Fed's Bowman warns AI debt tops $800 billion this year: what she said, and what she did not

A Federal Reserve official has put a number on how much of the AI buildout is being paid for with borrowed money, and the number is large. Michelle Bowman, the Fed's Vice Chair for Supervision, told an audience at the Council on Foreign Relations on November 10, 2026 that AI-related debt is on track to exceed $800 billion this year. What she said is more specific, and more measured, than some of the headlines that followed.

What Bowman said

The speech, titled "The U.S. Economy and Considerations for Monetary Policy," covers the whole economy. The AI passages are a few paragraphs. She credits "robust increases in AI-related capital expenditures" for supporting growth and says data center investment has "boomed." She then turns to how that investment is paid for. AI-related debt, she said, "is on track to exceed $800 billion this year," with rapid growth in corporate bonds, private credit and project finance.

She added two structural points. First, hyperscalers and large technology firms "have assumed off-balance-sheet liabilities" so that smaller, riskier companies can use collateral in asset-based financing, and other obligations such as lease commitments and residual value guarantees "have also surged." Second, equity investments by hyperscalers and chip designers in smaller companies that buy their products are "creating the appearance of circular financing." We explain how those arrangements work, using company documents, in our circular financing explainer.

The risk she named

Bowman's scenario is conditional. If AI revenues fall short of rising costs, "we could see a sharp correction in equity prices for AI firms, higher financing costs, and reduced access to credit." She argued that such a shock would likely slow the AI buildout and soften economic activity and inflation. That is a channel from markets to the real economy, the kind of exposure a supervisor is paid to watch. It is a warning about financing structure, not a forecast that revenues will disappoint.

What she did not say

Semafor's November 11 report put the speech under the headline that the Anthropic IPO "could lead to 'large correction' in markets, Fed official says." The speech text on the Fed's site does not mention Anthropic or an IPO. The article's own sourcing for the IPO angle is a Nomura analyst, Tore Opsahl, who is quoted saying the offering "might mean a selloff of the AI theme around the IPO." That is a market commentator's view, not the Fed's. The article also describes the offering with figures that do not obviously reconcile, a $1.4 trillion valuation in the lead and $90.4 billion elsewhere, without explaining the gap.

Anthropic's own filings show a company with very large compute commitments, as our coverage of its amended S-1 laid out. But connecting Bowman's general warning to one company is an inference by the press, and readers should label it as such. She spoke about a sector.

The numbers behind the warning

Company disclosures give a sense of the scale she was describing. CoreWeave, one of the most levered pure-play AI infrastructure companies, reported on November 10, 2026 that its total debt was about $46.9 billion at September 30, including roughly $41.6 billion of recourse debt. It also reported a net loss of $709 million for the third quarter on revenue of $3.66 billion, and net interest expense of $846 million. It expects about $9 billion in customer prepayments and deposits over the next three months, and our earlier report on the chip selloff and CoreWeave's debt has more on the sector's borrowing. One company is not the sector, but its interest bill shows why a supervisor cares about the structure.

The market has already had a rehearsal. Between the midterm elections and November 6, AI shares fell sharply and then rebounded, as a weekly roundup of those days noted. The Nasdaq fell 5.52% on November 5, and rose 1.73% the next day, in that roundup's account. The selloff was blamed on permitting risk for data centers, not on credit.

Three channels, three kinds of risk

Bowman named three places the borrowing is happening, and they fail differently. Corporate bonds are public and priced daily, so stress shows up quickly in yields. Private credit is opaque and held by funds that can restrict withdrawals, so stress shows up late and unevenly. Project finance is secured against a specific site and its customer contracts, which is safer if the customer is strong and dangerous if one tenant holds most of the revenue. A supervisor looking at the whole stack cares less about the total than about who holds each layer and how quickly they can sell. That is why the off-balance-sheet liabilities she described matter: they are obligations that do not appear in headline debt figures but still have to be paid.

Why it matters for the race

The AI race is increasingly a financing race. Training runs and data centers are priced in tens of billions, and the labs that win are the ones that can keep raising capital at acceptable rates. A credit squeeze would not stop research, but it would reorder who can afford the next generation of chips. That is the practical meaning of Bowman's warning: the weakest links are the borrowers with the thinnest revenue against fixed obligations, not the largest labs.

Our take

The speech is a useful, sober marker: a senior supervisor is on record counting AI debt and naming circular financing. It is not a prediction of a crash, and it is not about Anthropic. We would watch three things: whether the $800 billion figure appears in the Fed's next financial stability report, whether off-balance-sheet commitments get disclosed more fully, and whether the lenders behind projects like CoreWeave's start demanding shorter terms.