Big Tech hides $3 trillion in AI commitments off balance sheets

Alphabet, Microsoft, Amazon, Meta Platforms and Oracle have signed roughly $3 trillion in data-center leases and chip purchases that do not appear on their balance sheets, a Wall Street Journal analysis shows.

The Bank for International Settlements, the coordinating body for central banks, warned in a recent report that the financing echoes nineteenth-century railroad speculation and the dot-com bubble, with today's sums far exceeding those episodes.

The off-balance-sheet total of $1.65 trillion for the five hyperscalers exceeds the $1.35 trillion they report as debt, according to a Nikkei Asia analysis, and has grown roughly eightfold in four years. Meta carries about $420 billion in hidden obligations versus $140 billion in reported debt, while Oracle's have climbed about thirtyfold since 2022 to $273.3 billion, tied to the Stargate project with OpenAI. Alphabet's purchase commitments jumped to $811 billion at the end of the second quarter from $332.4 billion in the first, its quarterly filing shows.

The gap matters because lease and purchase contracts are recorded only in footnotes until assets begin operating, so a slowdown in AI revenue could convert billions in invisible obligations into balance-sheet burdens within a few quarters. The Financial Stability Board found AI projects accounted for more than a third of all private lending in 2025, up from 17 percent in the prior five years.

Special Purpose Vehicles Keep Debt Off the Books

The mechanism relies on special purpose vehicles, or SPVs. A technology company and a private lender establish a separate legal entity that owns the land, buildings, power supply and sometimes the chips of a data center; the SPV borrows the money while the tech company signs a long-term lease securing capacity. Institutional investors including PIMCO, BlackRock, Apollo and Blue Owl Capital, plus banks such as JPMorgan, provide debt and equity without the loan appearing on the parent's balance sheet.

Meta's Hyperion complex in Louisiana shows how the structure works. Meta and lender Blue Owl Capital created an SPV called Beignet Investor that raised about $27.3 billion in secured debt plus $2.5 billion in equity; Blue Owl controls roughly 80 percent while Meta holds 20 percent as sole tenant and operator. S&P assigned the bonds an A+ rating despite a yield near 6.58 percent, a level more typical of speculative debt. Because the obligations sit off Meta's books, the company raised an additional $30 billion on the regular bond market shortly after.

The accounting is legal: long-term lease and purchase obligations need only be disclosed in footnotes until a facility begins operating or chips are delivered, at which point they appear in full on the balance sheet. That time lag lets companies keep debt ratios low and credit ratings intact during the most capital-intensive construction phase.

Regulators Warn of a Systemic Credit Event

The investment wave behind these structures is enormous. The five largest hyperscalers are expected to spend $700 billion to $900 billion on capital in 2026, up about 36 percent from a year earlier, with Amazon alone planning roughly $200 billion and Alphabet nearly doubling its forecast to as much as $205 billion. Allianz Research puts investment intensity at 34 percent of revenue, more than double the 15 percent peak of the dot-com boom, while Sequoia estimates the gap between investment and revenue growth at about $600 billion annually.

The Financial Stability Board, which coordinates regulators from 24 countries, identified two risk channels that can reinforce each other: power-supply bottlenecks that trigger stricter loan conditions, and overcapacity if data-center expansion outpaces demand for AI computing. Moody's, in a more conservative estimate, puts off-balance-sheet agreements at about $1.2 trillion, with more than $820 billion tied to data centers under construction, while still calling hyperscaler balance sheets among the strongest in the corporate world.

A Bank of America survey of global fund managers found 34 percent now consider hyperscaler capital spending the most likely source of a future systemic credit event, double the prior month. Evercore ISI analysts flag that combined free cash flow among hyperscalers has fallen to levels last seen in 2022, and PIMCO calculates capital expenditures will consume about 94 percent of operating cash flows over the next two years, up from 40 percent in 2023.

For investors, the risk is that the market has priced in an uninterrupted continuation of this spending pace. Steve Eisman, the investor known for betting against the 2008 housing bubble, warned that a substantial reduction in investment by any major cloud provider would send the market into a downward spiral. If AI monetization falls short of the very short depreciation periods of three to five years, billions in currently invisible liabilities could convert to visible balance-sheet burdens within a few quarters, triggering the chain reaction regulators have flagged.

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