售給 I Squared Capital 的新聞稿提到「超過 1 GW 的規劃容量，以及超過 300 MW 位於電力有保障的土地上」[240]。比較一下規模：Equinix 的 SP6 是一座真實存在、已在運轉的資料中心，造價 1.14億美元，而該公司表示它在巴西「過去十年平均每年投資約 10億雷亞爾」[241]。

這些都不代表詐欺。宣布潛在規模在基礎建設領域很常見。問題在於，在狂熱的週期裡，標題上的數字會被當成已經完成的投資到處流傳，最後成為公共政策的依據。電力規劃顯示了這個落差有多大。在 2025年9月的一場簡報中，能源研究公司（EPE）指出，主管部會追蹤的資料中心申請案，到 2029 年合計 10,721 MW，到 2038 年合計 20,986 MW，其中 13,399 MW 位於東南部。依圖表判讀，EPE 自己對 2029 年主幹電網新增資料中心負載的預測約為平均 1,000 MW [242]。兩者單位不同，一個是申請的容量，一個是平均電量，因此這項比較的意義在於數量級。在巴西聯邦官方公報上搜尋，可以找到至少 15 項 2024 和 2025 年承認資料中心接入主幹電網的部令，其中沒有一項載明容量 [243]。為了過濾投機性的申請，ANEEL 開始要求申請接入輸電網者提供財務擔保 [244]，這與能源那一節所述俄亥俄州電價方案的邏輯相同。

電力過剩，電網不足

巴西的訴求是潔淨能源。2025 年，再生能源占電力結構的 86.8% [245]。而且電力還有剩。依我用 ONS 開放資料所做的計算，該調度機構監測的風力與太陽能電廠在 2025 年因削減指令少發了 37.2 TWh，相當於原本可發電量的 20.7%。高峰出現在 2025年10月，達 30.4%；當年被削減的電量有 84% 位於東北部。2026年1月至10月12日，初步資料合計 31.3 TWh，即 22.1% [246], [247]。根據第 15,269/2025 號法律，因「電力供給過剩，即因發電無法分配至負載而引發的事件」所造成的削減，其成本不能再由電力產業法律所規定的附加費支應 [248]。實務上，發電業者不會因這類削減獲得補償。

資料中心被描繪成這些過剩電力的買家，但這筆帳不會自己平衡。2025 年被削減的電量，約為 EPE 預測 2029 年資料中心新增負載的四倍（平均 1,000 MW 相當於每年約 8.8 TWh）。而且過剩電力集中在東北部，資料中心的申請案卻大多位於東南部。巴西的瓶頸和美國一樣，在於電網和地點，而不是發電。

把塞阿拉州和 AI 連在一起的債務

AI 週期與巴西長期債務之間最直接的連結，在於發電。2026年6月，Casa dos Ventos 在美國完成一筆 11億美元的私募融資，分兩批發行：8.25億美元，期限 24 年；2.52億美元，期限 17 年。獲得融資的電廠「屬於與 Ascenty、Omnia 和 TikTok 母公司 ByteDance 等企業簽訂的長期售電合約，也就是 PPA。」債務以美元計價，因為「售電合約同樣以美國貨幣計價」，該公司表示，這「消除了這筆交易的匯率風險」[249]。匯率風險消失了，但交易對手風險還在：17 至 24 年期的債務，靠的是與買家簽訂的合約，而這些買家的業務仰賴五或六年就折舊完畢的設備。這和美國核能合約的錯配如出一轍。根據該公司的說法，交易的另一端是「23 家來自不同領域的機構投資人，例如保險公司、資產管理公司和退休基金」[249]。

投資人與巴西中央銀行

巴西投資人也有曝險。根據 B3，BDR，也就是在巴西證券交易所交易的外國股票存託憑證，其存量從 2021 年的 340億雷亞爾降至 2022 年的 180億雷亞爾，2025 年則達到 500億雷亞爾，投資人約 100萬名。在報告涵蓋的期間，NVIDIA（NVDC34）是交易量第三大的 BDR，每日 1.16億雷亞爾 [250]。2022 年的下滑與美國科技股的下跌同時發生 [46]，但匯率和贖回也可能有影響，而我沒有找到能區分各項因素的研究。以一般大眾為對象的基金，最多可將淨資產的 20% 投資於海外資產，專業投資人基金則最多 40% [251]。封閉式退休基金最多可將 10% 配置於海外，另外還可將最多 10% 投資於在 B3 交易的 BDR 和國際 ETF [252]。這些是上限，不是實際曝險，而我也沒有找到實際配置的彙總資料。

巴西中央銀行有在關注這個議題。它在 2026年5月的《金融穩定報告》中寫道，「高估值以及投資更集中於特定領域，提高了市場對預期落空的敏感度，並帶來價格更劇烈修正的風險」。關於為資料中心提供資金的美國私募信貸，它指出「與集中度和不透明性相關的重大風險」。它也提醒，新興經濟體「仍暴露於匯率壓力和資本外流之下」。但它的結論是：「鑑於外匯曝險有限，對外部資金的依賴也小，國家金融體系對源自國外之金融波動的曝險很低。」[253]。

一旦出現修正，巴西會間接感受到 AI 週期的衝擊：透過資本外流和匯率，透過已宣布但會縮水或停留在紙上的計畫，透過債務仰賴資料中心合約的發電業者，以及透過投資 BDR 的散戶，他們在 2022 年已經看過存量幾乎減半。Redata 估計的財政成本為三年 72.5億雷亞爾，與宣布的金額相比很小。巴西並不背負這個週期的債務。它背負的是一部分敘事，以及一部分能源。

什麼可能推翻這個論點

一個無法被反駁的論點，就不是分析。本節整理反對 AI 週期具有信貸泡沫形態這一看法的最有力論點，連同支持它們的數據，以及這些數據無法證明的事。

需求確實存在，而且正在成長

最有力的論點是營收。Microsoft 表示，Azure 在 2026 財年突破 1,000億美元，成長 41%，而且「Customer demand continues to exceed available capacity」[13]。Google Cloud 的營收在 2026 年第二季成長 82%，達到 248億美元，待履行合約達 5,140億美元 [254], [14]。AWS 同一季成長 37%，年化規模達 1,690億美元，並表示其 AI 業務已超過每年 250億美元 [25]。Meta 的營收成長 28%，Mark Zuckerberg 把這個成果歸功於 AI：「AI is accelerating our core business today」[16]。在使用方面，Gemini 應用程式的每月使用者達到 9.5億 [254]，OpenAI 則提到「our collective 1.2B weekly users」[255]。

這並不能證明營收足以支應投資，也不能證明需求與融資無關。2026年1月，Microsoft 表示「Approximately 45% of our commercial RPO balance is from OpenAI」[84]。在 Amazon，隨著與 OpenAI 和 Anthropic 的合約擴大，尚未認列的長期合約平均期限在一年內從 4.0 年拉長到 6.4 年 [7]。Microsoft 本身也提供了反面的論據：2026年7月，它表示上一季待履行合約的成長全部來自前沿實驗室以外的客戶，而且當年度雲端營收有「nearly 90%」來自這些其他客戶 [13]。這是最值得追蹤的數據。如果 AI 實驗室以外客戶的需求繼續以這種速度成長，本文描述的循環就沒有看起來那麼舉足輕重。

資產負債表很穩健

依我根據各公司提交給美國證券交易委員會（SEC）的資產負債表所做的計算，2026年6月，四大業者合計持有約 5,330億美元的現金與短期投資：Microsoft 768億美元、Alphabet 2,425億美元、Amazon 1,230億美元、Meta 903億美元 [3], [5], [7], [9]。歐洲中央銀行和國際清算銀行（BIS）提醒，與 2000 年不同，這次股價上漲伴隨著獲利：依 BIS 的說法，這「dissimilar to the dotcom bubble of the late 1990s, which was largely fuelled by over-optimistic expectations that were not underpinned by realised earnings growth.」[211], [215]。

這並不能證明現金來自營運。Alphabet 在 2026年6月透過股票淨募得 496億美元，用途「including capital expenditures to scale AI infrastructure」；上半年發行了 562億美元的債務，並認列 980億美元的利得，「primarily the result of net unrealized gains on our equity securities」[254]。財務長這樣總結公司的債務：「a year ago, we were at about $16 billion; and we went to about $100 billion across multiple currencies and geographies.」[14]。Meta 在 2026 年第二季的自由現金流為 7.84億美元，903億美元的現金已經和 837億美元的長期債務相差無幾 [16]。資產負債表依然穩健，但緩衝空間縮小得很快。

相對於整個體系，債務仍然不大

英格蘭銀行寫道，以現金和股票為主的融資方式「limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term」[166]。歐洲中央銀行指出，整體而言，企業債務的成長仍然偏低 [212]；它在 2025年11月估算，科技業占主要公司債指數的 1.8% 至 8.3% [215]。2008 年那一節指出了另一個有利的因素：私募信貸的資金主要來自長期資本，沒有那種曾拖垮 2007 年各類融資工具的極短期資金依賴。

這並不能證明風險已經被衡量清楚。債務統計並未計入尚未開始的租約，這類租約在最大型的企業中超過 1兆美元；也沒有計入「循環」與「表外」兩節所描述的擔保和採購承諾。而且速度和存量一樣重要：2026 年與 AI 相關的發行量，到 9月時已超過 2025 年的兩倍 [213]。

生產力終將顯現

任務層級的研究顯示了實際的效益。在一項針對 5,172 名客服人員、發表於 Quarterly Journal of Economics 的研究中，能使用 AI 助理「increases worker productivity, as measured by issues resolved per hour, by 15% on average」[256]。聖路易聯邦準備銀行的研究人員估計，生成式 AI 的使用者平均節省了 5.4% 的工時，相當於總體生產力提高約 1.1% [257]。企業的採用也在增加：在美國人口普查局於 2026年10月4日結束的調查中，23.8% 的企業表示在前兩週內使用過 AI，相較之下，2025年11月為 17.3%，當時已採用相同的問題 [258]。BIS 總結了個體與總體之間的落差：任務層級的研究顯示效益「between 20 and 50% in time savings」，但總體生產力的估計「tend to be more conservative at less than 1% over a long horizon」[122]。

這並不能證明為基礎建設出資的人會賺錢。這正是鐵路和光纖的教訓：技術可以改變經濟，最早的投資人卻仍可能賠錢。Microsoft 執行長 Satya Nadella 似乎也認同這個類比。2026年7月被問到產能過剩的風險時，他提到的正是美國鐵路恐慌的那一年：「All of us are reading this 1873 as the book to be read.」財務長 Amy Hood 給出了務實的答案：資本支出已轉向壽命短的資產，而且「if the demand environment changes, you just slow down what is, in fact, the largest component and the driver of COGS.」[13]。對 Microsoft 來說，這是個好論點。但購買 GPU 一方的彈性，正是為這些 GPU 出資者，以及為它們興建建築者的風險。

Token 的價格可能創造自己的需求

使用 AI 的成本下降得非常快。NVIDIA 表示，Blackwell 推出後，光靠軟體，每個 token 的成本就降到原來的五分之一 [132]；國際能源總署（IEA）則估計，每項 AI 任務的能耗正以「by at least an order of magnitude annually」的幅度下降 [22]。如果需求的彈性夠大，每一次降價都會讓使用量倍增，已建置的容量也會被吸收。這是建設者的賭注。問題在於時間：如果需求來得比債務期限還慢，出資者會在容量填滿之前就倒下，就像電信業當年發生的那樣。

我會視為反證的情況

根據這些論據，以下是會讓我放棄這個論點的訊號：

• AI 產品的營收，而不是一般雲端的營收，連續數季支應資本支出中愈來愈大的比例；
• OpenAI 早於媒體所報導文件中的預期，轉為正現金流；
• 前幾代 GPU 的租金連續三年以上維持穩定，與六年的耐用年限相符；
• Oracle 和新興雲端業者在為 2027 與 2028 年到期的債務再融資時，信用利差持續下降；
• 超大規模雲端業者待履行合約的成長，由 AI 實驗室以外的客戶支撐。

如果這些訊號同時出現，AI 週期就算代價高昂，也不是信貸泡沫。截至 2026年10月，情況好壞參半。Microsoft 的 AI 年化營收每年成長 123% [20]，超過資本支出的成長，後者在 2026 財年約增加 80%；而且最近一季，待履行合約的成長完全來自實驗室以外的客戶。H100 是 2022 年的晶片，它的租金在 2026 年再度上漲，如折舊那一節所示，這是支持耐用性的另一個論據。反觀 Oracle 和新興雲端業者的利差上升，OpenAI 也還在燒錢。

情境與警訊

泡沫不會事先約好破裂的日期。Greenspan 在 1996 年談到非理性繁榮，股市卻又漲了三年；BIS 的統計檢定，正如 BIS 自己所提醒，也無法告訴我們泡沫何時結束。因此本節不預測日期。它根據兩次已經發生的預演，描述修正會是什麼樣子，並列出顯示週期走向的指標。

兩次預演

第一次來自技術。2024年12月底，中國的 DeepSeek 發表 V3 模型的技術報告，算出最終訓練的成本為 557.6萬美元，前提是 H800 GPU 每小時租金 2 美元。報告本身也提醒，這個數字不包括「the costs associated with prior research and ablation experiments on architectures, algorithms, or data」[259]。2025年1月27日，市場作出反應：「Nvidia lost close to $600 billion in market cap on Monday, the biggest drop for any company on a single day in U.S. history.」Broadcom 下跌 17%，Oracle 下跌 14%。Cantor 的分析師提到「great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs.」[260]。復甦來得很快：2025年10月，NVIDIA 成為第一家市值突破 5兆美元的公司 [261]。這起事件顯示，效率上的一次變化，就可能讓已建置的容量及其賣方在一夜之間被重新定價。

第二次來自金融。如總體經濟那一節所示，2026年7月，AI 相關股票重挫，一檔槓桿基金倒閉，超大規模雲端業者的股票也改以較低的倍數交易，但並未蔓延到市場其他部分。兩次預演都是股市事件。沒有一次波及信貸。問題是，什麼會讓它波及信貸。

三種情境

在第一種情境中，需求追上產能。AI 產品營收的成長足以支應相當一部分的投資，GPU 租金即使在短缺結束後仍維持堅挺，企業也能順利為 2027 與 2028 年到期的債務再融資。虧損由部分新興雲端業者和較晚進場的投資人承擔。這是企業自己描述的情境，上一節的雲端數據也為它提供了一些支持。

在第二種情境中，修正發生在股市，就像 2000 年那樣。AI 的變現令人失望，股價下跌，超大規模雲端業者照 Microsoft 財務長所描述的做：放慢採購壽命短的資產 [13]。晶片和記憶體供應商失去訂單，負債最重的新興雲端業者進行重整，經濟成長放緩。英格蘭銀行模擬過這樣的情況，美股下跌 45%，利差擴大 350 個基點 [1]。國際貨幣基金（IMF）估計，AI 投資的退潮可能使全球 GDP、投資和消費減少「by as much as 0.4, 1.3, and 0.6 percent, respectively, by 2028」[97]。BIS 也提醒，過去幾次投資熱潮結束時，GDP 成長平均放緩超過一個百分點 [208]。

在第三種情境中，修正波及信貸。一家大型實驗室無法完成下一輪募資，於是重新協商自己的承諾。Oracle 和新興雲端業者在不利的時機面臨再融資：例如 CoreWeave 在 2026 年剩餘期間有 44億美元債務到期，2027 年還有 62億美元 [93]。租金擔保和殘值保證開始被要求履行。以市價評估資料中心債券的基金調降價格；半流動性私募信貸基金在 2026 年第一季的贖回申請已經超過資金流入，將面臨更多流出 [70]。這就是錯向風險（wrong-way risk）出現的情境：擔保債務的一方，在擔保被要求履行的同時失去價值。

我不為這些情境設定機率。本文的論點更保守，也更可驗證：2025 與 2026 年建立起來的融資結構讓第三種情境成為可能，這在週期還靠現金和股票支付的時候是不可能的。

誰先虧損

依照前面各節描述的資本結構，損失會依序發生。首先是實驗室和新興雲端業者的股東，包括透過開放散戶參與的募資輪進場的個人投資人。接著是新興雲端業者的債權人，他們的貸款以 GPU 為擔保品。再來是提供擔保或入股客戶的供應商，例如 NVIDIA、AMD 和 Google，它們會同時失去營收和投資價值。然後，如果租金和殘值保證不足以支應債務，損失會落到資料中心融資工具的債券持有人身上，其中包括保險公司和固定收益基金。最後是超大規模雲端業者本身，它們仍有義務支付自己簽下的租金。這是可能的順序，而非必然：2007 年，看似遙遠的承諾因為合約或聲譽的緣故，提前回到了資產負債表上。

該追蹤什麼

以下指標都是公開的，可以逐季顯示週期處於哪一種情境：

這些訊號沒有一個能單獨定論。合在一起，它們能顯示這筆帳是由營收支付，還是被往後推延。

結論

本文要問的，從來不是人工智慧是否真實。鐵路是真實的，光纖也是，兩者都改變了經濟，但那是在讓相當一部分出資者破產之後。問題在於，如果營收來得比債務還慢，誰來付這筆帳。

「如何辨認泡沫」一節的六個標誌給出了部分答案。真實衝擊確實存在：四大雲端業者的投資從 2024 年的 2,280億美元增加到 2025 年的 3,760億美元，2026 年可能超過 7,000億美元，雲端與 AI 營收也在成長。以信貸支應擴張同樣存在：2026 年截至 5月，超大規模雲端業者占美國投資級債券發行量的 15% 以上，而到了 9月，與 AI 相關的發行量已經超過 2025 年全年的兩倍。負債端的創新顯而易見：從未合併的合資企業到殘值保證，從以 GPU 為擔保品的貸款到把晶片變成投資資產的平台，BlackRock 執行長把這些平台比作房貸債券市場的起步。仰賴自身融資的營收也已有數字可循：在 IMF 稱為循環式的交易中，承諾金額達 1.4兆美元；根據 BIS，AI 企業獲得的投資有 55.2% 來自其他 AI 企業。最後兩個標誌只部分出現。帳面上，擔保品老化得很快，晶片每年推出一代，折舊卻以六年計算，但由於短缺，舊 GPU 的租金在 2026 年上漲了。價格也很高，CAPE 接近 41，股市集中度處於歷史紀錄水準，但這些價格有真實存在的獲利支撐。

2008 年的教訓幾乎可以逐點套用。表外承諾會在危機中回到資產負債表內，而如今這類承諾包括超過 1兆美元尚未開始的租約，以及 Meta、NVIDIA、AMD 和 Google 提供的擔保。與市價連動的觸發條件會讓損失提前發生，而新興雲端業者的貸款規模是依 GPU 的帳面價值決定的。信用評等衡量的可能是擔保人，而不是資產本身：Beignet 的 A+ 債券資助的園區，唯一的承租人和保證人都是 Meta。會隨風險一起貶值的擔保，在被要求履行時就會失靈，而 IMF 已經用錯向風險（wrong-way risk）一詞來描述 AI 生態系。受市場緊盯的指標也會成為財務工程的對象：更長的耐用年限、一種降低所揭露資本支出卻不改變投資的租賃重分類，以及一種扣除客戶預付款的新資本支出衡量方式。

差異也必須說清楚。與 2007 年的銀行不同，超大規模雲端業者手握現金，槓桿也低；私募信貸不像商業本票融資工具那樣依賴極短期資金；資訊揭露更完善，合併報表的規則也更嚴格。末端也沒有數以百萬計的負債家庭。如果這個週期破裂，最可能的結果介於 2000 年和 2008 年之間：一場股市修正，信用損失集中在實驗室、新興雲端業者、提供擔保的供應商，以及買下它們債務的基金。

令人擔憂的是方向。這個週期一開始靠現金和股票支付，現在正快速轉向債務，其中很大一部分不在資產使用者的資產負債表上。截至 2026年10月，AI 這筆帳散落在財報附註、或有擔保，以及沒有人需要合併的融資工具之中。這正是那種在 2008 年要等到為時已晚、無法再分攤時，才完整浮現的帳。

參考文獻

1. [1]Bank of England, “Financial Stability Report – July 2026,” Bank of England, Jul. 7, 2026. [Online]. Available: https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026 ↩
2. [2]U.S. Securities and Exchange Commission, “Company facts: Microsoft Corporation (CIK 0000789019),” EDGAR XBRL API, Oct. 13, 2026. [Online]. Available: https://data.sec.gov/api/xbrl/companyfacts/CIK0000789019.json ↩
3. [3]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm ↩
4. [4]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Feb. 5, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm ↩
5. [5]Alphabet Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 23, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm ↩
6. [6]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Feb. 6, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm ↩
7. [7]Amazon.com, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 31, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm ↩
8. [8]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission, Jan. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm ↩
9. [9]Meta Platforms, Inc., “Form 10-Q for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission, Jul. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm ↩
10. [10]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2026,” U.S. Securities and Exchange Commission, Jun. 22, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm ↩
11. [11]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2023,” U.S. Securities and Exchange Commission, Jul. 27, 2023. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000095017023035122/msft-20230630.htm ↩
12. [12]Oracle Corporation, “Form 10-K for the fiscal year ended May 31, 2025,” U.S. Securities and Exchange Commission, Jun. 18, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000095017025087926/orcl-20250531.htm ↩
13. [13]Microsoft Corporation, “Microsoft Fiscal Year 2026 Fourth Quarter Earnings Conference Call,” Microsoft Investor Relations, Jul. 29, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 ↩
14. [14]Alphabet Inc., “2026 Q2 Earnings Call (transcrição),” Alphabet Investor Relations, Jul. 22, 2026. [Online]. Available: https://abc.xyz/investor/events/event-details/2026/2026-Q2-Earnings-Call-2026-GgTAq7Is0z/default.aspx ↩
15. [15]A. Palmer, “Amazon hikes 2026 capex to $220 billion due to higher memory costs,” CNBC, Jul. 30, 2026. [Online]. Available: https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html ↩
16. [16]Meta Platforms, Inc., “Meta Reports Second Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Jul. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026050596/meta-06302026xexhibit991.htm ↩
17. [17]H. Son, “Moody's says 'unprecedented' AI spending threatens credit quality of Amazon, Meta, Alphabet and others,” CNBC, Jul. 24, 2026. [Online]. Available: https://www.cnbc.com/2026/07/24/moodys-ai-spending-credit-quality-amazon-meta-alphabet.html ↩
18. [18]Oracle Corporation, “Form 10-Q for the quarterly period ended August 31, 2026,” U.S. Securities and Exchange Commission, Sep. 11, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm ↩
19. [19]Alphabet Inc., “2025 Q4 Earnings Call (transcrição),” Alphabet Investor Relations, Feb. 4, 2026. [Online]. Available: https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx ↩
20. [20]Microsoft Corporation, “Microsoft Fiscal Year 2026 Third Quarter Earnings Conference Call,” Microsoft Investor Relations, Apr. 29, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3 ↩
21. [21]Meta Platforms, Inc., “Meta Reports First Quarter 2026 Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Apr. 29, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028364/meta-03312026xexhibit991.htm ↩
22. [22]International Energy Agency, “Key Questions on Energy and AI,” executive summary, Apr. 16, 2026. [Online]. Available: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ↩
23. [23]U.S. Bureau of Economic Analysis, “National Income and Product Accounts: arquivos de dados anuais e trimestrais (NipaDataA.txt e NipaDataQ.txt),” version of Sep. 30, 2026. [Online]. Available: https://apps.bea.gov/national/Release/TXT/NipaDataA.txt ↩
24. [24]A. Odlyzko, “Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s,” University of Minnesota, preliminary version, Jan. 15, 2010. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/hallucinations.pdf ↩
25. [25]Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Jul. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm ↩
26. [26]C. P. Kindleberger and R. Z. Aliber, “Financial Crisis: A Hardy Perennial,” ch. 1 in Manias, Panics, and Crashes: A History of Financial Crises, 5th ed. Hoboken: John Wiley & Sons, 2005 (publisher's excerpt). [Online]. Available: https://media.wiley.com/product_data/excerpt/46/04714671/0471467146.pdf ↩
27. [27]C. P. Kindleberger and R. Z. Aliber, “Anatomy of a Typical Crisis,” ch. 2 in Manias, Panics and Crashes: A History of Financial Crises, 5th ed. Basingstoke: Palgrave Macmillan, 2005 (publisher's preview). [Online]. Available: https://page-one.springer.com/pdf/preview/10.1057/9780230628045_2 ↩
28. [28]H. P. Minsky, “The Financial Instability Hypothesis,” Working Paper No. 74, The Jerome Levy Economics Institute of Bard College, May 1992. [Online]. Available: https://www.levyinstitute.org/pubs/wp74.pdf ↩
29. [29]C. Perez, “Great Surges of development and alternative forms of globalization,” Working Papers in Technology Governance and Economic Dynamics No. 15, The Other Canon Foundation and Tallinn University of Technology, Jul. 2007. [Online]. Available: https://hum.ttu.ee/wp/paper15.pdf ↩
30. [30]C. Perez, “The double bubble at the turn of the Century: Technological roots and structural implications,” CFAP/CERF Working Paper No. 31, University of Cambridge, May 10, 2009; published in Cambridge Journal of Economics, vol. 33, no. 4, pp. 779–805, 2009. [Online]. Available: https://cp.informaticaprojects.com/wp-content/downloads/publications/theoretical-framework/C.PEREZ_CJE_DoubleBubble_WP%20CERF%20No%2031.pdf ↩
31. [31]International Monetary Fund, “Chapter 1,” in Global Financial Stability Report: Global Financial Markets Confront the War in the Middle East and Amplification Risks, Apr. 14, 2026. [Online]. Available: https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf ↩
32. [32]R. J. Shiller, “Speculative Asset Prices,” Prize Lecture, Dec. 8, 2013. [Online]. Available: https://www.nobelprize.org/uploads/2018/06/shiller-lecture.pdf ↩
33. [33]R. J. Shiller, “U.S. Stock Markets 1871-Present and CAPE Ratio,” spreadsheet ie_data.xls, updated Oct. 5, 2026. [Online]. Available: https://shillerdata.com/ ↩
34. [34]A. Greenspan, “The Challenge of Central Banking in a Democratic Society,” speech at the American Enterprise Institute, Board of Governors of the Federal Reserve System, Dec. 5, 1996. [Online]. Available: https://www.federalreserve.gov/boarddocs/speeches/1996/19961205.htm ↩
35. [35]A. Odlyzko, “This time is different: An example of a giant, wildly speculative, and successful investment mania,” The B.E. Journal of Economic Analysis & Policy, vol. 10, no. 1, art. 60, 2010. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/mania01.pdf ↩
36. [36]A. Odlyzko, “The Railway Mania: Fraud, disappointed expectations, and the modern economy,” Journal of the Railway & Canal Historical Society, no. 215, pp. 2–12, Nov. 2012 (revised version of Aug. 14, 2012). [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/mania06.pdf ↩
37. [37]C. P. Kindleberger, “Bubbles in History,” in The New Palgrave Dictionary of Economics, 2nd ed. Basingstoke: Palgrave Macmillan, 2008 (publisher's preview). [Online]. Available: https://page-one.springer.com/pdf/preview/10.1057/978-1-349-95121-5_3073-1 ↩
38. [38]G. Richardson and T. Sablik, “Banking Panics of the Gilded Age,” Federal Reserve History, Federal Reserve System, Dec. 4, 2015. [Online]. Available: https://www.federalreservehistory.org/essays/banking-panics-of-the-gilded-age ↩
39. [39]K. Giesecke, F. A. Longstaff, S. Schaefer and I. Strebulaev, “Corporate Bond Default Risk: A 150-Year Perspective,” NBER Working Paper no. 15848, Mar. 2010 (© 2010 by Kay Giesecke, Francis A. Longstaff, Stephen Schaefer, and Ilya Strebulaev). [Online]. Available: https://www.nber.org/system/files/working_papers/w15848/w15848.pdf ↩
40. [40]National Bureau of Economic Research, “US Business Cycle Expansions and Contractions,” data updated Mar. 14, 2023. [Online]. Available: https://www.nber.org/research/data/us-business-cycle-expansions-and-contractions ↩
41. [41]S. Mixon, “The Crisis of 1873: Perspectives from Multiple Asset Classes,” The Journal of Economic History, vol. 68, no. 3, pp. 722–757, Sep. 2008. [Online]. Available: https://doi.org/10.1017/S0022050708000624 ↩
42. [42]E. A. Couper, J. P. Hejkal and A. L. Wolman, “Boom and Bust in Telecommunications,” Economic Quarterly, Federal Reserve Bank of Richmond, vol. 89, no. 4, fall 2003. [Online]. Available: https://www.richmondfed.org/-/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf ↩
43. [43]A. M. Odlyzko, “Internet traffic growth: Sources and implications,” in Optical Transmission Systems and Equipment for WDM Networking II, Proc. SPIE, vol. 5247, pp. 1–15, 2003. [Online]. Available: https://www-users.cse.umn.edu/~odlyzko/doc/itcom.internet.growth.pdf ↩
44. [44]J. M. Kraushaar, “Fiber Deployment Update — End of Year 1998,” Industry Analysis Division, Common Carrier Bureau, Federal Communications Commission, Sep. 9, 1999, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20000817010609/http://www.fcc.gov:80/Bureaus/Common_Carrier/Reports/FCC-State_Link/Fiber/fiber98.pdf ↩
45. [45]M. Doms, “The Boom and Bust in Information Technology Investment,” FRBSF Economic Review, Federal Reserve Bank of San Francisco, pp. 19–34, 2004. [Online]. Available: https://www.frbsf.org/wp-content/uploads/sites/4/er19-34bk.pdf ↩
46. [46]NASDAQ OMX Group, “NASDAQ Composite Index [NASDAQCOM],” FRED, Federal Reserve Bank of St. Louis, daily series accessed Oct. 2026. [Online]. Available: https://fred.stlouisfed.org/series/NASDAQCOM ↩
47. [47]Bank for International Settlements, 72nd Annual Report, 1 April 2001–31 March 2002, Basileia, Jul. 8, 2002. [Online]. Available: https://www.bis.org/publications/aer-2002.pdf ↩
48. [48]Lucent Technologies Inc., “Form 10-K405 for the fiscal year ended September 30, 2001,” U.S. Securities and Exchange Commission, Dec. 28, 2001. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1006240/000095011701501896/a31752.txt ↩
49. [49]Nortel Networks Corporation, “Form 10-K for the fiscal year ended December 31, 2001,” U.S. Securities and Exchange Commission, Mar. 11, 2002. [Online]. Available: https://www.sec.gov/Archives/edgar/data/72911/000113031902000168/t06646e10-k.htm ↩
50. [50]Cisco Systems, Inc., “Form 10-K for the fiscal year ended July 28, 2001,” U.S. Securities and Exchange Commission, Sep. 24, 2001. [Online]. Available: https://www.sec.gov/Archives/edgar/data/858877/000109581101505065/f75710e10-k.txt ↩
51. [51]J. Frost, R. Kansal, K. Rishabh, V. Shreeti and L. S. Y. Zhang, “Circular relationships among AI firms,” BIS Bulletin, no. 137, Bank for International Settlements, Oct. 1, 2026. [Online]. Available: https://www.bis.org/publications/bulletin-137-circular-relationships-among-ai-firms.pdf ↩
52. [52]U.S. Securities and Exchange Commission, “SEC Charges Qwest Communications International Inc. with Multi-Faceted Accounting and Financial Reporting Fraud,” Litigation Release no. 18936, Oct. 21, 2004. [Online]. Available: https://www.sec.gov/litigation/litreleases/lr18936.htm ↩
53. [53]U.S. Securities and Exchange Commission, “In the Matter of Global Crossing Ltd., Thomas J. Casey, Dan J. Cohrs, and Joseph P. Perrone,” Exchange Act Release no. 51517, Apr. 11, 2005. [Online]. Available: https://www.sec.gov/files/litigation/admin/34-51517.pdf ↩
54. [54]U.S. Securities and Exchange Commission, “SEC Charges WorldCom with $3.8 Billion Fraud,” Litigation Release no. 17588, Jun. 27, 2002. [Online]. Available: https://www.sec.gov/litigation/litreleases/lr17588.htm ↩
55. [55]WorldCom, Inc., “Form 10-K for the fiscal year ended December 31, 2002,” U.S. Securities and Exchange Commission, Mar. 12, 2004. [Online]. Available: https://www.sec.gov/Archives/edgar/data/723527/000119312504039709/d10k.htm ↩
56. [56]W. C. Powers Jr., R. S. Troubh and H. S. Winokur Jr., “Report of Investigation by the Special Investigative Committee of the Board of Directors of Enron Corp.,” Feb. 1, 2002, Exhibit 99.2 to Enron Corp.'s Form 8-K of Feb. 7, 2002. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1024401/000090951802000089/big.txt ↩
57. [57]U.S. Securities and Exchange Commission, Office of the Chief Accountant, Office of Economic Analysis and Division of Corporation Finance, “Report and Recommendations Pursuant to Section 401(c) of the Sarbanes-Oxley Act of 2002 On Arrangements with Off-Balance Sheet Implications, Special Purpose Entities, and Transparency of Filings by Issuers,” Jun. 2005. [Online]. Available: https://www.sec.gov/news/studies/soxoffbalancerpt.pdf ↩
58. [58]Financial Accounting Standards Board, “Summary of Interpretation No. 46: Consolidation of Variable Interest Entities—an interpretation of ARB No. 51,” Jan. 2003, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20040611063520/http://www.fasb.org:80/st/summary/finsum46.shtml ↩
59. [59]Financial Accounting Standards Board, “Summary of Interpretation No. 46 (revised December 2003),” Dec. 2003, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20040611052814/http://www.fasb.org:80/st/summary/finsum46r.shtml ↩
60. [60]Financial Crisis Inquiry Commission, The Financial Crisis Inquiry Report: Final Report of the National Commission on the Causes of the Financial and Economic Crisis in the United States. Washington: U.S. Government Publishing Office, Jan. 2011. [Online]. Available: https://www.govinfo.gov/content/pkg/GPO-FCIC/pdf/GPO-FCIC.pdf ↩
61. [61]Board of Governors of the Federal Reserve System, “Asset-Backed Commercial Paper Outstanding [ABCOMP],” FRED, Federal Reserve Bank of St. Louis, weekly series accessed Oct. 13, 2026. [Online]. Available: https://fred.stlouisfed.org/series/ABCOMP ↩
62. [62]D. M. Covitz, N. Liang and G. A. Suarez, “The Evolution of a Financial Crisis: Panic in the Asset-Backed Commercial Paper Market,” Finance and Economics Discussion Series 2009-36, Federal Reserve Board, Aug. 18, 2009. [Online]. Available: https://www.federalreserve.gov/pubs/feds/2009/200936/200936pap.pdf ↩
63. [63]Citigroup Inc., “Citi Commits Support Facility for Citi-Advised SIVs,” press release (Exhibit 99.1 to Form 8-K), Dec. 13, 2007. [Online]. Available: https://www.sec.gov/Archives/edgar/data/831001/000114420407067635/v097312_ex99-1.htm ↩
64. [64]A. R. Valukas, “Report of Anton R. Valukas, Examiner,” vol. 1, In re Lehman Brothers Holdings Inc., Chapter 11 Case No. 08-13555, U.S. Bankruptcy Court, S.D.N.Y., Mar. 11, 2010, Internet Archive copy. [Online]. Available: https://web.archive.org/web/20100331000019/http://lehmanreport.jenner.com:80/VOLUME%201.pdf ↩
65. [65]Basel Committee on Banking Supervision, “CRE50 – Counterparty credit risk definitions and terminology,” Basel Framework, version published Jul. 5, 2024. [Online]. Available: https://www.bis.org/basel_framework/chapter/CRE/50.htm ↩
66. [66]Basel Committee on Banking Supervision, “Strengthening the resilience of the banking sector,” documento consultivo, Dec. 2009. [Online]. Available: https://www.bis.org/publ/bcbs164.pdf ↩
67. [67]Financial Accounting Standards Board, “Accounting Standards Update No. 2009-17, Consolidations (Topic 810): Improvements to Financial Reporting by Enterprises Involved with Variable Interest Entities,” Dec. 2009. [Online]. Available: https://storage.fasb.org/ASU2009-17.pdf ↩
68. [68]Board of Governors of the Federal Reserve System, “H.8 Assets and Liabilities of Commercial Banks in the United States,” Apr. 9, 2010. [Online]. Available: https://www.federalreserve.gov/releases/h8/20100409/ ↩
69. [69]International Monetary Fund, “Chapter 2: The Rise and Risks of Private Credit,” in Global Financial Stability Report: The Last Mile: Financial Vulnerabilities and Risks, Apr. 16, 2024. [Online]. Available: https://www.imf.org/-/media/Files/Publications/GFSR/2024/April/English/ch2.ashx ↩
70. [70]Board of Governors of the Federal Reserve System, Financial Stability Report, May 2026. [Online]. Available: https://www.federalreserve.gov/publications/files/financial-stability-report-20260508.pdf ↩
71. [71]P. Abbassi, I. Aldasoro and S. Doerr, “Financing the digital economy: the role of private credit,” BIS Quarterly Review, Bank for International Settlements, Sep. 14, 2026. [Online]. Available: https://www.bis.org/publications/financing-digital-economy-role-private-credit.pdf ↩
72. [72]Financial Stability Board, “Report on Vulnerabilities in Private Credit,” May 6, 2026. [Online]. Available: https://www.fsb.org/uploads/P060526.pdf ↩
73. [73]NVIDIA, “OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems,” NVIDIA Newsroom, Sep. 22, 2025. [Online]. Available: https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems ↩
74. [74]NVIDIA Corporation, “Form 10-Q — quarterly report for the period ended October 26, 2025,” U.S. Securities and Exchange Commission (EDGAR), Nov. 19, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581025000230/nvda-20251026.htm ↩
75. [75]OpenAI, “Scaling AI for everyone,” OpenAI, Feb. 27, 2026. [Online]. Available: https://openai.com/index/scaling-ai-for-everyone/ ↩
76. [76]A. Capoot, “Nvidia CEO Huang says $30 billion OpenAI investment 'might be the last',” CNBC, Mar. 4, 2026. [Online]. Available: https://www.cnbc.com/2026/03/04/nvidia-huang-openai-investment.html ↩
77. [77]NVIDIA Corporation, “Form 10-Q for the quarterly period ended July 26, 2026,” U.S. Securities and Exchange Commission, Aug. 26, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm ↩
78. [78]NVIDIA Corporation, “Form 10-K — annual report for the fiscal year ended January 25, 2026,” U.S. Securities and Exchange Commission (EDGAR), Feb. 25, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm ↩
79. [79]Advanced Micro Devices, Inc., “Form 8-K” and Exhibit 99.1, “AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs,” U.S. Securities and Exchange Commission, Oct. 6, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000119312525230895/d28189d8k.htm ↩
80. [80]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended September 27, 2025,” U.S. Securities and Exchange Commission, Nov. 5, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248825000166/amd-20250927.htm ↩
81. [81]Advanced Micro Devices, Inc., “Form 8-K” (warrant issued to Meta Platforms), U.S. Securities and Exchange Commission, Feb. 24, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248826000045/amd-20260223.htm ↩
82. [82]Advanced Micro Devices, Inc., “Form 10-Q for the quarterly period ended June 27, 2026,” U.S. Securities and Exchange Commission, Aug. 5, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/2488/000000248826000123/amd-20260627.htm ↩
83. [83]Microsoft, “The next chapter of the Microsoft–OpenAI partnership,” Official Microsoft Blog, Oct. 28, 2025. [Online]. Available: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/ ↩
84. [84]Microsoft, “Microsoft Fiscal Year 2026 Second Quarter Earnings Conference Call,” Microsoft Investor Relations, Jan. 28, 2026. [Online]. Available: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q2 ↩
85. [85]OpenAI, “OpenAI and Amazon announce strategic partnership,” OpenAI, Feb. 27, 2026. [Online]. Available: https://openai.com/index/amazon-partnership/ ↩
86. [86]Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute,” Anthropic, Apr. 20, 2026. [Online]. Available: https://www.anthropic.com/news/anthropic-amazon-compute ↩
87. [87]Anthropic, “Anthropic raises $65B in Series H funding at $965B post-money valuation,” Anthropic, May 28, 2026. [Online]. Available: https://www.anthropic.com/news/series-h ↩
88. [88]Anthropic, “Microsoft, NVIDIA, and Anthropic announce strategic partnerships,” Anthropic, Nov. 18, 2025. [Online]. Available: https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships ↩
89. [89]Anthropic, “Expanding our use of Google Cloud TPUs and Services,” Anthropic, Oct. 23, 2025. [Online]. Available: https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services ↩
90. [90]CoreWeave, Inc., “CoreWeave Announces Agreement with OpenAI to Deliver AI Infrastructure,” CoreWeave Investor Relations, Mar. 10, 2025. [Online]. Available: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Announces-Agreement-with-OpenAI-to-Deliver-AI-Infrastructure/default.aspx ↩
91. [91]CoreWeave, Inc., “CoreWeave Expands Agreement with OpenAI by up to $6.5B,” CoreWeave Investor Relations, Sep. 25, 2025. [Online]. Available: https://investors.coreweave.com/news/news-details/2025/CoreWeave-Expands-Agreement-with-OpenAI-by-up-to-6-5B/default.aspx ↩
92. [92]CoreWeave, Inc., “Current Report (Form 8-K) — Order Form with NVIDIA Corporation,” U.S. Securities and Exchange Commission (EDGAR), Sep. 15, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm ↩
93. [93]CoreWeave, Inc., “Quarterly Report (Form 10-Q) for the quarterly period ended June 30, 2026,” U.S. Securities and Exchange Commission (EDGAR), Aug. 12, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm ↩
94. [94]OpenAI, “OpenAI and SoftBank Group partner with SB Energy,” OpenAI, Jan. 9, 2026. [Online]. Available: https://openai.com/index/stargate-sb-energy-partnership/ ↩
95. [95]J. Bort, “Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments,” TechCrunch, Nov. 6, 2025. [Online]. Available: https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/ ↩
96. [96]A. Capoot and K. Rooney, “Nvidia, Oracle, CoreWeave and other AI stocks sink on OpenAI revenue report,” CNBC, Oct. 8, 2026. [Online]. Available: https://www.cnbc.com/2026/10/08/open-ai-revenue-nvidia-oracle-coreweave.html ↩
97. [97]International Monetary Fund, “Global Financial Stability Report, October 2026: Resilience, Risk, and Transformation,” IMF, Oct. 12, 2026. [Online]. Available: https://doi.org/10.5089/9798229055673.082 ↩
98. [98]A. Bao, “IMF warns bond yields could see a 'disorderly' spike as debt, inflation risks worsen,” CNBC, Oct. 13, 2026. [Online]. Available: https://www.cnbc.com/2026/10/13/treasury-yields-imf-debt-inflation-.html ↩
99. [99]OpenAI and Broadcom, “OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators,” OpenAI, Oct. 13, 2025. [Online]. Available: https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/ ↩
100. [100]Meta Platforms, “Meta Announces Joint Venture With Funds Managed by Blue Owl Capital to Develop Hyperion Data Center,” Meta Newsroom, Oct. 21, 2025. [Online]. Available: https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/ ↩
101. [101]C. Arroyo and L. Benitez, “Meta, Blue Owl Seal $30 Billion Private Capital Deal for AI,” Bloomberg, Oct. 16, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-16/blue-owl-seals-largest-private-capital-deal-for-meta-s-ai-growth ↩
102. [102]Blue Owl Real Estate Net Lease Trust, “Form 10-Q: quarterly report for the period ended September 30, 2025,” U.S. Securities and Exchange Commission (EDGAR), Nov. 7, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1944366/000194436625000126/osnl-20250930.htm ↩
103. [103]Meta Platforms, Inc., “Form 10-Q: quarterly report for the period ended March 31, 2026,” U.S. Securities and Exchange Commission (EDGAR), Apr. 30, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm ↩
104. [104]Deloitte, “Roadmap: Consolidation — 7.1 Introduction (Determining the Primary Beneficiary),” DART – Deloitte Accounting Research Tool, Sep. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc810-10/roadmap-consolidation/chapter-7-determining-primary-beneficiary/7-1-introduction ↩
105. [105]J. Weil, “AI Meets Aggressive Accounting at Meta’s Gigantic New Data Center,” The Wall Street Journal, Nov. 24, 2025. [Online]. Available: https://www.wsj.com/tech/meta-ai-data-center-finances-d3a6b464 ↩
106. [106]Deloitte, “Roadmap: Leases — 6.7 Amounts That It Is Probable That the Lessee Will Owe Under a Residual Value Guarantee,” DART – Deloitte Accounting Research Tool, Aug. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-6-lease-payments/6-7-amounts-that-it-is ↩
107. [107]Deloitte, “Roadmap: Leases — 8.4 Recognition and Measurement,” DART – Deloitte Accounting Research Tool, Aug. 2026. [Online]. Available: https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc842-10/roadmap-leasing/chapter-8-lessee-accounting/8-4-recognition-measurement ↩
108. [108]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q2_nport.zip,” SEC Division of Economic and Risk Analysis, Jul. 8, 2026. [Online]. Available: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q2_nport.zip ↩
109. [109]U.S. Securities and Exchange Commission, “Form N-PORT Data Sets: 2026q1_nport.zip,” SEC Division of Economic and Risk Analysis, Apr. 3, 2026. [Online]. Available: https://www.sec.gov/files/dera/data/form-n-port-data-sets/2026q1_nport.zip ↩
110. [110]C. Arroyo and E. Ludlow, “xAI to Raise $20 Billion After Nvidia and Others Boost Round,” Bloomberg, Oct. 7, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-07/musk-s-xai-nears-20-billion-capital-raise-tied-to-nvidia-chips ↩
111. [111]C. Arroyo, “How AI Companies Are Keeping Debt Off Their Balance Sheets,” Bloomberg, Oct. 31, 2025. [Online]. Available: https://www.bloomberg.com/news/articles/2025-10-31/meta-xai-starting-trend-for-billions-in-off-balance-sheet-debt ↩
112. [112]Apollo Global Management, “Apollo Backs $5.4 Billion Valor and xAI Data Center Compute Infrastructure Transaction with $3.5 Billion Capital Solution,” Apollo (GlobeNewswire), Jan. 7, 2026. [Online]. Available: https://www.apollo.com/insights-news/pressreleases/2026/01/apollo-backs-5-4-billion-valor-and-xai-data-center-compute-infrastructure-transaction-with-3-5-billion-capital-solution-3214463 ↩
113. [113]NVIDIA Corporation, “Form 8-K (Items 1.01, 2.03 and 7.01),” U.S. Securities and Exchange Commission, Aug. 17, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm ↩
114. [114]A. Capoot and K. Rooney, “Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans,” CNBC, Jul. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html ↩
115. [115]NVIDIA, “NVIDIA Guarantees SB Energy’s PORTS-Pike Technology Campus in Ohio to Exclusively Host NVIDIA AI Compute,” NVIDIA Newsroom, Aug. 17, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-guarantees-sb-energy-s-ports-pike-technology-campus-in-ohio-to-exclusively-host-nvidia-ai-compute ↩
116. [116]TeraWulf Inc., “TeraWulf Announces Fluidstack Expansion with 160 MW CB-5 Lease at Lake Mariner,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Aug. 18, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1083301/000110465925079463/tm2523651d3_ex99-1.htm ↩
117. [117]Cipher Mining Inc., “Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Sep. 25, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1819989/000095010325012168/dp234624_ex9901.htm ↩
118. [118]Hut 8 Corp., “Hut 8 Signs 15-Year, 245 MW AI Data Center Lease at River Bend Campus with Total Contract Value of $7.0 Billion,” Form 8-K, Exhibit 99.1 (SEC EDGAR), Dec. 17, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1964789/000110465925122052/hut-20251217xex99d1.htm ↩
119. [119]NVIDIA, “NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital,” NVIDIA Newsroom, Aug. 10, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital ↩
120. [120]K. Leswing, “Wall Street just endorsed Jensen Huang's ‘big concept’ for AI. What now?,” CNBC, Aug. 11, 2026. [Online]. Available: https://www.cnbc.com/2026/08/11/wall-street-endorsed-jensen-huangs-big-concept-for-ai-what-now.html ↩
121. [121]Bank for International Settlements, “Markets recalibrate amid shifting currents,” BIS Quarterly Review, Mar. 16, 2026. [Online]. Available: https://www.bis.org/publ/qtrpdf/r_qt2603a.htm ↩
122. [122]Bank for International Settlements, “I. Progress and peril,” BIS Annual Economic Report 2026, Jun. 28, 2026. [Online]. Available: https://www.bis.org/publications/aer-2026/progress-peril ↩
123. [123]D. Gautam, J. Kramer, H. Kraus and A. Yokoyama, “Background Analysis 1.1. Rising Leverage in the AI Ecosystem,” Global Financial Stability Report, October 2026 (IMF), Oct. 12, 2026. [Online]. Available: https://www.elibrary.imf.org/display/book/9798229055673/9798229055673.xml ↩
124. [124]CoreWeave, Inc., “Form S-1 Registration Statement,” U.S. Securities and Exchange Commission, Mar. 3, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm ↩
125. [125]Amazon.com, Inc., “Form 10-K for the fiscal year ended December 31, 2024,” U.S. Securities and Exchange Commission, Feb. 7, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm ↩
126. [126]B. Caulfield, “‘Accelerate Everything,’ NVIDIA CEO Says Ahead of COMPUTEX,” NVIDIA Blog, Jun. 2, 2024. [Online]. Available: https://blogs.nvidia.com/blog/computex-2024-jensen-huang/ ↩
127. [127]NVIDIA, “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing,” press release, Mar. 18, 2024. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing ↩
128. [128]NVIDIA, “NVIDIA GB200 NVL72,” product page, version of Aug. 27, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/gb200-nvl72/ ↩
129. [129]NVIDIA, “NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer,” NVIDIA Technical Blog, Mar. 16, 2026. [Online]. Available: https://developer.nvidia.com/blog/nvidia-vera-rubin-pod-seven-chips-five-rack-scale-systems-one-ai-supercomputer/ ↩
130. [130]S. Koparkar, “Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents,” NVIDIA Blog, Aug. 24, 2026. [Online]. Available: https://blogs.nvidia.com/blog/vera-rubin-nvl72-efficiency-ai-agents/ ↩
131. [131]NVIDIA, “NVIDIA H100 GPU,” product page, version of Jul. 1, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/h100/ ↩
132. [132]NVIDIA, “NVIDIA DGX B200,” product page, version of Apr. 29, 2026, accessed Oct. 13, 2026. [Online]. Available: https://www.nvidia.com/en-us/data-center/dgx-b200/ ↩
133. [133]Llama Team, AI @ Meta, “The Llama 3 Herd of Models,” arXiv:2407.21783, Jul. 23, 2024. [Online]. Available: https://arxiv.org/abs/2407.21783 ↩
134. [134]M. Blake, M. Hsu, H. Petty and J. Huntington, “NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories,” NVIDIA Technical Blog, May 20, 2025. [Online]. Available: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories/ ↩
135. [135]Alphabet Inc., “Form 10-K for the fiscal year ended December 31, 2023,” U.S. Securities and Exchange Commission, Jan. 31, 2024. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm ↩
136. [136]Microsoft Corporation, “Form 10-K for the fiscal year ended June 30, 2022,” U.S. Securities and Exchange Commission, Jul. 28, 2022. [Online]. Available: https://www.sec.gov/Archives/edgar/data/789019/000156459022026876/msft-10k_20220630.htm ↩
137. [137]Meta Platforms, Inc., “Form 10-K for the fiscal year ended December 31, 2022,” U.S. Securities and Exchange Commission, Feb. 2, 2023. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1326801/000132680123000013/meta-20221231.htm ↩
138. [138]M. Burry (@michaeljburry), “Understating depreciation by extending useful life of assets artificially boosts earnings…,” X, Nov. 10, 2025. [Online]. Available: https://x.com/michaeljburry/status/1987918650104283372 ↩
139. [139]J. Sor, “Michael Burry's latest argument against hyperscalers like Meta and Oracle echoes one made by an investor who shorted Enron,” Business Insider, Nov. 11, 2025. [Online]. Available: https://www.businessinsider.com/michael-burry-ai-stocks-nvidia-meta-oracle-hyperscalers-jim-chanos-2025-11 ↩
140. [140]Y. Li, “Nvidia name-checks Michael Burry in secret memo pushing back on AI bubble allegations,” CNBC, Nov. 25, 2025. [Online]. Available: https://www.cnbc.com/2025/11/25/nvidia-pushes-back-on-charges-that-ai-investment-is-a-bubble.html ↩
141. [141]C. Yun, “Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances,” AWS News Blog, Jun. 5, 2025. [Online]. Available: https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/ ↩
142. [142]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2023-08-23),” AWS Pricing (Wayback capture), Aug. 25, 2023. [Online]. Available: https://web.archive.org/web/20230825130912/https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
143. [143]Amazon Web Services, “EC2 On-Demand price map — US East (N. Virginia), Linux (published 2026-10-12),” AWS Pricing, Oct. 12, 2026. [Online]. Available: https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20%28N.%20Virginia%29/Linux/index.json ↩
144. [144]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS (Wayback capture), Jun. 9, 2025. [Online]. Available: https://web.archive.org/web/20250609034829/https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
145. [145]Amazon Web Services, “Amazon EC2 Capacity Blocks for ML Pricing,” AWS, Oct. 13, 2026. [Online]. Available: https://aws.amazon.com/ec2/capacityblocks/pricing/ ↩
146. [146]Lambda, “GPU cloud pricing,” Lambda (Wayback capture), Nov. 5, 2025. [Online]. Available: https://web.archive.org/web/20251105190524/https://lambda.ai/pricing ↩
147. [147]Lambda, “GPU cloud pricing: rent NVIDIA H100, H200, and B200,” Lambda (Wayback capture), Oct. 1, 2026. [Online]. Available: https://web.archive.org/web/20261001114850/https://lambda.ai/pricing ↩
148. [148]K. Leswing, “The question everyone in AI is asking: How long before a GPU depreciates?,” CNBC, Nov. 14, 2025. [Online]. Available: https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html ↩
149. [149]E. Lopatto, “Nvidia’s new financial strategy does not compute,” The Verge, Aug. 19, 2026. [Online]. Available: https://www.theverge.com/ai-artificial-intelligence/981668/nvidias-goldman-blackrock-gpu-compute-asset ↩
150. [150]CME Group, “CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks,” CME Group Press Release, Aug. 11, 2026. [Online]. Available: https://www.cmegroup.com/media-room/press-releases/2026/8/11/cme_group_and_silicondatatolaunchcomputefuturesonoctober5tounloc.html ↩
151. [151]S. Koparkar, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” NVIDIA Blog, Oct. 1, 2026. [Online]. Available: https://blogs.nvidia.com/blog/productive-durable-fungible-ai-factories/ ↩
152. [152]International Energy Agency, “Energy and AI,” executive summary, Apr. 10, 2025 (license CC BY 4.0). [Online]. Available: https://www.iea.org/reports/energy-and-ai/executive-summary ↩
153. [153]A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet and D. Sartor, “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, LBNL-2001637, Dec. 2024. [Online]. Available: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf ↩
154. [154]J. Rand, A. Cheyette, C. Talley, S. Zhang, W. Gorman, R. H. Wiser, J. Seel, S. Jeong and F. Kahrl, “Queued Up: 2026 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2025,” Lawrence Berkeley National Laboratory, Jun. 2026. [Online]. Available: https://emp.lbl.gov/queues ↩
155. [155]NVIDIA, “NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint With Broad Industry Support,” press release, Mar. 16, 2026. [Online]. Available: https://nvidianews.nvidia.com/news/nvidia-releases-vera-rubin-dsx-ai-factory-reference-design-and-omniverse-dsx-digital-twin-blueprint-with-broad-industry-support ↩
156. [156]GE Vernova Inc., “GE Vernova reports fourth quarter and full year 2025 financial results; increasing outlook with Prolec GE acquisition,” Exhibit 99 to Form 8-K, U.S. Securities and Exchange Commission, Jan. 28, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000012/gevpressrelease4q25.htm ↩
157. [157]GE Vernova Inc., “GE Vernova reports second quarter 2026 financial results and raises 2026 financial guidance,” Exhibit 99 to Form 8-K, U.S. Securities and Exchange Commission, Jul. 22, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1996810/000199681026000147/gevpressrelease2q26.htm ↩
158. [158]Constellation Energy, “Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to The Grid,” press release, Sep. 20, 2024. [Online]. Available: https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html ↩
159. [159]M. Terrell, “New nuclear clean energy agreement with Kairos Power,” The Keyword (Google), Oct. 14, 2024. [Online]. Available: https://blog.google/company-news/outreach-and-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/ ↩
160. [160]Constellation Energy, “Constellation, Meta Sign 20-Year Deal for Clean, Reliable Nuclear Energy in Illinois,” press release, Jun. 3, 2025. [Online]. Available: https://www.constellationenergy.com/news/2025/constellation-meta-sign-20-year-deal-for-clean-reliable-nuclear-energy-in-illinois.html ↩
161. [161]Talen Energy, “Talen Energy Expands Nuclear Energy Relationship with Amazon,” press release, Jun. 11, 2025. [Online]. Available: https://ir.talenenergy.com/news-releases/news-release-details/talen-energy-expands-nuclear-energy-relationship-amazon ↩
162. [162]Constellation Energy, “U.S. Government Backs Constellation's Plan to Launch Crane Clean Energy Center, Adding 835 MWs of New Baseload Power to the Grid,” press release, Nov. 18, 2025. [Online]. Available: https://www.constellationenergy.com/news/2025/11/us-government-backs-constellations-plan-to-launch-crane-clean-energy-center-adding-835-mws-of-new-baseload-power-to-the-grid.html ↩
163. [163]American Electric Power, “AEP Ohio, PUCO Staff, Ohio Consumers’ Counsel, Ohio Energy Group and Others Agree on How to Address Growing Data Center Power Needs,” Oct. 23, 2024. [Online]. Available: https://www.aep.com/news/stories/view/9829/aep-ohio-puco-staff-ohio-consumers-counsel-ohio-energy-group-and-others-agree-on-how-to-address-growing-data-center-power-needs/ ↩
164. [164]Ohio Power Company, “Data Center Tariff,” web page, accessed Oct. 13, 2026. [Online]. Available: https://www.aepohio.com/company/about/rates/data-center-tariff ↩
165. [165]American Electric Power Company, Inc., “Form 10-Q for the quarterly period ended June 30, 2025,” U.S. Securities and Exchange Commission, Jul. 30, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/4904/000000490425000141/aep-20250630.htm ↩
166. [166]Bank of England, “Financial Stability Report – December 2025,” Bank of England, Dec. 2, 2025. [Online]. Available: https://www.bankofengland.co.uk/financial-stability-report/2025/december-2025 ↩
167. [167]S. Friar, “A business that scales with the value of intelligence,” OpenAI, Jan. 18, 2026. [Online]. Available: https://openai.com/index/a-business-that-scales-with-the-value-of-intelligence/ ↩
168. [168]A. Capoot et al., “OpenAI closes record-breaking $122 billion funding round as anticipation builds for IPO,” CNBC, Mar. 31, 2026. [Online]. Available: https://www.cnbc.com/2026/03/31/openai-funding-round-ipo.html ↩
169. [169]OpenAI, “OpenAI raises $122 billion to accelerate the next phase of AI,” OpenAI, Mar. 31, 2026. [Online]. Available: https://openai.com/index/accelerating-the-next-phase-ai/ ↩
170. [170]S. Muppidi, “OpenAI Says Its Business Will Burn $115 Billion Through 2029,” The Information, Sep. 6, 2025. [Online]. Available: https://www.theinformation.com/articles/openai-says-business-will-burn-115-billion-2029 ↩
171. [171]S. Muppidi and S. Palazzolo, “OpenAI Boosts Revenue Forecasts, Predicts $111 Billion More Cash Burn Through 2030,” The Information, Feb. 20, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-boost-revenue-forecasts-predicts-112-billion-cash-burn-2030 ↩
172. [172]A. Gardizy and A. Efrati, “OpenAI CEO and CFO Diverge on IPO Timing,” The Information, Apr. 5, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-ceo-cfo-diverge-ipo-timing ↩
173. [173]E. Woo, “OpenAI Burned $3.7 Billion in First Three Months of 2026,” The Information, Jun. 16, 2026. [Online]. Available: https://www.theinformation.com/articles/openai-burned-3-7-billion-first-three-months-2026 ↩
174. [174]OpenAI, “Confidential submission of draft S-1 to the SEC,” OpenAI, Jun. 8, 2026. [Online]. Available: https://openai.com/index/openai-submits-confidential-s-1/ ↩
175. [175]SoftBank Group Corp., “Follow-on Investments in OpenAI,” press release, Feb. 27, 2026. [Online]. Available: https://group.softbank/en/news/press/20260227 ↩
176. [176]SoftBank Group Corp., “Execution of Bridge Facility Agreement Primarily for the Follow-on Investments in OpenAI,” press release, Mar. 27, 2026. [Online]. Available: https://group.softbank/en/news/press/20260327 ↩
177. [177]SoftBank Group Corp., “Issuance of Foreign Currency-Denominated Senior Notes,” press release, Sep. 24, 2026. [Online]. Available: https://group.softbank/en/news/press/20260924 ↩
178. [178]CNBC, “Mega-IPOs could signal market top, say analysts as SpaceX and OpenAI prep record floats,” CNBC, May 22, 2026. [Online]. Available: https://www.cnbc.com/2026/05/22/ipo-flurry-top-market-analysts-ai-spacex-musk-altman.html ↩
179. [179]The Information, “How OpenAI’s Web of Business Relationships Could Complicate Its IPO,” The Information, Jun. 22, 2026. [Online]. Available: https://www.theinformation.com/articles/openais-light-balance-sheet-face-hard-look-ipo ↩
180. [180]CNBC, “OpenAI CFO Sarah Friar says company isn't seeking government backstop, clarifying prior comment,” CNBC, Nov. 6, 2025. [Online]. Available: https://www.cnbc.com/2025/11/06/openai-cfo-sarah-friar-says-company-is-not-seeking-government-backstop.html ↩
181. [181]A. Capoot and M. Sigalos, “Sam Altman says OpenAI will top $20 billion in annualized revenue this year, hundreds of billions by 2030,” CNBC, Nov. 6, 2025. [Online]. Available: https://www.cnbc.com/2025/11/06/sam-altman-says-openai-will-top-20-billion-annual-revenue-this-year.html ↩
182. [182]P. Singh et al., “An analyst worries OpenAI will miss its 'extravagant' spending commitments. These stocks would be the most vulnerable,” CNBC, Nov. 7, 2025. [Online]. Available: https://www.cnbc.com/2025/11/07/these-stocks-are-at-risk-if-openai-misses-its-extravagant-spending-plans.html ↩
183. [183]S. Subin, “OpenAI's revenue, growth estimates fall short as company races toward IPO: Report,” CNBC, Apr. 28, 2026. [Online]. Available: https://www.cnbc.com/2026/04/28/openais-revenue-growth-estimates-fall-short-report.html ↩
184. [184]Oracle Corporation, “Oracle Announces Fiscal Year 2026 First Quarter Financial Results,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Sep. 9, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312525199175/orcl-ex99_1.htm ↩
185. [185]U.S. Securities and Exchange Commission, “Company facts: Oracle Corporation (CIK 0001341439),” EDGAR XBRL API, Oct. 13, 2026. [Online]. Available: https://data.sec.gov/api/xbrl/companyfacts/CIK0001341439.json ↩
186. [186]Oracle Corporation, “Oracle Announces Q1 Results Driven by Triple Digit Growth in Cloud Infrastructure Revenues,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Sep. 10, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526387905/orcl-ex99_1.htm ↩
187. [187]Oracle Corporation, “Oracle Announces Record Q4 and FY 2026 Results Driven by Cloud Infrastructure & Cloud Applications,” Exhibit 99.1 to Form 8-K, U.S. Securities and Exchange Commission, Jun. 10, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526265848/orcl-ex99_1.htm ↩
188. [188]S. Mody, “Oracle's credit default swaps are plummeting as financing plan boosts investor confidence,” CNBC, Feb. 2, 2026. [Online]. Available: https://www.cnbc.com/2026/02/02/oracles-credit-default-swaps-are-plummeting-after-financing-plans.html ↩
189. [189]J. Novet, “Oracle beats on earnings, but stock drops on plans to raise another $20 billion,” CNBC, Jun. 10, 2026. [Online]. Available: https://www.cnbc.com/2026/06/10/oracle-orcl-q4-earnings-report-2026.html ↩
190. [190]J. Novet, “Oracle CEO Magouyrk: 'Of course' OpenAI can pay $60 billion per year for cloud infrastructure,” CNBC, Oct. 13, 2025. [Online]. Available: https://www.cnbc.com/2025/10/13/oracle-ceo-magouyrk-of-course-openai-can-pay-60-billion-per-year.html ↩
191. [191]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, Sep. 24, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312525215807/d49252dfwp.htm ↩
192. [192]Oracle Corporation, “Final Pricing Term Sheet” (Form FWP), U.S. Securities and Exchange Commission, Feb. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526033882/d44245dfwp.htm ↩
193. [193]Oracle Corporation, release of Feb. 1, 2026 on the 2026 funding plan, filed as Form FWP, U.S. Securities and Exchange Commission, Feb. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1341439/000119312526032650/d36096dfwp.htm ↩
194. [194]J. Novet, “Oracle stock has worst week since 2001 dot-com bust as AI financing concerns escalate,” CNBC, Jun. 26, 2026. [Online]. Available: https://www.cnbc.com/2026/06/26/oracle-stock-ends-worst-week-since-2001-as-investors-dwell-on-finances.html ↩
195. [195]T. Burns, “What more expensive corporate debt could mean for the AI buildout,” CNBC, Jul. 26, 2026. [Online]. Available: https://www.cnbc.com/2026/07/26/what-more-expensive-corporate-debt-could-mean-for-the-ai-buildout.html ↩
196. [196]Reuters, “Oracle sued by bondholders over losses tied to AI buildout,” published by CNBC, Jan. 14, 2026. [Online]. Available: https://www.cnbc.com/2026/01/14/oracle-sued-by-bondholders-over-losses-tied-to-ai-buildout.html ↩
197. [197]S. Subin, “Oracle sends 'force majeure' notice about data center project — stock drops 3%,” CNBC, Sep. 24, 2026. [Online]. Available: https://www.cnbc.com/2026/09/24/oracle-data-center-force-majeure.html ↩
198. [198]CoreWeave, Inc., “CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing,” CoreWeave Investor Relations, Mar. 31, 2026. [Online]. Available: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx ↩
199. [199]CoreWeave, Inc., “CoreWeave Closes $2.6 Billion Loan Facility, Expanding Financing Flexibility for AI Infrastructure,” CoreWeave Investor Relations, Aug. 10, 2026. [Online]. Available: https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-2-6-Billion-Loan-Facility-Expanding-Financing-Flexibility-for-AI-Infrastructure/default.aspx ↩
200. [200]CoreWeave, Inc., “Annual Report (Form 10-K) for the fiscal year ended December 31, 2025,” U.S. Securities and Exchange Commission (EDGAR), Mar. 2, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm ↩
201. [201]Nebius Group N.V., “Unaudited Condensed Consolidated Financial Statements for the Three and Six Months Ended June 30, 2026,” Form 6-K, Exhibit 99.2 (SEC EDGAR), Aug. 12, 2026. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d2.htm ↩
202. [202]Nebius Group N.V., “Report of Foreign Private Issuer (Form 6-K) — Commercial Agreement with Microsoft,” U.S. Securities and Exchange Commission (EDGAR), Sep. 8, 2025. [Online]. Available: https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm ↩
203. [203]K. Leswing and I. O'Brien, “Debt-hungry AI companies face increased risk as bond yields spike,” CNBC, Sep. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/09/27/debt-hungry-data-center-companies-increased-risk-bond-yields-spike.html ↩
204. [204]N. Lichtenberg, “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says,” Fortune, Oct. 7, 2025. [Online]. Available: https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist/ ↩
205. [205]Federal Reserve Bank of St. Louis, “ALFRED — Real GDP (GDPC1) and Real Private Fixed Investment: Information Processing Equipment and Software (A679RX1Q020SBEA), data vintage of Sep. 25, 2025,” ALFRED, Sep. 25, 2025. [Online]. Available: https://alfred.stlouisfed.org/series?seid=A679RX1Q020SBEA ↩
206. [206]U.S. Bureau of Economic Analysis, “NIPA Handbook: Concepts and Methods of the U.S. National Income and Product Accounts — Chapter 4: Estimating Methods,” BEA, Dec. 2023. [Online]. Available: https://www.bea.gov/resources/methodologies/nipa-handbook/pdf/chapter-04.pdf ↩
207. [207]International Monetary Fund, “Global Financial Stability Report, October 2025: Shifting Ground beneath the Calm,” IMF, Oct. 14, 2025. [Online]. Available: https://www.imf.org/en/Publications/GFSR/Issues/2025/10/14/global-financial-stability-report-october-2025 ↩
208. [208]I. Aldasoro, S. Doerr and D. Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin, no. 120, Jan. 7, 2026. [Online]. Available: https://www.bis.org/publications/bulletin-120-financing-ai-boom-cash-flows-debt ↩
209. [209]U.S. Census Bureau, “Monthly Construction Spending, August 2026 (Release CB26-158),” U.S. Census Bureau, Oct. 1, 2026. [Online]. Available: https://www.census.gov/construction/c30/pdf/release.pdf ↩
210. [210]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – October 2025,” Bank of England, Oct. 8, 2025. [Online]. Available: https://www.bankofengland.co.uk/financial-policy-committee-record/2025/october-2025 ↩
211. [211]Bank for International Settlements, “Volatility challenges risk-taking,” BIS Quarterly Review, Dec. 8, 2025. [Online]. Available: https://www.bis.org/publ/qtrpdf/r_qt2512a.htm ↩
212. [212]European Central Bank, “Financial Stability Review, May 2026,” ECB, May 27, 2026. [Online]. Available: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202605~50566915a7.en.html ↩
213. [213]Bank of England – Financial Policy Committee, “Financial Policy Committee Record – September 2026,” Bank of England, Sep. 30, 2026. [Online]. Available: https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026 ↩
214. [214]Board of Governors of the Federal Reserve System, “Financial Stability Report – November 2025,” Federal Reserve Board, Nov. 7, 2025. [Online]. Available: https://www.federalreserve.gov/publications/files/financial-stability-report-20251107.pdf ↩
215. [215]European Central Bank, “Financial Stability Review, November 2025,” ECB, Nov. 26, 2025. [Online]. Available: https://www.ecb.europa.eu/press/financial-stability-publications/fsr/html/ecb.fsr202511~263b5810d4.en.html ↩
216. [216]A. Bailey (Financial Stability Board), “FSB Chair’s letter to G20 Finance Ministers and Central Bank Governors: August 2026,” Financial Stability Board, Aug. 28, 2026. [Online]. Available: https://www.fsb.org/uploads/P310826.pdf ↩
217. [217]Bank for International Settlements, “Yields climb, yet risk appetite holds firm,” BIS Quarterly Review, Sep. 14, 2026. [Online]. Available: https://www.bis.org/publications/qr-202609/yields-climb-yet-risk-appetite-holds-firm ↩
218. [218]K. Leswing, “Apple ends day as world's most valuable company, passing Nvidia,” CNBC, Jul. 27, 2026. [Online]. Available: https://www.cnbc.com/2026/07/27/apple-most-valuable-company-nvidia.html ↩
219. [219]Presidência da República, “Medida Provisória nº 1.318, de 17 de setembro de 2025,” Diário Oficial da União, Sep. 18, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/mpv/mpv1318.htm ↩
220. [220]G. Alckmin, F. Haddad and A. Silveira, “EMI nº 00010/2025 MDIC MF MME (Explanatory Memorandum of MP 1.318/2025),” Presidência da República, Sep. 16, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/Exm/Exm-1318-25.pdf ↩
221. [221]MDIC, “MP cria o Redata, que estimula datacenters e impulsiona economia digital no Brasil,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, Sep. 17, 2025. [Online]. Available: https://web.archive.org/web/20250929123902/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/setembro/mp-cria-o-redata-que-estimula-datacenters-e-impulsiona-economia-digital-no-brasil ↩
222. [222]Congresso Nacional, “Ato Declaratório do Presidente da Mesa do Congresso Nacional nº 11, de 2026,” Diário Oficial da União, Feb. 27, 2026. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/congresso/adc-11-mpv1.318.htm ↩
223. [223]W. Máximo, “Governo vai buscar diálogo para reativar incentivo para data centers,” Agência Brasil, Feb. 25, 2026. [Online]. Available: https://web.archive.org/web/20260226002349/https://agenciabrasil.ebc.com.br/economia/noticia/2026-02/governo-vai-buscar-dialogo-para-reativar-incentivo-para-data-centers ↩
224. [224]Câmara dos Deputados, “PL 278/2026: legislative record,” Câmara dos Deputados, Sep. 28, 2026. [Online]. Available: https://www.camara.leg.br/proposicoesWeb/fichadetramitacao?idProposicao=2600838 ↩
225. [225]Senado Federal, “PL 278/2026 (Redata),” Senado Federal, Sep. 28, 2026. [Online]. Available: https://www25.senado.leg.br/web/atividade/materias/-/materia/172786 ↩
226. [226]Presidência da República, “Lei nº 15.504, de 15 de setembro de 2026,” Diário Oficial da União (Extra Edition), Sep. 15, 2026. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2026/lei/l15504.htm ↩
227. [227]Presidência da República, “Lei Complementar nº 237, de 15 de setembro de 2026,” Diário Oficial da União (Extra Edition), Sep. 15, 2026. [Online]. Available: https://www.in.gov.br/web/dou/-/lei-complementar-n-237-de-15-de-setembro-de-2026-731891987 ↩
228. [228]P. Peduzzi, “Sancionado regime especial para incentivar instalação de datacenters,” Agência Brasil, Sep. 15, 2026. [Online]. Available: https://agenciabrasil.ebc.com.br/economia/noticia/2026-09/sancionado-regime-especial-para-incentivar-instalacao-de-datacenters ↩
229. [229]Imprensa Nacional, “DOU search: ‘Redata’ (Sep. 15 to Oct. 13, 2026),” Diário Oficial da União, accessed Oct. 13, 2026. [Online]. Available: https://www.in.gov.br/consulta/-/buscar/dou?q=Redata&s=todos&exactDate=personalizado&publishFrom=15-09-2026&publishTo=13-10-2026 ↩
230. [230]Microsoft News Center Brasil, “Microsoft anuncia investimento de R$ 14,7 bilhões ao longo de três anos em infraestrutura de nuvem e IA e oferece treinamento em IA em larga escala para capacitar 5 milhões de pessoas no Brasil,” Microsoft, Sep. 26, 2024. [Online]. Available: https://news.microsoft.com/pt-br/microsoft-anuncia-investimento-de-r-147-bilhoes-ao-longo-de-tres-anos-em-infraestrutura-de-nuvem-e-ia-e-oferece-treinamento-em-ia-em-larga-escala-para-capacitar-5-milhoes-de-pessoas-no-brasil/ ↩
231. [231]About Amazon Brasil, “AWS investirá mais de R$10 bilhões para expandir infraestrutura no Brasil,” Amazon, Sep. 11, 2024. [Online]. Available: https://www.aboutamazon.com.br/noticias/aws/aws-investira-mais-de-r-10-bilhoes-para-expandir-infraestrutura-no-brasil ↩
232. [232]MDIC, “CZPE aprova 12 projetos para R$ 585 bilhões em novos investimentos, incluindo datacenter, metais e amônia verde,” Ministério do Desenvolvimento, Indústria, Comércio e Serviços, Nov. 3, 2025. [Online]. Available: https://web.archive.org/web/20251104021618/https://www.gov.br/mdic/pt-br/assuntos/noticias/2025/outubro/czpe-aprova-12-projetos-para-r-585-bilhoes-em-novos-investimentos-incluindo-datacenter-metais-e-amonia-verde ↩
233. [233]Governo do Estado do Ceará, “Com R$ 200 bilhões de investimento, Data Center do TikTok no Ceará será o maior do Brasil,” Portal do Governo do Ceará, Dec. 3, 2025. [Online]. Available: https://web.archive.org/web/20251203170929/https://www.ceara.gov.br/2025/12/03/com-r-200-bilhoes-de-investimento-data-center-do-tiktok-no-ceara-sera-o-maior-do-brasil/ ↩
234. [234]Governo do Estado do Ceará (Ascom ZPE), “Com novo data center, ZPE Ceará consolida o Estado como hub tecnológico do País,” Portal do Governo do Ceará, Feb. 5, 2026. [Online]. Available: https://web.archive.org/web/20260205195108/https://www.ceara.gov.br/2026/02/05/com-novo-data-center-zpe-ceara-consolida-o-estado-como-hub-tecnologico-do-pais/ ↩
235. [235]Scala Data Centers, “Com investimento inicial de R$ 3 bilhões, governo do RS e Scala Data Centers assinam acordo para o maior projeto de infraestrutura digital do estado,” Scala Data Centers, Sep. 11, 2024. [Online]. Available: https://scaladatacenters.com/com-investimento-inicial-de-r-3-bilhoes-governo-do-rs-e-scala-data-centers-assinam-acordo-para-o-maior-projeto-de-infraestrutura-digital-do-estado/ ↩
236. [236]Governo do Estado do Rio Grande do Sul, “Ministério de Minas e Energia autoriza reserva energética para cidade de data centers no Rio Grande do Sul,” Portal do Estado do RS, May 14, 2025. [Online]. Available: https://estado.rs.gov.br/ministerio-de-minas-e-energia-autoriza-reserva-energetica-para-cidade-de-data-centers-no-rio-grande-do-sul ↩
237. [237]MME/SNTEP, “Portaria SNTEP/MME nº 2.942, de 9 de maio de 2025,” Diário Oficial da União, May 13, 2025. [Online]. Available: https://www.in.gov.br/web/dou/-/portaria-sntep/mme-n-2.942-de-9-de-maio-de-2025-628902215 ↩
238. [238]Elea Data Centers, “Elea assina MoU com Tapestry para assegurar capacidade energética de 1,5 GW da Rio AI City,” Elea Data Centers, Nov. 6, 2025. [Online]. Available: https://eleadatacenters.com/2025/11/06/elea-assina-mou-com-tapestry-para-assegurar-capacidade-energetica-de-15-gw-da-rio-ai-city/ ↩
239. [239]Elea Data Centers, “Elea anuncia a Rio AI City, projeto com capacidade de até 3,2GW em energia verde para impulsionar a Inteligência Artificial no mundo,” Elea Data Centers, May 8, 2025. [Online]. Available: https://eleadatacenters.com/2025/05/08/elea-anuncia-a-rio-ai-city/ ↩
240. [240]Elea Data Centers, “Elea Data Centers é adquirida pela I Squared Capital e inicia nova fase de expansão na América Latina,” Elea Data Centers, Apr. 28, 2026. [Online]. Available: https://eleadatacenters.com/2026/04/28/elea-data-centers-e-adquirida-pela-i-squared-capital-e-inicia-nova-fase-de-expansao-na-america-latina/ ↩
241. [241]Equinix, “Equinix fortalece liderança na América Latina com inauguração de novo data center SP6 em São Paulo,” Equinix Newsroom, Apr. 16, 2026. [Online]. Available: https://newsroom.equinix.com/2026-04-16-Equinix-fortalece-lideranca-na-America-Latina-com-inauguracao-de-novo-data-center-SP6-em-Sao-Paulo ↩
242. [242]EPE (Superintendência de Estudos Econômicos e Energéticos), “A Expansão da Carga de Data Centers no mercado elétrico brasileiro (III COPAM),” Empresa de Pesquisa Energética, Sep. 10, 2025. [Online]. Available: https://web.archive.org/web/20251004104216/https://www.epe.gov.br/sites-pt/areas-de-atuacao/energia-eletrica/WorkshopCOPAM/%5bEPE%5d%20Datacenter_AllexYukizaki%20III%20COPAM_09-2025.pdf ↩
243. [243]Imprensa Nacional, “DOU search: ‘alternativa de acesso à Rede Básica’ (Jan. 1, 2024 to Oct. 13, 2026),” Diário Oficial da União, Oct. 13, 2026. [Online]. Available: https://www.in.gov.br/consulta/-/buscar/dou?q=%22alternativa%20de%20acesso%20%C3%A0%20Rede%20B%C3%A1sica%22&s=todos&exactDate=personalizado&publishFrom=01-01-2024&publishTo=13-10-2026 ↩
244. [244]ANEEL, “Resolução Normativa ANEEL nº 1.122, de 20 de maio de 2025,” Diário Oficial da União, May 26, 2025. [Online]. Available: https://www.in.gov.br/web/dou/-/resolucao-normativa-aneel-n-1.122-de-20-de-maio-de-2025-631621948 ↩
245. [245]EPE, “Balanço Energético Nacional 2026: Relatório Síntese, ano base 2025,” Empresa de Pesquisa Energética, Aug. 3, 2026. [Online]. Available: https://www.epe.gov.br/pt/publicacoes-dados-abertos/publicacoes/balanco-energetico-nacional-2026 ↩
246. [246]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Eólicas (open data),” Operador Nacional do Sistema Elétrico, Oct. 12, 2026. [Online]. Available: https://dados.ons.org.br/dataset/restricao_coff_eolica_usi ↩
247. [247]ONS, “Dados de Restrição de Operação por Constrained-off de Usinas Fotovoltaicas (open data),” Operador Nacional do Sistema Elétrico, Oct. 12, 2026. [Online]. Available: https://dados.ons.org.br/dataset/restricao_coff_fotovoltaica ↩
248. [248]Presidência da República, “Lei nº 15.269, de 24 de novembro de 2025 (conversion of MP 1.304/2025),” Diário Oficial da União, Nov. 24, 2025. [Online]. Available: https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/lei/l15269.htm ↩
249. [249]Casa dos Ventos, “Casa dos Ventos bate recorde com captação de US$ 1,1 bi nos EUA,” Casa dos Ventos, Jun. 18, 2026. [Online]. Available: https://casadosventos.com.br/noticias/captacao-nos-estados-unidos ↩
250. [250]B3, “Relatório anual de BDRs 2025,” B3, Dec. 8, 2025. [Online]. Available: https://www.b3.com.br/data/files/5D/D3/83/92/6430B9105B12E5A9AC094EA8/Relatorio%20Anual%20BDRs%202025.pdf ↩
251. [251]CVM, “Resolução CVM nº 175, de 23 de dezembro de 2022: Normative Annex I (consolidated text),” Comissão de Valores Mobiliários, Mar. 5, 2025. [Online]. Available: https://conteudo.cvm.gov.br/export/sites/cvm/legislacao/resolucoes/anexos/100/resol175consolid_Anexo01.pdf ↩
252. [252]Conselho Monetário Nacional, “Resolução CMN nº 4.994, de 24 de março de 2022 (as amended by CMN Resolution nº 5.202/2025),” Banco Central do Brasil, Mar. 24, 2022. [Online]. Available: https://www.bcb.gov.br/estabilidadefinanceira/exibenormativo?tipo=Resolu%C3%A7%C3%A3o%20CMN&numero=4994 ↩
253. [253]Banco Central do Brasil, “Relatório de Estabilidade Financeira, vol. 25, no. 1,” Banco Central do Brasil, May 22, 2026. [Online]. Available: https://www.bcb.gov.br/content/publicacoes/ref/202605/RELESTAB202605-refPub.pdf ↩
254. [254]Alphabet Inc., “Alphabet Announces Second Quarter 2026 Results,” Alphabet Investor Relations, Jul. 22, 2026. [Online]. Available: https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf ↩
255. [255]OpenAI, “DevDay 2026 Recap,” OpenAI, Sep. 29, 2026. [Online]. Available: https://openai.com/index/devday-2026-recap/ ↩
256. [256]E. Brynjolfsson, D. Li and L. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889–942, Feb. 4, 2025. [Online]. Available: https://doi.org/10.1093/qje/qjae044 ↩
257. [257]A. Bick, A. Blandin and D. Deming, “The Impact of Generative AI on Work Productivity,” Federal Reserve Bank of St. Louis — On the Economy, Feb. 27, 2025. [Online]. Available: https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity ↩
258. [258]U.S. Census Bureau, “Business Trends and Outlook Survey (BTOS): National estimates (AI use, collection ending Oct. 4, 2026),” U.S. Census Bureau, Oct. 8, 2026. [Online]. Available: https://www.census.gov/hfp/btos/downloads/National.xlsx ↩
259. [259]DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437, Dec. 27, 2024. [Online]. Available: https://arxiv.org/abs/2412.19437 ↩
260. [260]S. Subin, “Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history,” CNBC, Jan. 27, 2025. [Online]. Available: https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html ↩
261. [261]S. Meredith, “Nvidia becomes first company to reach $5 trillion valuation, fueled by AI boom,” CNBC, Oct. 29, 2025. [Online]. Available: https://www.cnbc.com/2025/10/29/nvidia-on-track-to-hit-historic-5-trillion-valuation-amid-ai-rally.html ↩