The European Central Bank has warned that the financing of the artificial intelligence boom is beginning to resemble the debt-fuelled excesses seen before the global financial crisis, with major technology companies carrying more than $4 trillion (£3.1 trillion) in off-balance-sheet commitments.
The figure, contained in the ECB’s twice-yearly Financial Stability Review, covers Alphabet, Amazon, Meta, Microsoft, Nvidia, Oracle, Broadcom and Anthropic. The central bank said those commitments were “well in excess” of the companies’ recorded on-balance-sheet liabilities.
“The rapid shift towards debt finance invokes historical parallels with the dotcom and subprime bubbles,” ECB authors wrote in an assessment of the borrowing supporting the AI build-out.
The warning follows a separate assessment by the US Federal Reserve, which identified more than $3 trillion in signed leases, take-or-pay contracts and residual value guarantees at seven major technology companies. The two estimates use different methods and do not cover exactly the same businesses.
ECB highlights hidden liabilities in AI expansion
The ECB’s concern is that arrangements which function economically like debt are not always recorded as liabilities. Hyperscalers are signing long-term contracts to use data centres and chips owned by other companies, while those owners borrow against the expected payments from banks, bond investors and private credit funds.
Under US accounting rules, the commitments may not appear on a hyperscaler’s balance sheet until the relevant asset enters use. The ECB said that meant cash-flow obligations could be debt-like in practice without being reported as debt.
One example cited in the review is Meta’s $27 billion Hyperion data centre in Louisiana. It was built through a joint venture in which Blue Owl Capital owned 80 per cent, leaving the full debt off Meta’s balance sheet. The bonds were rated A+ and carried a yield of 6.58 per cent.
The central bank said a downturn in earnings could expose the scale of the borrowing. “The full debt incurred for the AI buildout would gradually be recognised on balance sheets,” it said, increasing leverage and putting pressure on credit ratings.
The ECB drew a comparison with Citigroup’s decision in December 2007 to bring $49 billion of assets held by structured investment vehicles back on to its balance sheet. The vehicles had previously been kept off the bank’s books, but the move was intended to prevent ratings downgrades and forced asset sales.
Concerns over circular financing and credit ratings
The review also questioned whether reported demand for AI infrastructure is being amplified by financing arrangements between companies in the sector.
Cloud providers and AI laboratories buy or lease chips from chipmakers, which may also help finance those purchases. The ECB said such deals could “inflate the expected revenues of both the chipmaker and the service provider”, because demand reported by one company may partly result from the other’s financing decisions.
The review noted that “other income” at Alphabet and Amazon included the revaluation of stakes in Anthropic. It also referred to Nvidia’s investments in CoreWeave, OpenAI and Anthropic.
ECB vice-president Boris Vujčić said in the review’s foreword that “circular financing is making intrinsic demand hard to gauge”.
The central bank also identified signs that investors may be less confident in the risk assessments attached to technology debt. Bond spreads for hyperscalers have widened against those of companies with the same credit ratings, particularly for longer-dated debt.
That, the ECB said, “could be an indication that markets currently view rating agencies’ risk assessments as too benign”. Credit default swap spreads for several major technology companies have also risen to record levels, it said.
Short-lived equipment funded by long-term debt
Another concern is the mismatch between the lifespan of the equipment being bought and the duration of the borrowing used to fund it.
“The maturity of hyperscalers’ outstanding debt exceeds the economic lifetimes of a substantial share of the physical capital funded by this debt, especially in the form of advanced chips,” the ECB authors wrote.
Market intelligence cited by the central bank suggests AI equipment may last for only one to three years, while the International Monetary Fund puts the lifespan of graphics processing units at about two years. By contrast, standard equipment is estimated to last seven years.
Most hyperscalers have extended the accounting life assigned to their servers from three years in 2020 to six years in 2026, according to the review.
The ECB said non-bank investors, including funds, had significantly increased their exposure to AI and begun buying debt issued by AI companies. It pointed to the near-collapse of the US hedge fund Situational Awareness in July 2026 as an example of leverage and concentrated exposure combining outside the banking sector.
ECB says earnings are still keeping pace
The central bank stressed that the comparison with earlier crises has limits. “So far, however, hyperscalers’ earnings have broadly kept up with expectations,” the authors wrote.
Vujčić said banks might not be materially exposed to a correction in AI-related assets, but could still face spillovers through their links with non-bank financial institutions.
The review distinguished between the dotcom crash and the subprime-led financial crisis. The dotcom bust produced only a shallow recession, the ECB said, partly because a large share of the financing had come from equity rather than borrowing.
By contrast, credit booms that end in a bust have been associated with a significant and prolonged contraction in economic output. The authors warned that a shock originating in the United States would probably cause a significant contraction in euro area gross domestic product.
“The rapid acceleration of debt financing raises concerns that the AI buildout might become debt-driven and unsustainable,” they wrote.
The scale of the commitments was identified by ECB staff using company filings, with an in-house model helping to extract off-balance-sheet figures before the results were checked manually. Anthropic’s figures were taken from its initial public offering filing.
For now, the review said, earnings remain the main measure supporting the expansion, while the other risks identified by the ECB depend on whether those earnings continue to meet expectations. The central bank is expected to reassess the position in May if it maintains this year’s review schedule.
