AI Banking Risk Grows as Fed Hikes and Regulators Loosen Rules

Calling it a hidden crisis misses the point

Talk of an AI banking crisis usually assumes officials are hiding something. The record from the past six months points the other way. Supervisors in Washington, London, Frankfurt and Tokyo have published a steady run of warnings about AI-linked debt, cyber risk and private credit. What deserves more attention is the timing. Some of the rules that would govern this risk are being loosened, borrowing costs are rising again, and the AI infrastructure boom depends heavily on debt. The danger is not a secret. It is a mismatch between how fast risk is building and how quickly oversight is keeping up.

Interest rates are going up again

The Federal Reserve raised its policy range by a quarter point to 3.75%–4.00% on September 16 in a 12-0 vote. It was the Fed's first hike since 2023.1620 The median projection for the year-end rate rose to 4.1% from 3.8%, and 16 of 18 participants expected at least one more increase in 2026.16 Minutes released October 7 said every participant backed a higher range. A couple of officials also said tighter policy would help keep price pressures from energy disruptions and AI-related demand from spreading into broader inflation.18

Forecasts for the next move differ. In late September, futures traders put roughly 70% odds on an October hike.14 Fixed-income markets were reported to be pricing hikes at both remaining 2026 meetings.15 By October 8, prediction-market contracts leaned heavily toward an October hold, with a December move much more likely.16 A soft September jobs report and the meeting's closeness to the midterm elections help explain the shift.18 U.S. Bancorp's economists expect 75 basis points of total tightening between October 2026 and March 2027. That would bring the peak range to 4.25%–4.50%.17

The exact date of the next hike matters less than where rates are heading. Long-term borrowing costs have risen. The 10-year Treasury yield reportedly reached 5.35%, its highest level since 2002.18 Daily Fed data show it hovering between about 5.2% and 5.3% in early October, with the bank prime rate at 7%.11 Futures imply an effective fed funds rate near 4.7% by October 2027.13 For long-dated infrastructure loans that depend on refinancing, a market that is still tightening is a real problem.

Banks' exposure to AI debt is growing

This is where interest rates and AI meet. The Bank for International Settlements estimates that the five largest tech companies' AI capital spending will top $1 trillion across 2025 and 2026. That money moves through a chain of hyperscalers, project vehicles, private-credit funds, insurers and banks.24 The Bank of England counts about $450 billion in global AI-related debt issuance by early September, more than double the total for all of 2025. It flagged AI and debt as a top stability risk on September 30.25

Estimates of how much of this sits with banks vary widely, and the gap is the most important open question in this story. The Chicago Fed puts direct bank exposure to AI-adjacent industries at about 0.8% of total assets, with delinquency rates close to those of wider loan books.24 Measured against capital, the picture looks less comfortable. Large banks' outstanding loans to that group average about 9% of Tier 1 capital, and committed exposure is closer to 25%.22 Columbia finance professor Stijn Van Nieuwerburgh adds up mortgages, syndicated loans and project debt and gets roughly $800 billion. He says many banks are close to their concentration limits for data-center lending.30

The two sets of figures measure different things. The Chicago Fed counts direct, reportable lending. The larger figures try to capture the whole financing web. The Chicago Fed itself admits that banks probably have more exposure through loans to nonbank lenders, and that these indirect channels are hard to measure with regulatory data.2224 The smaller number is a minimum, not the full picture.

Signs of strain

The warnings now include some market evidence. Banks in a syndicate including Santander and Jefferies reportedly quoted about $18 billion of loans tied to an Oracle-leased data center in New Mexico at 89 to 91 cents on the dollar.22 JPMorgan marked down loans to private-credit firms in March, especially software loans seen as vulnerable to AI. That cut the amount private-credit funds could borrow against that collateral.26 Fitch puts the trailing 12-month U.S. private-credit default rate at 6.3% through August.28

Some of the bluntest warnings come from inside the industry. Carlyle says AI compute may need about $1 trillion of private-credit financing, more than half of what the industry manages. It urges lenders to cap exposure so they do not repeat the concentration mistakes of the software-loan boom.23 Carlyle's Jason Thomas compared some data-center underwriting to pre-2008 mortgage lending. His point was that loans are priced on the hyperscaler sponsor's credit rating rather than on whether the project can pay its own way.25 Fortress co-CEO Jack Neumark told lenders their returns are capped while they still carry the full risk of falling asset values and illiquid positions.28 Van Nieuwerburgh notes that debt can make up as much as 90% of a project's cost, so it would not take a large shock to put that debt in trouble.30

The Bank of England has described how rates and technology combine. Long-term debt is financing buildings that could lose value sooner than expected if they cannot support newer AI hardware. Uncertain demand for computing, limited power supply and fast depreciation add to the risk.22 Moody's has warned against underwriting these facilities on projected demand rather than signed contracts.21

Regulators are moving in opposite directions

The regulatory picture is mixed: supervisors are watching more closely while some formal standards get lighter.

On the supervision side, the New York Fed has been asking JPMorgan, Wells Fargo, Barclays and Morgan Stanley since the spring about their lending to private-credit firms, their risk controls and the quality of their collateral.26 Japan's Financial Services Agency is looking at how its largest banks and life insurers fund mostly U.S. data centers. It says it is not trying to choke off lending.29 The Financial Stability Board, a G20 body, has proposed 12 AI governance practices, with a final report due this month. Its chair, Andrew Bailey, told G20 ministers in August that cyber risk from frontier AI models is the most immediate threat to financial stability.1

On the rules side, the April 17 interagency model-risk guidance, SR 26-2, replaced the 2011 standard, SR 11-7. It leaves generative and agentic AI out of its scope, describing them as novel and rapidly evolving.37 The Center for Democracy and Technology points out that SR 26-2 also dropped earlier language telling banks to use models more conservatively when they are uncertain about them. In its view, that could raise banks' appetite for risk.8 Fed Vice Chair for Supervision Michelle Bowman described the parallel rewrite of third-party risk guidance as an effort to remove vagueness and avoid holding back innovation.4 The September 11 proposal for that rewrite takes a technology-neutral approach and sets no AI-specific requirements.7 The Financial Stability Oversight Council set up a standing AI working group rather than writing rules.4

State supervisors are partly filling the gap. The Conference of State Bank Supervisors released an AI examination framework on September 16 that covers banks and nonbanks. Analysts say it addresses third-party AI risks that federal guidance does not.62 Capital rules may push the other way. American Banker reports that proposals to lower risk-weight floors for some securitizations and corporate loans could make private-credit assets more attractive for banks to hold.30

What to watch

The evidence does not show a crisis underway. Direct delinquencies look normal, and analysts tracking data-center finance describe a mapped vulnerability rather than a chain of failures.2224 But three trends are converging. Rates are rising, and markets expect them to stay restrictive through 2027.13 Highly leveraged AI projects depend on cheap refinancing and assets that hold their value. And the formal rules for AI, third-party risk and capital are becoming lighter in ways that reduce how cautious banks are required to be.

The weakest link is the least visible one. Risk that banks pass on through synthetic risk transfers, in which a bank keeps its loans but sells protection against losses, can come back to them through their loans to the funds that bought it. The Bank of England warns this makes the distribution of risk harder to see.22 Over the next few months, watch the prices of syndicated data-center loans, any markdowns of private-credit collateral during bank earnings season, and whether the Fed's December projections assume rates stay high. If AI debt does cause serious problems, it will most likely come through these channels, which today's monitoring covers least well.