The Federal Reserve has identified more than $3 trillion in potential obligations tied to the construction of artificial intelligence data centres that do not appear as conventional debt on the balance sheets of seven major technology companies.
The commitments include more than $1 trillion in signed leases that have yet to begin, alongside more than $2 trillion in take-or-pay contracts and guarantees covering the residual value of assets, according to the Fed’s November Financial Stability Report.
The figures relate to Amazon, Alphabet, Broadcom, Meta, Microsoft, Nvidia and Oracle. Fed governor Michael Barr said in a speech in Milan that the “potential maximum nominal value” of the obligations exceeded $3 trillion.
He added that on-balance-sheet debt at the seven biggest AI firms was expected to surpass $800 billion before the end of the year. The central bank cautioned that the off-balance-sheet figures are undiscounted and “not directly comparable” with debt.
Federal Reserve warns over AI data-centre debt
The report’s new section, on leverage in the data-centre buildout, examines how major technology companies are supporting projects without immediately recording the full exposure on their own balance sheets.
Data-centre developers use leases and other arrangements to finance construction, while the riskier portions of the resulting debt can be held by high-yield bond funds, life insurers, business development companies and private debt vehicles.
The Fed said the arrangements allowed companies with stronger credit ratings to support projects backed by weaker borrowers “without having an immediate, direct effect on the balance sheets” of the larger firms.
America’s largest bank holding companies have close to $1 trillion of exposure to technology companies, data-centre developers, AI-related special purpose entities, private credit vehicles and utilities, the report said.
Mr Barr warned that the lack of transparency “can exacerbate selloffs when investor sentiment shifts, as it is hard to know where risk resides”. Fed chair Kevin Warsh said staff had begun modelling a range of AI-related scenarios, from a growth boom to a disappointing outcome.
Borrowing linked to advanced computing has accelerated sharply. Debt issuance reached more than $680 billion by October, around five times the $130 billion recorded during the whole of 2023.
Advanced computing accounted for about 31% of non-financial investment-grade corporate debt issuance this year, compared with less than 6% in 2023. The Fed estimated that outstanding public debt linked to the sector had risen above $1.1 trillion, while noting that private credit deals made the overall total difficult to measure.
Concern among financial market participants has also increased. In a survey of 23 contacts at broker-dealers, banks, investment funds and advisory firms, 87% identified AI as a significant risk to the financial system, up from 50% in the previous survey in May. A broad correction in risk assets was cited by 52%.
The structure of the financing means that losses could spread beyond the technology companies if demand for computing capacity falls short. Mr Barr said guarantees showed the extent to which hyperscalers and chip producers could be required to absorb losses if borrowers defaulted.
Nvidia’s filings illustrate the scale of some of those commitments. The chipmaker agreed in August to residual value guarantees connected with an AI campus in Ohio, where an OpenAI affiliate is expected to lease about 4.25 gigawatts of IT capacity.
If the tenant became insolvent or failed to meet lease payments, Nvidia would generally cover the difference between the guaranteed minimum value and the amount recovered. The initial commitment was capped at $105 billion, subject to conditions including the sites becoming ready for service, which was expected in 2028. OpenAI agreed to reimburse Nvidia for any payments made by the chipmaker.
Nvidia’s latest quarterly filing also listed $3.5 billion in guarantees for data-centre leases held by AI cloud partners. Those partners had placed $712 million in escrow against the guarantees.
Banks judged resilient despite rising exposure
Mr Barr said he was less concerned about direct losses to banks, arguing that the amount of debt supporting the current boom remained limited. The Fed described the banking system as sound and resilient, with high regulatory capital ratios.
However, a separate annual assessment by the New York Fed found that a measure of banks’ liquidity stress had reached a level not seen since 2014. In a scenario repeating the interest-rate shock of 2022, banks would face a capital shortfall of $151.1 billion, broadly unchanged from the previous year.
The central bank’s report said major technology companies might absorb losses during a future downturn, but warned that their economic exposure could be “substantially larger than suggested by their on-balance-sheet liabilities”.
