
MarketPedia Inc.™
Finance · 17 November 2026
FinanceExplainer
Fed warns opaque AI debt could threaten financial stability
Governor Michael Barr's speech in Milan, built on the central bank's twice-yearly report, details billions in off-balance-sheet liabilities and financing loops among the biggest tech firms.
Governor Michael Barr told an audience in Milan on Monday that the Federal Reserve is watching three features of artificial intelligence financing with particular care: how fast the debt is growing, how much of it loops back through the same handful of companies, and how many obligations sit off the balance sheet entirely.
He was expanding on the Fed's semi-annual Financial Stability Report, published three days earlier, on November 13. The report marks the first time the central bank has placed AI financing structures among the risks it tracks most closely.
Why the Fed is raising its voice now
Market contacts surveyed by the Federal Reserve Bank of New York in September and October named AI as a risk to financial stability 87% of the time, up from 50% six months earlier. That is a sharp jump for a single survey cycle. It suggests sentiment inside markets is moving faster than the underlying debt figures themselves.
Barr's own words matched that mood. "In the case of a significant repricing of AI investments, stress from these three aspects of AI financing could amplify the repricing of AI investments," he said in Milan.
How big is the debt behind the boom
On-balance-sheet debt at the seven largest AI companies is set to top $800bn by the end of 2026, the Fed estimates. Large as that figure is, it is dwarfed by what sits off the books.
Off-balance-sheet liabilities tied to AI, including long-term contracts, lease commitments and residual value guarantees on computing equipment, now exceed $3tn, according to the Fed. None of that shows up as debt on a balance sheet. It still creates a legal duty to pay that behaves like debt the moment revenue falls short of expectations.
Separately, debt issuance tied to advanced computing, mostly data centers and chips, reached more than $680bn by October 2026, a fivefold rise from $130bn in 2023. The category has climbed from under 6% of investment-grade nonfinancial corporate debt issuance three years ago to roughly 31% now.
The Fed's own tally does not fully capture private credit, where disclosure is thinner and lending to AI infrastructure projects is harder to track. The report does not attempt to size that gap, and figures beyond October are not yet available.

What the Fed means by circular financing
The report also describes a lending pattern the Fed calls circular financing. Hyperscalers and chipmakers make equity investments in smaller AI firms, which then use that money to buy computing services or hardware from the very companies that funded them.
The arrangement can make the sector's growth look larger than outside demand alone would justify, since revenue booked by one firm can trace back to money supplied by another company in the same network. That inference follows from the mechanics the Fed describes; the report itself puts no number on the effect.
Separate coverage of the report notes that a small number of hyperscale cloud providers account for most of the capital spending behind data-center construction, concentrating funding risk among very few companies.
What exposure the banks are carrying
The largest US bank holding companies have about $1tn of exposure to large technology firms, data center developers and related special purpose entities, the Fed found. A special purpose entity is a separate legal structure set up to hold assets or debt off a parent company's books, often used to finance a single project such as a data center.
That is a meaningful slice of balance sheets that were, until recently, largely insulated from the AI buildout.
Barr, for his part, played down the chance of an immediate credit event. "Given the still-limited amount of debt backing the current boom, I'm less concerned about direct losses," he said. His comfort rests on the relatively small size of on-balance-sheet debt compared with the broader corporate bond market, not on the off-balance-sheet liabilities or the circular financing loops he flagged elsewhere in the same speech.
Barr also warned that AI trading agents deployed across banks and asset managers could react to shocks in similar ways at the same time, speeding bank runs and pushing markets toward herd-like selling in a downturn.

What analysts make of the numbers
Analysts who track corporate credit read the numbers differently from a simple default count. On-balance-sheet debt remains modest next to the broader bond market, so a wave of AI-company bankruptcies looks like the less likely outcome.
The bigger danger, in that reading, is a repricing shock that moves at once through equity markets, off-balance-sheet contracts, private credit funds and bank exposure, each channel amplifying the others in the way Barr described in Milan.
Why the dot-com era keeps coming up
The Fed has drawn a comparison between the current AI buildout and the telecommunications investment cycle of the late 1990s and early 2000s, when debt-financed infrastructure spending outran the revenue that eventually arrived. The parallel is not exact. Telecom debt sat mostly on balance sheets. A sizable share of AI financing does not.
That difference sits at the center of the Fed's concern. A shock that hits equity valuations for AI companies would not stay confined to equity markets. It would run through the credit backing those firms, the off-balance-sheet contracts tied to data centers, and the bank exposures layered underneath, according to the structure Barr laid out in Milan.

What the Fed has not spelled out
The Fed has not detailed how supervisors might address circular financing arrangements, nor how future stress tests would separate genuine productivity gains from speculative froth in AI valuations. Barr's speech and the report both describe the risk; neither lays out a specific supervisory response. Those questions stay open for now.
Three years ago, advanced-computing debt was a rounding error in investment-grade issuance. By October, it was nearly one bond in three.