# 87% of the Fed's contacts see AI as a financial risk

> The Federal Reserve's financial stability report survey jumps from 50% to 87% in six months and puts the focus on data center debt.
> By José Maldonado · Finance · 2026-11-16

Eighty-seven percent. That is the percentage of market contacts who, when asked by the Federal Reserve Bank of New York, named artificial intelligence as a risk to US financial stability. In the spring it was 50%. The jump, of almost forty points in six months, appears in the [Federal Reserve's financial risk report](https://www.federalreserve.gov/publications/files/financial-stability-report-20261113.pdf), published on November 13.

The figure should be read with its fine print before turning it into a headline. The survey was conducted by analysts at the regional bank between September and October, with 23 contacts from banks, brokerages, funds and consultancies. The report itself warns that this summary reflects the opinion of those surveyed and not that of the Federal Reserve board or the New York Fed.

With that caveat, the data point remains relevant. A small sample that moves so much in six months describes a shift in mood, and mood moves markets before balance sheets do. The next most cited risk, a correction in risk assets, was mentioned by 52%, far behind.

## What worries them is not the model, it is the bill

The report breaks the concern down into three fronts. The first is stock valuations: if results disappoint, shares can correct as quickly as they rose. The second is cybersecurity, with a new nuance: the risk comes not only from a state attack, but also from a frontier model causing harm on its own.

The third is the one that generates the fewest headlines and carries the most weight. The contacts point to the complexity of financing structures tied to AI, a euphemism for describing debt piled on debt in an industry that has yet to demonstrate stable revenue. There the report stops being a survey and becomes an X-ray.

The chapter on leverage in data center construction provides the figures the survey hints at. Debt issuance earmarked for advanced computing exceeded $680 billion in October 2026, about five times the $130 billion for all of 2023. Its share of investment-grade non-financial corporate debt issuance rose from less than 6% in 2023 to around 31% in 2026.

## A risk that is already on the balance sheet

That 31% is the uncomfortable part. When a technology is financed with debt that accounts for a third of investment-grade issuance, its fate stops being a sector issue. A dip in results does not only affect the companies selling models: it affects the bonds bought by insurers, pension funds and banks.

Government debt issued for advanced computing was estimated at over a trillion dollars in October, according to the same chapter. It is not a bankruptcy figure or a prophecy: it is the recognition that the AI rollout has been built on a maturity schedule, and maturities do not negotiate with the technology cycle.

The report adds a detail worth keeping in mind. The survey measures frequency, not severity. Although AI is the most cited risk, contacts considered that cyberattacks and the sustainability of fiscal debt would have more severe consequences if they materialize. A risk that is frequent in the conversation is not the same as a catastrophic risk.

Risk cited

Spring 2026

Autumn 2026

Artificial intelligence

50%

87%

Correction in risk assets

35%

52%

Higher long-term rates

20%

48%

Cyberattacks

30%

43%

Sustainability of fiscal debt

20%

43%

## The role of leverage

Until now, the public debate on AI focused on capabilities and employment. The report moves the conversation to older, more tested ground: who lends, with what collateral and at what term. It is the same kind of question asked in 2007 about mortgages and in 2022 about [stablecoins](https://observatorioblockchain.com/en/crypto/bitcoin-sube-industria-cripto-negociara-democratas/), and the answer usually comes late.

The difference is speed. A data center takes years to build and the debt financing it starts accruing before the first machine is switched on. If revenue arrives more slowly than expected, the mismatch is not resolved by renegotiating a cloud contract: it is resolved by selling assets or refinancing at a worse rate.

The report also mentions that hedge fund leverage is at its highest level since 2013 and concentrated among few players. That combination, little diversification and a lot of leverage, is the classic recipe for a forced sale when the market turns, and explains why regulators look more at the structure than at the technology.

None of this implies that AI will trigger a crisis. It implies that its financing is already part of the risk map that regulators monitor, and that map is updated twice a year with industry data.

## What the regulator will monitor

Financial regulators have three tools and none of them is fast. The first is requesting information: requiring banks to declare their exposure to the AI value chain, something that already happens with other concentrated sectors. The second is reviewing collateral: if a loan is backed by chips that depreciate in three years, the valuation matters more than for a property.

The third is the stress test. Simulating a 30% correction in tech stocks and seeing what happens to advanced computing debt is a cheap and politically uncomfortable exercise, because the results usually point to specific institutions. The [discussion on how public treasuries use AI models](https://observatorioblockchain.com/en/ai/ocde-haciendas-ia-fraude-fiscal/) shows that the public sector is already in that territory, albeit with less urgency.

For the individual investor, the report does not say what to buy or what to sell. It says something more useful: the part of the market growing fastest is also the part most dependent on outside financing, and that dependence introduces a risk that does not appear in product metrics.

## What the 87% does not say

It does not measure how many of those 23 contacts changed their minds because of a balance sheet analysis and how many because of reading headlines. It does not measure whether the concern will translate into smaller positions or only into more conversation. And above all, it does not measure the risk that no one has identified yet, which is the one that historically breaks the models.

What it does measure is the consensus. When a concern goes from 50% to 87% across two surveys, it stops being a minority opinion and becomes the baseline scenario the industry works with. Baseline scenarios do not cause crises on their own, but they change how those who have to decide with incomplete information behave.

The Federal Reserve published the report two weeks after the legislative elections and closed it with data up to October 28. It is a technical document, without prescriptions, that leaves one clear idea: AI is no longer discussed as a technological promise, it is discussed as a balance sheet position. And balance sheet positions, unlike promises, have a maturity date.
