The Fed asked Wall Street what scares it most. 87% said AI
Six months ago half the central bank's market contacts named AI as a threat to the financial system. Now nearly all of them do, and the Fed has counted $680 billion of data center borrowing this year alone.
Twice a year the Federal Reserve rings round its contacts in the markets and asks a simple question: what could break the financial system? In the spring, half of them mentioned artificial intelligence. In the survey published on Friday, 87% did.
Nothing else came close. A general slump in risky assets was second at 52%. Higher long-term interest rates got 48%. Cyberattacks and the sustainability of US government debt, which has passed $40 trillion, tied at 43%. The people the Fed talks to are more frightened of chatbots than of the national debt.
A caveat before the panic. This is a small poll: 23 contacts at banks, funds and advisory firms, interviewed in September and October. It measures mood. But these are the people who lend the money, and the rest of the report suggests their mood has a basis.
Three ways to get hurt
The Fed's Financial Stability Report sorts the AI worries into three piles. In its words, respondents flagged:
"(1) elevated equity valuations and the possibility that lower-than-expected earnings could result in lower valuations and a market correction; (2) cybersecurity risk, stemming from either a malicious attack by sophisticated state actors or damage caused by frontier AI models themselves; and (3) risks associated with complex forms of AI-related financing."
The first is the familiar bubble argument. The second deserves a second read. A document from the US central bank now lists "damage caused by frontier AI models themselves" as a financial stability concern, separate from anything a hacker might do with one. Elsewhere the report says some respondents "pointed to the risk of unintended damage" from the models, and worried that a disruption to trading, settlement or payment systems could shake confidence in financial institutions.
The third pile is where the Fed did its own homework.
The $680 billion tab
A box in the report titled "Leverage in the Data Center Buildout" adds up the borrowing. Debt issuance tied to advanced computing passed $680 billion by October, about five times the $130 billion issued in all of 2023. Three years ago this kind of borrowing was less than 6% of investment-grade corporate debt issuance outside the financial sector. This year it is about 31%.
The Fed puts outstanding public debt linked to advanced computing at more than $1.1 trillion and admits the true figure is hard to measure, because sizable deals also happen in private credit, where nobody has to publish anything. Count the debt, guarantees and leases that have been signed but haven't started yet and the total could top $3 trillion.
The report compares all this to the broadband boom of the late 1990s, the one that ended with fiber in the ground and telecom companies in bankruptcy court. Capital spending by five hyperscalers has been tracking the telecom industry's spending from that cycle, scaled to the size of the economy, and the Fed says the current pace "may surpass that of the 1990s telecom cycle relatively soon."
For most of this decade the big cloud companies paid for their data centers out of cash flow. The report notes, dryly, that their profits reduced the need to borrow "at least until this year."
The most interesting passage is about what the industry calls circular financing. The Fed lays out the mechanics. A hyperscaler signs a long lease for capacity in a data center that a smaller developer hasn't built yet. The developer takes that lease to lenders and borrows against it. Chip makers, meanwhile, guarantee the resale value of their own GPUs so that AI companies can borrow to buy them. Then the hyperscalers and chip designers take equity stakes in the smaller firms on the other side of these deals.
There is good news in this, sort of. If things go wrong, the richest companies in the world are standing behind a lot of the paper, so they "may absorb losses" before the damage reaches anyone else. The bad news is the next sentence: their "economic exposure to such a downturn may be substantially larger than suggested by their on-balance-sheet liabilities."
Put another way, the balance sheets investors are reading do not show the full bill. The bonds these companies have issued sit mostly in mutual funds, pension funds and insurance companies. The Fed's conclusion is written in regulator: the links between data center builders, public credit markets, private lenders and special purpose vehicles "create vulnerabilities that necessitate close monitoring."
The report adds that hedge fund leverage is at its highest since the data series began in 2013, which is the sort of thing that turns a bad week into a bad year.
One more detail. The survey closed in October. Its respondents gave their answers before the midterms, and before Nvidia's founder resigned on November 12 over an affair with a suspected Chinese spy. The 87% is what Wall Street thought when things were calm.
Sources: Federal Reserve Financial Stability Report (PDF), CNBC, The Nation.
More from the feed
United just banned your AI assistant from buying its tickets
Your agent can still look up the fare. As of this week, United says only you get to press buy, and its CEO admits he doesn't want another middleman.
CoreWeave beat its numbers. Its bonds still trade like a company in trouble
Revenue up 168%, a $132.5 billion backlog, and a promise to stop borrowing at the rates lenders are now demanding. The stock and the bonds are reading the same night very differently.
OpenAI built a stop button for its own research. It takes 35 seconds to work
In September an OpenAI training run kept going for hours after a monitor caught an agent slipping past its network rules. Control Room is the fix, and it comes with a 35-second gap.