Several Forces Converged as Markets Plunged
US markets sold off across the board on November 20. By AP's tally, the Nasdaq Composite closed down 5.3% from the previous day, the S&P 500 fell 3.7%, and the Dow Jones Industrial Average lost 1.9%. For the week, the Nasdaq Composite fell 10%, the S&P 500 8.6%, and the Dow 8.0%, and the S&P 500 and the Dow, each more than 10% below its record high, entered correction territory. In Tokyo the same day, the Nikkei Stock Average (Nikkei 225) fell 3,007.56 points (5.06%) from the previous day to 56,436.65, a decline of 22.0% from its closing high of 72,366.34 on June 25.
The selling centered on AI and chips. On the 20th, NVIDIA fell 13.8% by Reuters' count, while by AP's count Micron fell 13.9%, AMD 13.3%, Broadcom 12.2%, and the Philadelphia Semiconductor Index (SOX) 10.7%, all posting steep losses. Anthropic, which had gone public the day before, on the 19th, in the largest listing ever by amount raised, also gave ground on the 20th, falling 8.5% to $88.91 by AP's corrected tally (CNBC also reported an 8.5% drop, while Reuters put it at 9.1%).
Several developments overlapped in the background. The US Federal Reserve (the Fed) had raised rates on September 16, and the October consumer price index, released on November 10, had risen 0.4% from the previous month, keeping markets focused on the prospect of rates staying high.
Then, on November 11, NVIDIA announced that its board had accepted Jensen Huang's resignation as chief executive officer. On the 17th, The Wall Street Journal reported that the US Securities and Exchange Commission (SEC) had opened a civil investigation into the company over its disclosures, though neither the company nor the SEC has confirmed the investigation.
These individual news items did act as triggers to some extent. But what lay beneath the market's anxiety was the enormous investment in AI infrastructure and the swelling debt that supports it.
Commitments Swelling Off the Books, and a Mismatch of Lifespans
Just before the plunge, central banks in the US and Europe had, one after another, sounded the alarm about AI-related financing. According to the Financial Stability Report the Fed published on November 13, advanced computing-related debt issuance rose to more than $680 billion by October 2026, about five times the total for all of 2023. Its share of all nonfinancial investment-grade debt issuance jumped from less than 6% in 2023 to about 31% in 2026.
What both central banks focused on in particular were off-balance-sheet commitments, obligations that never appear on the balance sheet. The Fed tallied the off-balance-sheet commitments of seven large US companies, including Amazon, Microsoft, and NVIDIA, at more than $3 trillion as of the April–June quarter of 2026 (more than $1 trillion in leases that have not yet commenced, and more than $2 trillion in take-or-pay contracts, residual value guarantees, and the like). The European Central Bank (ECB), in its November Financial Stability Review, likewise estimated more than $4 trillion for eight large firms including Anthropic, but the two are separate tallies covering overlapping sets of companies, and they cannot simply be added together.
Corporate filings show that the issue today is not that revenue has failed to materialize, but that future commitments are even larger than the surge in revenue. According to Anthropic's IPO prospectus, its revenue for January–September 2026 was $33.24 billion, up sharply from $2.55 billion a year earlier. Meanwhile, its commitments to pay for compute through 2045, including those entered into after the end of September, total $900.9 billion ($451.6 billion of it by the end of 2030 and $70.7 billion within the next 12 months); its cash on hand, estimated at $232.4 billion on a post-IPO pro forma basis, covers the near-term payments, but the long-term contractual burden is extremely large.
The gap between the term of the borrowed money and the lifespan of AI equipment has also become an issue. The ECB report points out that the maturities of large companies' debt exceed the economic lifetimes of physical capital such as advanced chips, and a note to one of its charts cites market intelligence putting the economic life of AI equipment at one to three years. By contrast, the four big US firms (Microsoft, Alphabet, Amazon, and Meta) extended the accounting useful lives of their servers and network equipment from three years to five or six years between 2020 and 2025, disclosing the amounts of those changes in estimates; as the exception, Amazon alone shortened part of them from six years to five starting in 2025, citing the pace of technology development in artificial intelligence and machine learning.
What Sets This Apart From a Lehman-Scale Crisis
The ECB report wrote of the current trend that "the rapid shift towards debt finance invokes historical parallels with the dotcom and subprime bubbles." An economics study by Jordà and colleagues (2015), which examined data from 17 countries over 140 years, likewise shows that what makes bubbles dangerous is credit (borrowing), and that the collapse of a credit-fueled bubble tends to lead to financial crises, deep recessions, and slow recoveries. In the ECB's view, the bursting of the dot-com bubble, which was financed largely with equity, brought only a shallow recession, whereas a collapse on the debt side has longer-lasting effects.
So is the structure this time the same as the 2008 collapse of Lehman Brothers? At the core of what made the 2008 crisis so severe was a run, in which the rollover of short-term funding such as asset-backed commercial paper (ABCP) and repo transactions suddenly stopped. At the end of November 2007, Lehman Brothers relied on secured short-term funding for 37.3% of its total assets and its leverage had reached 30.7 times, while banks had extended liquidity guarantees to off-balance-sheet vehicles (conduits), so losses flowed back onto the banks themselves.
By contrast, the picture of the financial system that emerges from the primary sources this time differs greatly from 2008. In its November 2026 report, the Fed assesses the banking sector as sound and resilient overall, with capital near historically high levels, and leverage at broker-dealers as low. In addition, AI-related on-balance-sheet debt is held primarily by mutual funds, pension funds, and insurance companies, and bonds issued by data-center special purpose vehicles include long-dated examples such as Meta's joint-venture deal, which matures in 2049.
In the 2008 crisis, the S&P 500 fell 56.8% from its October 2007 peak to its March 2009 trough (Federal Reserve Bank of Atlanta), and CNBC reported at the time that it dropped 18.19% in a single week in October 2008. By comparison, the mechanisms through which short-term funding can dry up within days are still limited this time, and the scale of the stock decline is also different. That said, the Fed also cautions that channels of contagion remain, such as bank credit commitments to other financial institutions exceeding $2.8 trillion in the April–June quarter of 2026, bond and loan mutual funds that allow daily redemptions, and hedge fund leverage at its highest level since the data began.
A Closer Mirror in History: The Telecom Bubble of 2000
A precedent that stands out as closer to today's AI investment than the 2008 financial crisis is the telecom bubble of 2000–2002. The dot-com bubble is often described as a stock-only collapse, but in reality huge debts and extensions of credit built up around telecom infrastructure, and both stocks and debt collapsed. Not only did the Nasdaq Composite fall 78% from its peak (according to a National Bureau of Economic Research paper), but defaults on US and Canadian high-yield bonds hit a record $96.9 billion in 2002, with telecom and media companies accounting for more than half in both 2001 and 2002 (according to a tally by Altman and a coauthor at New York University).
Emblematic of the era was the vendor financing provided by telecom equipment makers such as Lucent. Lucent expanded credit lines and debt guarantees so that customers could buy its equipment, and its commitments as of the end of September 2000 stood at $8.1 billion, about 28% of its restated annual revenue ($28.9 billion). In the following fiscal year, however, customer bankruptcies came one after another, and the company had to book $2.2 billion in provisions for uncollectibles and customer financings.
In today's AI industry, too, a ring of credit extension has been spreading in a different form. According to NVIDIA's filings and earnings materials, the company's maximum exposure under guarantees grew from $108.5 billion as of July 26, 2026, to $118.0 billion as of October 25 (of which residual value guarantees for an OpenAI-related project in Ohio account for up to $105 billion, with payments expected to begin in 2028 or later). The company says it has made no payments under its guarantees to date, that any payment would be limited to the shortfall remaining after recovery through re-leasing or sale, and that it is also indemnified by OpenAI.
Even so, comparing the scale brings the overlap with history into view. NVIDIA's maximum guarantee exposure ($118.0 billion) amounts to about 25% of its total third-quarter revenue ($115.8 billion) on an annualized basis, close to Lucent's level at its peak (about 28% of total revenue).
On top of that, the company also carries investments in other companies ($99 billion as of July, plus $25 billion in investment commitments) and an obligation to buy unused capacity from the neocloud provider CoreWeave (an initial value of $6.3 billion, through April 2032). Most of the maximum guarantee exposure ($105 billion) is the residual value guarantee for a single project in Ohio, so the substance of the commitments differs from Lucent's, which centered on direct lending to customers and debt guarantees; even so, the structure, in which the seller's credit supports the money customers use to buy its chips, has elements in common.
On the other hand, the strongest argument from those who say this is not a bubble is the sheer size of the backlogs that have piled up. As of the end of September 2026, contracted amounts not yet recognized as revenue stood at $683 billion at Microsoft, $567.7 billion at Alphabet, and about $578 billion at Amazon, while Oracle had $638 billion as of the end of May. The ECB also judges that the big cloud companies' earnings have broadly kept up with expectations. Still, a large part of the backlog is tied to a small number of AI companies in which the big firms themselves have invested (Amazon, for instance, has disclosed expanded contracts on the order of $100 billion each with OpenAI and Anthropic), so the backlog and the circular structure are two sides of the same coin.
Japan's Market Links and What to Watch From Here
This US-born turbulence reaches Japan's market, too, through concrete channels of money. At the center is SoftBank Group (SBG), whose shares fell 14.48% on the 20th (as reported by CNBC). According to the company's earnings materials, its cumulative investment in OpenAI had reached $64.6 billion (an ownership interest of about 13%) as of October 1, 2026.
SBG's LTV (adjusted net interest-bearing debt ÷ adjusted equity value of holdings), the metric it uses to gauge its financial soundness, stood at 15.9% at the end of September. The company's stated policy is to keep LTV below 25% in normal circumstances (with an upper threshold of 35% even in extraordinary circumstances), but much of the equity value of holdings in the denominator consists of Arm shares and SoftBank Vision Fund 2 (SVF2), whose holdings include OpenAI. Indeed, in a pro forma estimate as of November 9 reflecting the decline in Arm's shares and other factors, net asset value (NAV) fell by about ¥15 trillion, from ¥59.8 trillion at the end of September to ¥44.7 trillion.
SBG's funding design is also closely tied to the price movements of AI-related stocks. In addition to a margin loan backed by Arm shares (a $25 billion facility, non-recourse to SBG), in August 2026 SBG entered into a $10 billion loan (guaranteed by SBG) that references the fair value of OpenAI preferred shares and carries cash-collateral and prepayment clauses that can be triggered by events such as a significant decline. On OpenAI's valuation as of the end of September, the company states that "the valuation was unchanged from the previous fiscal year-end," but at an investor briefing on November 10, an SBG official said that "Anthropic's listing and subsequent trading will be a factor when we reassess the current valuation, but it will be just one of several factors."
To judge whether this plunge remains a temporary correction or moves on to a shakeout of equipment and debt, one needs to watch the movement of credit more than share prices. For example, the neocloud provider CoreWeave's revenue for July–September rose 168% from a year earlier to $3.66 billion, while its net interest expense ($846 million) was 2.2 times its adjusted operating income ($379 million), and it faces $10.97 billion in principal repayments in 2027. The company expects about $9 billion in customer prepayments and deposits over the next three months and says it will not use the unsecured debt market at current yields; the refinancing of companies like this, movements in banks' credit lines, and the Fed's decision at its December meeting, with the central bank in a rate-hiking phase, will be the milestones ahead.
Because the Tokyo market is closed on November 23 for Labor Thanksgiving Day, a Japanese public holiday, the 24th will be its first trading day after the plunge in US markets on the 20th. History suggests that an infrastructure investment frenzy is truly tested at the moment the question becomes whether the countless commitments piled up off the books will turn into real cash.
References
- How major US stock indexes fared Friday 11/20/2026 (AP)
- S&P 500 sinks 3.7% and the Nasdaq composite goes negative for the year (AP)
- Wall St ends sharply lower on AI selloff, S&P 500, Dow in correction (Reuters, via StreetInsider)
- 日経平均株価 日次データ — Nikkei Stock Average daily data (Nikkei Indexes, Japanese)
- S&P 500 and Dow close in correction; Nasdaq erases 2026 gain as AI rout picks up steam (CNBC)
- Federal Reserve issues FOMC statement, September 16, 2026 (Federal Reserve)
- Consumer Price Index – October 2026 (US Bureau of Labor Statistics)
- Form 8-K, November 11, 2026 (NVIDIA)
- Report: The Wall Street Journal (2026-11-17), SEC Is Probing Nvidia Amid Departure of CEO Jensen Huang
- Financial Stability Report, November 2026 (Federal Reserve)
- The debt financing of AI and its implications for financial stability (ECB Financial Stability Review, November 2026)
- Prospectus (Form 424B4), November 19, 2026 (Anthropic)
- Form 10-K, fiscal year 2021 (Microsoft)
- Form 10-K, fiscal year 2023 (Microsoft)
- Form 10-K, 2021 (Alphabet)
- Form 10-K, 2023 (Alphabet)
- Form 10-K, 2020 (Amazon)
- Form 10-K, 2022 (Amazon)
- Form 10-K, 2024 (Amazon)
- Form 10-K, 2025 (Amazon)
- Form 10-K, 2021 (Meta Platforms)
- Form 10-K, 2022 (Meta Platforms)
- Form 10-K, 2025 (Meta Platforms)
- Leveraged Bubbles — Jordà, Schularick and Taylor (NBER)
- Securitization without risk transfer — Acharya, Schnabl and Suarez (Journal of Financial Economics, author's version)
- Securitized Banking and the Run on Repo — Gorton and Metrick (NBER)
- Form 10-K, fiscal year 2007 (Lehman Brothers Holdings)
- Financing the AI Buildout — Stijn Van Nieuwerburgh (NBER)
- Stock Prices in the Financial Crisis — Gerald P. Dwyer (Federal Reserve Bank of Atlanta, Internet Archive copy)
- Quick Market Stats: Week Ending 10/10 (CNBC, 2008)
- Wealth Effects Revisited 1978-2009 — Case, Quigley and Shiller (NBER)
- Defaults and Returns on High Yield Bonds: The Year 2002 in Review and the Market Outlook — Altman and Bana (NYU Stern)
- Form 10-K, fiscal year 2000 (Lucent Technologies)
- Form 10-K, fiscal year 2001 (Lucent Technologies)
- Form 10-Q, quarter ended July 26, 2026 (NVIDIA)
- Form 8-K, August 17, 2026 (NVIDIA)
- CFO Commentary on Third Quarter Fiscal 2027 Results (NVIDIA)
- NVIDIA Announces Financial Results for Third Quarter Fiscal 2027 (NVIDIA)
- Form 8-K, September 15, 2025 (CoreWeave)
- Form 10-Q, quarter ended September 30, 2026 (Microsoft)
- Form 10-Q, quarter ended September 30, 2026 (Alphabet)
- Form 10-Q, quarter ended September 30, 2026 (Amazon)
- Form 10-K, fiscal year 2026 (Oracle)
- Consolidated Financial Report for the Six-Month Period Ended September 30, 2026 (IFRS) (SoftBank Group, English translation)
- Q2 FY2026 Earnings Results Investor Briefing, Finance (SoftBank Group)
- Q2 FY26 Investor Briefing Q&A (SoftBank Group)
- Form 10-Q, quarter ended September 30, 2026 (CoreWeave)
- CoreWeave Reports Strong Third Quarter 2026 Results (CoreWeave)
- CoreWeave's Q3 shows revenue surge, CEO aims to allay debt worries (Constellation Research)
- Meeting calendars and information (Federal Reserve)
- 営業時間・休業日一覧 — Trading hours and holidays (Japan Exchange Group, Japanese)
This article was written by Google's Gemini 4 Neon and fact-checked by Anthropic's Claude.