AI Economics
The Fed Survey and the AI Financing Chain
A survey of 23 market contacts puts AI at the top of perceived financial risks. The financing data tell a more complicated story than a crisis probability.
The Federal Reserve Bank of New York asked 23 market contacts about potential shocks to financial stability in September and October. Twenty mentioned artificial intelligence. In the spring survey, 10 of 20 had done so. The resulting jump from 50% to 87% makes a striking headline, but a survey of contacts is a measure of what those people chose to name, not a 87% probability of a financial crisis.
The Federal Reserve’s November financial stability report gives the concern substance. Respondents pointed to elevated equity valuations, cyber threats and complex financing. Those are different mechanisms. A model security failure, an earnings disappointment and a data-center default do not have the same path into a bank’s balance sheet. Treating all three as one risk can make the headline look more precise than the survey is.
The optimistic case deserves room. Companies with large existing cash flows can finance infrastructure that smaller firms cannot easily build on their own. Long leases give developers some certainty about future revenue and can make expensive facilities financeable. If demand materializes, those commitments support useful capacity, while the stronger companies absorb ordinary project volatility.
A promise of future revenue also creates an exposure. A developer can borrow against a hyperscaler’s lease; the lender’s comfort depends partly on the tenant’s ability and willingness to meet that lease. If the same tenant also invests in the developer, its debt, equity and future capacity costs may all become sensitive to the same demand shock. Multiple contracts can still point back to one economic source of cash.
The numbers have different units
The Fed estimates that advanced-computing-related debt issuance exceeded $680 billion through October 2026, compared with $130 billion issued in all of 2023. The 2026 figure is issuance during a period, not the total amount outstanding. The report separately estimates more than $1.1 trillion in outstanding public debt related to advanced computing. Some private placements are excluded from its issuance measure, so neither number captures every private deal.
A different chart tracks seven hyperscalers and chip companies. At the second quarter of 2026 it shows roughly $880 billion of debt, $1.02 trillion of leases yet to commence, and $2.08 trillion of guarantees and contractual commitments. The latter two are undiscounted amounts, and the Fed cautions that they cannot be directly compared with discounted balance-sheet debt. A future lease payment is also different from a guarantee that pays only under specified conditions. Adding these figures and calling the sum a present-day loss would be an accounting error.
Nor should the $680 billion issuance figure be added to the $880 billion debt figure as if they described separate pools. They cover different periods and samples, with potential overlap. The more useful exercise is to trace a commitment to its counterparty, maturity and trigger, then ask what happens to the same company’s cash needs across a range of outcomes.
Where losses might travel
A less profitable data center does not automatically produce a bank crisis. Equity owners can take the first loss, and a large tenant may keep paying its lease even when capacity earns less than expected. Strong cash flow can give a company time to absorb a bad project. The Fed explicitly notes that large firms taking losses may dampen the effect on the wider system.
There is another side to that protection. If several projects depend on the same tenant or on chip values supported by similar expectations, the tenant or guarantor can face pressure across several contracts at once. The Fed describes GPU residual-value guarantees, where a chip maker supports financing by guaranteeing a value for the hardware. A decline in expected resale value changes the economics of that promise; the size and timing of any payment depend on the actual contract.
The financing can also pass through special-purpose entities with senior and subordinated claims. The report says riskier debt and equity tranches are held by high-yield bond funds, life insurers, business development companies and other private debt vehicles. It estimates that the largest US bank holding companies have close to $1 trillion of exposure across a broad set of technology firms, developers, financing vehicles and utilities. That is an exposure measure across categories, not an estimate of losses or even a measure exclusively attributable to AI.
A single project therefore cannot be assessed from the headline size of its lease alone. Who owns the junior claim, whether the lease can be terminated, and whether a guarantee is capped all affect where a shortfall lands. Public reporting may not disclose enough detail to resolve each point.
Concern is not a forecast
The Fed’s contact survey asked about potential shocks over the next 12 to 18 months. Its fall sample had 23 respondents; the spring sample had 20. The report says the survey summarizes respondents’ views and does not represent the views of the Federal Reserve Board or the New York Fed. Twenty people naming AI is meaningful evidence of what this small group is watching. It is not evidence that 20 of them expect the same failure, or that a crisis is imminent.
For an investor reading the report, the next useful update is a change in the contracts or the cash flows behind them. A refinancing on worse terms, a revised lease commitment or a disclosed guarantee payment can narrow the range of outcomes more than another round of survey anxiety. The credit analysis starts with whose obligation comes due, and who has the cash to pay it.