The Guarantee Machine: FT Finds $300 Billion of AI Debt Hiding in Big Tech's Fine Print

There is a number buried in the footnotes of Big Tech’s securities filings that you will not find on any balance sheet: up to $300 billion. That, according to a Financial Times investigation published September 20, 2026, is the total value of guarantees that Meta, Nvidia, Broadcom and other technology companies have issued over the past year to back debt raised for AI data centres and chips. The borrowing itself sits mostly in special-purpose vehicles — legally separate entities that own the buildings and the GPUs — while the tech giants stand behind it with promises that accounting rules treat as contingent liabilities rather than ordinary debt.

The result is an AI buildout that looks dramatically cheaper on paper than it is. Alphabet’s data-centre guarantees jumped from $16.9 billion to $43.8 billion in just six months, the FT reports — a near-tripling of the commitment. Yet less than 2% of that figure shows up as an actual liability on Alphabet’s balance sheet. The rest lives in disclosure tables that no leverage ratio, covenant test, or screening filter consumes automatically.

How the guarantee machine works

The key instrument is called a residual value guarantee. The mechanics are simple: investors lend billions to build an AI data centre, and the tech company promises that if the facility or its chips are eventually sold for less than an agreed value, the company will cover part of the shortfall. Because the debt belongs primarily to the special-purpose vehicle rather than the guarantor, the exposure does not initially appear as conventional borrowing.

Meta popularized the structure with its Hyperion data-centre project in Richland Parish, Louisiana. An SPV called Beignet Investor — majority-owned by private credit giant Blue Owl Capital, with Meta holding a 20% stake — raised roughly $30 billion, including about $27 billion in loans from lenders such as Pimco, BlackRock and Apollo. The vehicle owns the data centre; Meta leases it. When state regulators approved the gas plants powering the campus, Meta had already restructured the deal through a Delaware holding company the same day.

Broadcom later took on a similar guarantee around financing for chips destined for Anthropic, while Nvidia provided guarantees connected to a SoftBank data-centre development for OpenAI. The FT tally — up to $300 billion across the sector — comes on top of roughly $3 trillion in broader off-balance-sheet commitments that a Wall Street Journal analysis surfaced in August.

Google’s three-sided structure

The FT’s clearest illustration of how risk migrates is the Fluidstack arrangement. Fluidstack, an AI cloud provider, leases and builds data centres stocked with Google’s TPU chips, then rents the capacity to Anthropic. Google guarantees Fluidstack’s lease payments, which lets lenders finance the buildout at lower rates because Alphabet is effectively co-signing the loan.

The structure puts three companies between the debt and the balance sheet: the SPV that owns the building, the cloud provider that leases it, and the chipmaker that guarantees the payments. Each layer is legal and disclosed. But the credit exposure has quietly migrated from hyperscaler balance sheets to vehicle lender syndicates — asset managers and private credit funds become the effective infrastructure creditors of the AI era.

Why now: the Oracle stress signal

The timing of the FT report is not accidental. On Friday, about $18 billion of loans tied to a data centre leased to Oracle in New Mexico slid into stressed territory, with the debt quoted at 89 to 91 cents on the dollar as investors grew wary of construction delays and local opposition. It is the most concrete signal yet that AI infrastructure debt can wobble — and that when it does, the guarantees behind it suddenly matter.

The guarantees only stay cheap on paper while tenants like Anthropic and OpenAI keep paying their leases and the data centres keep generating revenue that justifies their cost. If AI demand slows while enormous amounts of capacity are already built, unused facilities could fall in value, older GPUs could depreciate faster than expected, and Big Tech could suddenly have to absorb losses that investors never saw counted as debt.

Legal, disclosed — and structurally invisible

None of this is concealment in the fraudulent sense. Companies publish the figures that made the FT’s reporting possible; the structures are permitted under current GAAP, and credit-rating agencies do examine the guarantees when they assess creditworthiness. The issue is architectural: a number on the balance sheet propagates automatically into every ratio and model built on top of it, while a number in a footnote enters those calculations only when a human reads it, judges it, and manually adds it back.

Both sets of numbers are public. Only one of them shows up by default in every leverage screen. Regulators have begun to notice — a letter to Treasury Secretary Scott Bessent earlier this year flagged concerns about AI-related debt structures, and Senator Elizabeth Warren has pressed the Financial Stability Oversight Council to probe what she calls an “AI debt bubble.” Whether accounting standards evolve before the next wave of commitments is contracted is now an open question.

For an industry whose bulls argue that AI capex is self-funding and whose bears warn of a Enron-style financing spiral, the FT’s $300 billion figure lands somewhere clarifying: the AI buildout is not necessarily over-leveraged, but a substantial share of its financial weight now lives in the part of the reporting stack that requires deliberate human retrieval rather than automatic propagation. The companies published the numbers. The question is which analysts — and which risk models — are actually reading them.