Analysis
The headline AI capex numbers are the small ones. A Wall Street Journal analysis of footnotes in the most recent securities filings of nine large technology companies found roughly $3 trillion of off-balance-sheet commitments mostly tied to AI, Axios reported. Reported capital expenditure for the same group over the trailing year was about $600 billion. The off-balance-sheet number is roughly five times larger, growing faster, and about triple the combined total of outstanding leases and long-term borrowings.
How $3 trillion disappears
Most of it is mechanical rather than sinister. About $1.2 trillion consists of leases that have not yet commenced, and under current accounting a lease stays off the balance sheet until it begins. The rest is purchase obligations, capacity reservations, power contracts and guarantees.
“## How $3 trillion disappears Most of it is mechanical rather than sinister.”
The Hyperion campus in Louisiana is the cleanest illustration. Meta owns 20% of it. Blue Owl, the private-credit manager, owns 80%. The $27 billion of construction debt sits on neither Meta's public balance sheet nor anywhere else investors routinely look -- yet Meta has guaranteed to be the sole tenant. Economically, Meta bears the risk. Accounting-wise, Meta reports a fraction of it.
Why private credit is in the middle of this
The structure is not an accident. Blue Owl, Apollo, Ares and Blackstone have raised enormous private-credit vehicles specifically to fund infrastructure that investment-grade corporates would rather not carry. The corporate gets capacity without balance-sheet consolidation; the credit fund gets a long-dated, effectively investment-grade-guaranteed asset at a spread. Both sides are rational. The systemic question is that the risk has moved from public disclosure into private funds whose marks are not observable in real time.
The comparison people will reach for
The 2008 analogy is available and mostly wrong. These are operating commitments to build real assets with a named, creditworthy tenant, not synthetic exposures layered on a mortgage pool. A better comparison is telecom in 1999-2001: enormous long-dated commitments to build capacity on a demand forecast, financed off balance sheet, where the assets were genuinely real and the demand curve arrived late. Fiber got built. WorldCom and Global Crossing still failed.
What to actually do with this
If you are an LP or a public-markets investor, the practical implication is that AI capex is more leveraged than the headline capex figures suggest, and that the leverage sits with counterparties you cannot see. If you are a founder in data centers, power or cooling, the implication runs the other way: the money committed to this buildout is larger and more contractually locked than the quarterly capex guides imply, and the customers cannot easily walk.
The specific disclosure to push for is when those $1.2 trillion of unstarted leases commence. That is the quarter the accounting catches up with the economics.
Who else is exposed
The nine companies in the WSJ analysis are the obvious names, but the second-order exposure runs through the credit funds and insurance balance sheets that hold the debt. Private credit has grown to roughly $2 trillion in assets over the past decade, and data-center lending has become one of its largest new origination categories. Insurance affiliates of the large alternative managers hold meaningful portions of that paper. None of it is marked daily. In a scenario where AI capex guidance is cut, the equity reprices in an afternoon and the credit reprices whenever the next valuation committee meets -- which is exactly the lag that makes private-credit exposure hard to hedge.
