A share of a small room

The headline number is a share of a small room. New York Fed staff asked 23 market contacts in September and October which shocks over the next 12 to 18 months concerned them. The published 87% corresponds to 20 of the 23 naming artificial intelligence (ALFW's arithmetic from the reported percentages). In the spring round, the published 50% corresponds to 10 of 20 contacts.

What the contacts meant by AI was specific: equity valuations high enough that disappointing earnings could trigger a correction; cybersecurity risk, from state actors using AI or from damage caused by frontier models themselves; and complex forms of AI-related financing. On a correction in risk assets, the respondents' discussion adds the mechanism: stretched valuations, circular financing arrangements, and heavy reliance on debt could amplify losses and spill into business investment and household consumption.

Two limits sit inside the same box. The Fed states plainly that the box is a summary of respondents' views and should not be interpreted as representing the views of the Board or the New York Fed. And on severity rather than frequency, AI did not lead: cyberattacks and fiscal debt sustainability were viewed as having the most severe potential consequences. AI is the most mentioned risk in this survey, which is a different claim from the most dangerous one. The survey also closed before recent weeks: responses were gathered in September and October, and the report reflects information as of October 28, 2026.

The debt figures are the report's own estimates

Box 2.2, "Leverage in the Data Center Buildout," is staff analysis rather than survey sentiment, and its figures are the report's own estimates rather than a poll. Debt issuance to finance advanced computing infrastructure rose to more than $680 billion by October 2026, about five times the $130 billion issued in all of 2023. Its share of nonfinancial investment-grade issuance rose from under 6% in 2023 to about 31% in 2026. Outstanding public debt tied to advanced computing is estimated at more than $1.1 trillion. The mix is shifting as it grows, from mostly corporate bonds toward public and private debt, structured credit, and project finance. The $680 billion is a floor for the trackable market: the Fed's footnote says it covers public deals and Rule 144A private securities, and excludes Section 4(a)(2) private placements it cannot count.

Off the balance sheet, the commitments are larger. Leases not yet commenced carry undiscounted values above $1 trillion, and take-or-pay contracts and residual value guarantees carry undiscounted values above $2 trillion. The Fed warns that these undiscounted figures are not directly comparable with on-balance-sheet debt, so a single combined headline would misread the report. What they show is direction: in the Fed's words, other obligations have grown even more rapidly than debt balances.

Circular, by the Fed's own description

The report describes how strong balance sheets are being used to finance weaker ones. Hyperscalers sign long-term leases with third-party data center developers, and the developers use those leases to secure construction financing. GPU makers guarantee the residual value of their chips so that AI firms with thinner credit can borrow against them. These financings are often supplemented with equity investments by the same hyperscalers and GPU designers, which creates what the Fed calls a "circular financing" dynamic.

Its judgment on that structure is two-sided. The largest technology firms may absorb losses in a downturn, cushioning the rest of the system. The same links mean their real exposure may be substantially larger than their balance sheets suggest. The debt sits mostly with mutual funds, pension funds, and insurers; the riskier subordinated tranches sit with high-yield funds, life insurers, business development companies, and private debt vehicles. The largest U.S. bank holding companies hold close to $1 trillion of exposure across large technology firms, data center developers, advanced computing special purpose entities and private credit vehicles, and utilities, evaluated in the annual stress test.

For scale, the Fed compares the buildout with the late-1990s broadband telecom cycle, plotting five hyperscalers (Amazon, Alphabet, Meta Platforms, Microsoft, and Oracle) against the telecom industry's cumulative spending and borrowing. The report says the current cycle's investment pace may surpass that of the 1990s telecom cycle. The Fed chose to print that comparison.

What to watch

The next survey arrives with the spring report in about six months and will provide another snapshot; respondent composition and question wording need comparing before anyone calls it a trend. The nearer test is issuance: whether advanced computing borrowing holds near a third of nonfinancial investment-grade supply, and whether the private, harder-to-count share keeps growing. A survey of 23 people can reverse in one round. $1.1 trillion of outstanding debt moves more slowly.

Sources

  1. [1] Federal Reserve Board, Financial Stability Report, November 2026 (released November 13, 2026), Box 2.2 "Leverage in the Data Center Buildout" and Box 5.1 "Survey of Salient Risks to Financial Stability"Read source
  2. [2] Reuters, "AI rockets to top of worries about financial system in latest Fed survey" (November 13, 2026)