The Fed's AI Warning, Explained: Big Tech's Unstarted Leases Now Top $1.3 Trillion

The Federal Reserve's new Financial Stability Report puts artificial intelligence at the top of Wall Street's worry list. The headline survey is smaller than it sounds. The Fed's own analysis is more important: Big Tech's commitments that are not counted as debt have grown even faster than its borrowing, and company filings released since the Fed's data cutoff show one of them, leases signed but not yet started, is still growing fast.

The Federal Reserve Board published its twice-yearly Financial Stability Report on Friday, November 13. Within hours, the line spreading across markets and social media was that the Fed now sees AI as the biggest threat to the financial system. The report's own figure is that 87% of the market contacts surveyed named AI as a risk, up from 50% in the spring.

That figure is real, but it is the weakest part of the report. The more consequential material is in a box titled "Leverage in the Data Center Buildout." There, Fed staff document how the largest technology companies are financing the AI buildout through leases, guarantees and contracts that do not show up as debt on their balance sheets. We went back to those companies' latest filings. One of those obligations, leases signed but not yet started, has grown by roughly a fifth since the period the Fed measured.

What the Fed said, and what 23 people said

The 87% comes from the report's "Survey of Salient Risks to Financial Stability." Staff at the Federal Reserve Bank of New York collected responses in September and October from 23 contacts at broker-dealers, banks, investment funds and advisory firms. The report says this summary "should not be interpreted as representing the views of the Federal Reserve Board or the Federal Reserve Bank of New York."

In absolute terms, 87% of 23 is 20 people. The spring survey, published in the May 2026 report, drew on 20 contacts, and half of them, 10 people, cited AI. The direction of the change matters, because AI went from one concern among several to the most cited. But this is a poll of market professionals, not a Fed finding. The respondents named three AI concerns: stretched equity valuations that could fall if earnings disappoint, cyberattacks by state actors or "damage caused by frontier AI models themselves," and "complex forms of AI-related financing." The next most cited risks were a broad correction in risk assets (52%), higher long-term interest rates (48%), cyberattacks (43%) and U.S. fiscal debt sustainability (43%).

The Fed's own judgments are more measured. It describes equity valuations as elevated, "partly driven by market enthusiasm for super intelligence (SI)," and lists disappointing returns on business investment as one of three near-term risks. In that scenario, weaker-than-expected corporate earnings would lower asset prices, tighten credit and cut capital spending. On cyber risk, the report notes that LLMs and agentic systems have become better at finding and exploiting vulnerabilities. It also says reports of agentic systems being involved in cyber events have become more frequent since May.

The debt numbers

The Fed's figures on borrowing are striking on their own. According to the report:

• Debt issued to finance advanced computing infrastructure rose to more than $680 billion by October 2026, about five times the $130 billion issued in all of 2023. The estimate, drawn from Bank of America Global Research and Bloomberg data, does not capture all private deals, so it likely understates the total.
• That borrowing grew from less than 6% of investment-grade debt issued by non-financial companies in 2023 to about 31% in 2026.
• Outstanding public debt related to advanced computing is estimated at more than $1.1 trillion. The Fed adds that sizable private credit deals make the true total "more difficult to measure."

The Fed also compares the buildout with the telecom boom of the late 1990s. Measured against GDP, cumulative capital spending by five hyperscalers (Amazon, Alphabet, Meta, Microsoft and Oracle) has tracked the telecom industry's spending after 1995 relatively closely. The Fed says it has been accelerating lately, "suggesting that the pace of the current cycle's investment may surpass that of the 1990s telecom cycle relatively soon." The key difference so far is funding. The hyperscalers' strong cash flows have reduced their need to borrow to fund similar levels of investment, "at least until this year."

The obligations that are not debt

The section most coverage skims is where the Fed is most specific. Debt balances have grown, the report says, but "other obligations have grown even more rapidly." For seven companies (Amazon, Alphabet, Broadcom, Meta, Microsoft, Nvidia and Oracle), the Fed counts two categories:

• Leases that have not yet started, with undiscounted values above $1 trillion. These are contracts already signed, typically for data centers still being built. They stay off the balance sheet until the company takes possession of the building.
• Other obligations, such as take-or-pay contracts and residual value guarantees, with undiscounted values above $2 trillion. A take-or-pay contract obliges the buyer to pay whether or not it uses what it ordered, whether that is power, capacity or hardware. A residual value guarantee promises that an asset, such as a GPU used as loan collateral, will be worth at least a set amount later.

The Fed attaches an important caveat: these figures are undiscounted, so they are "not directly comparable with on-balance-sheet debt, which is discounted." A dollar of rent due in 2040 counts the same as a dollar due next year. Readers should not add the $1 trillion and $2 trillion to the $1.1 trillion of debt and call the result AI debt. The three numbers also cover different groups of companies.

The mechanism explains why the Fed cares. A hyperscaler signs a 15- or 20-year lease with a data center developer, and the developer uses that lease to raise money to build. A GPU maker guarantees what its chips will be worth so that a smaller AI company can borrow against them. The same large firms often invest equity in those smaller firms too, which the Fed calls a "circular financing" dynamic. The strong-credit company supports the weaker one's borrowing "without having an immediate, direct effect on the balance sheets of the stronger-credit firms."

The Fed is careful about what that means. It could be stabilizing: the biggest companies "may absorb losses in the event of a future downturn," shielding the rest of the system. But it also means their "economic exposure to such a downturn may be substantially larger than suggested by their on-balance-sheet liabilities."

The lease pile has grown since the Fed's snapshot

The Fed's chart of these obligations runs through the second quarter of 2026. Since then, all five hyperscalers have filed new quarterly reports, and each discloses its signed-but-not-yet-commenced leases in a single line of its notes. We compared each company's figure in the filing that covers roughly the Fed's period with its most recent one.

Sources: Microsoft 10-K and 10-Q; Meta Q2 and Q3 10-Qs; Oracle 10-K and 10-Q; Amazon Q2 and Q3 10-Qs; Alphabet Q2 and Q3 10-Qs. Figures are undiscounted future lease payments as each company reports them. Oracle's fiscal quarters end one month earlier than the others'.

Two things stand out. First, the earlier total for these five companies, about $1.09 trillion, is consistent with the Fed's "more than $1 trillion" for its seven-firm group. That suggests the Fed's number is mostly a hyperscaler number. We did not add Nvidia's or Broadcom's lease disclosures. Second, about three months later the five hyperscalers alone stand at roughly $1.3 trillion. Their pile of signed-but-unstarted leases grew by a net $213 billion in roughly three months.

The commitments are also large relative to the leases these companies already carry. On a like-for-like undiscounted basis, Microsoft's unstarted leases are about 2.8 times all the lease payments already on its books ($128.3 billion). Alphabet's and Oracle's are about 4.5 times theirs ($28.8 billion and $63.2 billion). Amazon is closer to parity, at about 1.2 times ($139.4 billion). Meta's quarterly filing has no comparable table, but its unstarted leases are more than four times the $84 billion of bonds it has outstanding. That comparison is rough, because a lease is not a bond.

The filings contain qualifiers that headline totals omit. Microsoft says some of its arrangements are "subject to certain contractual conditions being met." Amazon does not break its figure down by asset type. Meta's and Alphabet's latest descriptions newly include energy infrastructure alongside data centers. So not every dollar is a server hall, and not every dollar is certain to be paid. These are also separate from the companies' even larger purchase commitments. Alphabet, for example, reports $924.4 billion of material purchase commitments, and Meta reports $454.52 billion of non-cancelable contractual commitments. Those are different categories, and we have not added them to the lease totals.

Who ends up holding the risk

The Fed traces where the exposure sits. The hyperscalers' and GPU makers' own bonds are held mainly by mutual funds, pension funds and insurers. Much of the off-balance-sheet financing runs through special purpose entities that issue both senior and riskier subordinated debt. Those riskier tranches and equity slices are held by high-yield bond funds, life insurers, business development companies and other private debt vehicles. The largest U.S. bank holding companies have close to $1 trillion of exposure to large tech firms, data center developers, related special purpose entities, private credit vehicles and utilities. The Fed notes those exposures are evaluated in its annual stress tests.

In other words, if AI demand disappoints, the first losses may not fall on Microsoft or Meta shareholders. They may fall on the holders of the riskiest slices of data center financings, including life insurers and high-yield and private debt funds.

The case against alarm

There are good reasons not to read this as a 2000-style crash in waiting. The companies signing these leases are among the most creditworthy borrowers in the world, with cash flows the telecom carriers of the 1990s never had. A signed lease is a commitment, not a loss. If the computing capacity is used, it generates revenue. The Fed's own assessment is that the banking system remains sound and resilient. It also sees a plausible cushioning effect, in which the giants absorb losses that would otherwise land on weaker firms. And the survey that drove the headlines reflects the views of 23 market participants, not a formal Fed judgment.

The Fed's concern is narrower and harder to dismiss. The more a buildout depends on contingent, undiscounted and off-balance-sheet commitments, the harder it is for anyone, including the companies' own investors, to see how much is riding on AI demand arriving on schedule. Leases that start between 2026 and 2036 lock in costs for years, whatever happens to AI pricing in the meantime.

What to watch

Three markers will show whether this risk is growing or easing. The first is the next round of filings: Oracle typically reports its November quarter in December, and the calendar-year companies file annual reports early next year. Unstarted leases that keep growing faster than leases taking effect would mean the off-balance-sheet pile is still expanding. The second is how these commitments convert. As leases commence, they move onto balance sheets as liabilities, and depreciation and lease costs start hitting earnings. The third is the private credit and special purpose vehicle market the Fed says it cannot fully measure. Any stress in subordinated data center debt would show up there first.

For a running log of the deals behind these numbers, see our AI data center deals tracker.