Fed Financial Stability Report: AI Data Center Debt Hits 31% of Investment-Grade Issuance
13 hour ago / Read about 28 minute
Source:TechTimes

Pexels/Déji Fadahunsi

The Federal Reserve now counts borrowing for AI computing infrastructure at about 31% of all nonfinancial investment-grade corporate debt issuance, up from under 6% in 2023, according to the Financial Stability Report it published Friday, Nov. 13. Investment-grade debt is where the highest-rated companies borrow and where bond funds, insurers and pension managers buy, so a one-third share puts AI financing inside portfolios that were never built as technology bets.

The same report carries the New York Fed's fall survey of 23 market contacts, and AI was the most-cited risk. Secondary summaries put the share at 87% naming AI as a salient risk to U.S. financial stability, up from 50% in the spring. That works out to about 20 respondents, and the report states that the survey reflects respondents' views, not those of the Fed Board or the New York Fed.

The survey closed before Nvidia announced on Nov. 12 that CEO Jensen Huang had resigned effective immediately, and Nvidia is one of seven companies in the Fed's tally of data-center commitments.

Fed Counts More Than $680 Billion of AI Computing Debt Issued Through October

The Fed's boxed analysis, titled "Leverage in the Data Center Buildout," says debt issued to finance advanced computing rose to more than $680 billion by October 2026, roughly five times the $130 billion issued in all of 2023.

That borrowing accounts for about 31% of nonfinancial investment-grade corporate debt issuance this year, up from under 6% in 2023.

The report estimates that outstanding public debt tied to advanced computing stands at more than $1.1 trillion. Issuance is a one-year flow and outstanding debt is a stock, so the two figures should not be added together. The Fed also notes the mix is evolving from mostly corporate bonds toward a combination of public and private debt, which makes the total harder to measure.

Leases, Take-or-Pay Contracts and Guarantees Add About $3 Trillion, Outside Debt Counts

The Fed examined seven companies: Amazon, Alphabet, Broadcom, Meta, Microsoft, Nvidia and Oracle. Together they carry more than $1 trillion in data-center leases that have not yet started, plus more than $2 trillion in take-or-pay contracts and residual value guarantees, both measured without discounting. Those figures aren't directly comparable to balance-sheet debt, but together they amount to roughly $3 trillion in obligations beyond the debt conventional counts capture.

The structure works like this: a hyperscaler signs a long-term lease with a third-party developer, and the developer uses that lease to secure the loan that pays for construction. Hyperscalers and GPU makers often add equity stakes in the smaller developers, which the Fed calls "circular financing." The report says the big firms' true exposure may be much larger than their on-balance-sheet liabilities suggest.

Two features make these promises harder to size than debt. Special-purpose entities are legal shells commonly used to isolate a sponsor's financial risk, and consolidation rules decide whether their liabilities appear on the sponsor's books. A take-or-pay clause obliges a buyer to accept delivery or pay anyway, yet outside oil and gas, courts often reject such clauses as unenforceable penalties. How computing-capacity versions would fare in a dispute is an open question, not a Fed finding.

GPU Resale Prices Sit Under Loans That Outlast Chip Generations

Lenders size these loans against contracted customer payments plus what the hardware is expected to be worth when the financing matures, typically three to six years out, according to a Data Center Dynamics analysis.

Nvidia has introduced a new data-center GPU architecture about every two years: Ampere in 2020, Hopper in 2022 and Blackwell in 2024. A chip that is current at signing can therefore be two generations old before a five-year loan matures.

To persuade lenders, Nvidia has agreed to cover up to 25% of the difference if pledged chips lose value, TechCrunch reported in August. Analyst Dave Friedman argues the weak point is that customer behavior and collateral value can deteriorate together, which lenders call wrong-way risk. The same company that sells the chip and partly backstops its resale value also sets how quickly the next generation arrives.

Are Banks Exposed to AI Debt?

The largest U.S. bank holding companies have close to $1 trillion of exposure to large technology firms, data-center developers, advanced-computing special-purpose entities, private credit vehicles and utilities, the report says, and the Fed evaluates those exposures in its annual supervisory stress test.

The Fed and the Office of the Comptroller of the Currency describe the banking system as sound, profitable and generally well capitalized, and the OCC reports first-half 2026 bank net income of about $116.6 billion, up 23.1% from a year earlier.

The pressure points sit outside banks. Private credit vehicles held more than $2.6 trillion in financial assets at the end of 2025, and hedge-fund leverage reached its highest level in the first quarter of 2026 since the Fed began collecting the data in 2013.

A February analysis by Chicago Fed researchers found large banks' outstanding commercial and industrial loans to AI-adjacent industries such as software, energy and semiconductors average 9% of tier 1 capital, while their committed credit averages about 25%, and that lending to nonbank lenders adds exposure that regulatory data cannot easily measure. Forbes frames the open question as whether stress-test scenarios capture simultaneous, correlated distress across banks, private credit, hedge funds, insurers and data-center financing vehicles.

Share of Fed Contacts Citing AI Rose From 9% to 87% in Four Surveys

Survey roundShare citing AI as a salient risk
Spring 20259% cited AI
Fall 202530% cited AI
Spring 202650% cited AI
Fall 202687% cited AI

The fall round surveyed 23 contacts and the spring round 20, so each percentage point rests on very few people. The 87% figure comes from secondary summaries of the Fed's chart rather than from text in the report itself.

AI was the most-cited risk, according to the report. A Diario Bitcoin summary lists the next most common as a correction in risk assets (52%, up from 35% in spring), higher long-term interest rates (48%, up from 20%), cyberattacks (43%, up from 20%) and fiscal-debt sustainability (43%, up from 30%). The pairing of those figures with their risks comes from that one summary of the Fed's chart.

Respondents grouped their AI concerns into three areas: elevated equity valuations that could fall if earnings disappoint; cybersecurity, from state-actor attacks or damage caused by frontier AI models; and complex AI-related financing.

Williams Saw No Bubble in Summer Interview; Berkeley's Judge Warns of Credit Crisis Risk

New York Fed President John Williams, whose staff run the survey, told Reuters in an interview published in early August: "I don't see this as a bubble kind of situation." He said companies doing the borrowing have high earnings and he was less worried about leverage as a financial-stability threat at that point.

UC Berkeley's Brian Judge argues in a Project Syndicate column that the AI boom could spark a credit crisis as soaring data-center debt meets uncertain returns. The Fed's November numbers, a count of issuance and commitments rather than opinion, bear more directly on that credit concern than on the stock-valuation debate.

Survey Closed Before Nvidia's CEO Resigned, So Its 87% Predates the Leadership Change

The report reflects information through Oct. 28, and the survey ran in September and October. Nvidia announced on Nov. 12 that Huang had resigned effective immediately, with CFO Colette Kress serving as interim CEO while keeping her finance role and Harvey C. Jones named chair of the board for the first time. The filing gives no reason for the departure.

Huang had been a visible voice in technology policy. He told reporters in December 2025 that the proposed GAIN AI Act, which would have required chipmakers to serve U.S. customers first, would be more harmful than the AI Diffusion Act, and the White House named him to PCAST, the President's Council of Advisors on Science and Technology, on March 25, 2026.

Nvidia will report third-quarter results on Tuesday, Nov. 17, at about 4:20 p.m. ET, with a call at 5 p.m. ET. Respondents' views may have moved since the survey closed.

Nvidia Results, Stress-Test Scenarios and Spring 2027 Survey Will Test Whether 87% Holds

Salience is not likelihood: naming a risk says it is on people's minds, not that a shock is expected, and a 23-person survey cannot supply probabilities. The survey's valuation concern is that lower-than-expected earnings could trigger a market correction.

Three tests come next. Nvidia's Nov. 17 results will show how much demand sits behind the buildout. The Fed's 2027 stress test is expected to be the first run on revised models, and the first whose results count toward bank capital requirements again. And the Fed's next semiannual Financial Stability Report, expected in spring 2027, will show whether data-center debt keeps climbing as a share of investment-grade issuance.

In short: the Fed counts about $680 billion of new debt and about $3 trillion of commitments beyond it, AI leads its survey of market contacts, banks are described as sound while private credit and hedge-fund leverage carry the pressure, and the survey itself predates Nvidia's leadership change.

This article is for informational purposes only and does not constitute investment advice. Readers should conduct their own due diligence before making investment decisions.

Frequently Asked Questions

Is the $3 trillion of leases and guarantees the same as $3 trillion of debt?

No. The Fed measured leases not yet started and take-or-pay contracts and residual value guarantees without discounting, and says they are not directly comparable with balance-sheet debt. They are promises that only turn into payments under certain conditions, such as a customer walking away or chips reselling for less than lenders expected, which is why the Fed warns that exposure may be larger than balance sheets show.

Does the Fed expect an AI-driven financial crisis?

The report describes vulnerabilities that necessitate close monitoring of the links among data-center firms, credit markets, private lenders and financing vehicles, but the 87% figure is a survey of 23 contacts and not a Fed forecast.

How could AI borrowing reach my bond fund, pension or bank?

About 31% of this year's issuance of investment-grade corporate debt is tied to AI computing, close to $1 trillion of exposure sits at the largest bank holding companies, and the Fed says the debt mix is moving toward a combination of public and private debt. Asking a fund manager or plan administrator how much AI-linked debt they hold is a reasonable question.

What does 87% of 23 contacts actually mean?

It means about 20 professionals at broker-dealers, banks and funds named AI when asked which risks they consider salient, according to secondary summaries of the Fed's chart, a measure of attention rather than probability.