Big Tech has written $300 billion of AI guarantees that stay off its balance sheets

Technology companies have provided as much as $300 billion in guarantees over the past year to finance AI data centers and chips, supporting the construction boom without recording most of the exposure as debt, according to a Financial Times investigation reported this week. The number is an aggregate across several distinct arrangements rather than one company's hidden liability, which is the detail that determines who actually carries the risk.

The largest named structures sit with different guarantors. NVDA provided $105 billion in guarantees to SB Energy for an Ohio data center campus intended for OpenAI. META supplied a $28 billion guarantee for its Hyperion data center venture with Blue Owl in Louisiana, an arrangement that helped the project raise $27 billion at borrowing costs only modestly above Meta's own debt while keeping most of the obligation outside Meta's reported balance sheet. GOOGL data center guarantees are reported to have risen from $16.9 billion to $43.8 billion over six months, with less than 2% of that figure reaching the balance sheet. Meta, xAI, Oracle and CoreWeave have together moved more than $120 billion of data center financing debt to outside investors through special purpose vehicles.

The mechanism is a residual value guarantee or an SPV that borrows against the asset, so the obligation does not appear as ordinary on balance sheet debt for the company standing behind it. That is legal and disclosed in footnotes, but it means conventional leverage screens understate how much of the AI buildout these companies have underwritten.

Running alongside the financing story, though not caused by it, Anthropic CEO Dario Amodei published a proposal on September 12 arguing that frontier labs should slow the pace of capability gains. His plan is specific rather than rhetorical: embedded third party evaluators with employee-like access, domestic coordination among labs on safety standards, and narrow international coordination. OpenAI's Sam Altman agreed and committed OpenAI to match the independent evaluator step, saying the approach was a good idea. Elon Musk's endorsement was a two word reply on X, not a policy commitment of his own, and the difference is worth keeping straight.

For investors the practical question is whether the guarantee structures get repriced. If AI infrastructure demand softens, the residual value assumptions behind these guarantees are what would be tested first, and the exposure sits with the guarantor rather than the SPV's lenders. Watch the footnote disclosures in the next round of quarterly filings, where the guarantee balances are where this risk is actually visible.