Tech Giants Hide $2 Trillion in Off-Balance-Sheet AI Debt

Six AI hyperscalers carry nearly $2 trillion in off-balance-sheet lease and purchase commitments that could pressure liquidity and leverage if data center returns disappoint, Goldman Sachs and Morgan Stanley research shows.

The concentration of AI across technology, communication services and utilities reminded him "quite a lot of 1999," Mike Coop, portfolio manager at Morningstar, said last week.

Goldman Sachs tallied roughly $1 trillion in lease commitments that have not yet commenced and therefore sit only in financial statement footnotes, up from about $200 billion five years ago. Morgan Stanley separately counted $982 billion in purchase commitments across Google, Microsoft, Amazon, Nvidia and Oracle as of the first quarter.

The obligations come due as Big Tech's combined AI spending is set to surpass $730 billion this year, and bond markets are already showing early stress. Beignet bonds — issued by Meta's data center financing vehicle — have seen yields climb from 5.65 percent at issuance to about 6.95 percent.

The Accounting Loophole Behind the Hidden Debt

The mechanism traces to US GAAP ASC 842, which requires lease payment obligations to be recognized on the balance sheet only once the lease actually commences. Commitments that have not started are disclosed only in footnotes. Fitch and Moody's similarly exclude uncommenced leases from their rating metrics; only S&P takes a more conservative approach by including leases with "substantive debt characteristics."

Meta pioneered the structure with its Hyperion data center in Louisiana. The company formed Beignet, a joint venture with Blue Owl, holding just 20 percent equity while committing to lease the facility for at least 20 years. That lease guarantee enabled Beignet to issue $27 billion in amortizing bonds — a record issuance — without the debt appearing on Meta's balance sheet. The structure proved so effective that other hyperscalers have since adopted similar off-balance-sheet arrangements.

Purchase Commitments Add Another $982 Billion

Beyond leases, hyperscalers have made extensive purchase commitments for compute, chips, equipment and power. Morgan Stanley credit analysts counted $982 billion in such commitments across Google, Microsoft, Amazon, Nvidia and Oracle as of the first quarter. Combined with Goldman's $1 trillion lease estimate, the potential undisclosed financial obligations approach $2 trillion.

The concern is not purely theoretical. Meta's free cash flow dropped almost $8 billion in one year, while Google's free cash flow turned negative for the first time in the company's history. Nvidia, meanwhile, has been expanding its own financing role — working with Apollo Global, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR on a reported $500 billion AI infrastructure funding package, and separately in talks for a $250 billion data center financing deal with OpenAI. The chipmaker also announced a $500 billion deal with South Korean memory maker SK Hynix in late July.

Traditional credit metrics still look solid. Morgan Stanley data shows the hyperscaler group's average net leverage at just 0.5 times, below the broader tech sector's 0.8 times and far below the 1.8 times average for US non-financial companies. Cash holdings still exceed book debt. But analysts argue these metrics miss the growing web of contractual obligations that will eventually surface.

For bond investors, the question is whether data center capital expenditure will deliver the returns needed to service these obligations. Beignet bond yields have already moved from 5.65 percent at issuance last fall to about 6.95 percent, though analysts note some of the pressure reflects the broader rise in US Treasury yields rather than credit deterioration.

Nvidia shares fell 2.4 percent to about $218 on Monday, erasing $130 billion in market value, after the Financial Times reported the $500 billion Wall Street financing package. The company reports second-quarter earnings on Aug. 26, which will give investors a clearer view of whether AI chip demand justifies the infrastructure boom. Nvidia posted $81.6 billion in revenue, up 85 percent year-over-year, with net income of $58.3 billion in its first quarter.

This article is for informational purposes only and does not constitute investment advice.