The $3 Trillion Blind Spot: WSJ Exposes Big Tech's Hidden AI Debt Machine

Here is a number that puts every AI capex headline you have read this year in the shade: roughly $3 trillion. That, according to a Wall Street Journal analysis published August 18, is the combined total of off-balance-sheet commitments — buried in the footnotes of securities filings — carried by just nine US technology companies as they race to build artificial-intelligence infrastructure. It is not on their balance sheets. It does not appear in the capex figures that analysts quote every quarter. And it is roughly five times the approximately $600 billion in capital expenditure those same companies reported over the trailing twelve months.

The nine companies the WSJ examined are Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, and AMD. Together they form the backbone of the AI buildout — the buyers of chips, the builders of data centers, and, increasingly, the financiers of each other’s growth. The finding lands at a moment when AI financing is under intensifying scrutiny: Broadcom is seeking more than $60 billion in fresh debt for AI chip financing, Alibaba’s AI-driven capex surge has cratered its quarterly profit, and US Senator Elizabeth Warren has pressed the FSOC to probe the systemic risks of an “AI debt bubble.” The WSJ analysis gives that debate its most concrete number yet.

What is actually in the footnotes

The hidden liabilities fall into two broad buckets, and both exploit the same accounting principle: under current rules, a commitment does not have to be recorded on the balance sheet until goods are delivered or an asset is put into use.

The larger bucket — about $1.9 trillion — consists of purchase commitments. These are non-cancellable long-term contracts to buy equipment: Nvidia’s hard-to-source AI accelerators, memory chips, networking gear, and the electricity that data centers will consume for decades. Alphabet’s power-purchase agreements, remarkably, stretch as far out as 2054. As of June 30, Alphabet alone listed $811 billion in purchase and contractual commitments, according to the WSJ’s reading of its filings.

The second bucket — about $1.2 trillion — is lease obligations that have not yet commenced. When a company signs a 15-year lease on a data center that opens in 2029, the entire liability sits in the footnotes until the lease begins. The total value of these future lease commitments across the nine companies has roughly quadrupled in a single year, a growth rate that far outpaces the already-staggering expansion of reported capital spending.

Meta’s Hyperion: a case study in vanishing debt

The WSJ’s most instructive example is Meta’s Hyperion data-center project in Louisiana. Meta is the developer and anchor tenant of the facility, and the project was financed with $27 billion of debt. Yet none of that debt — and not the data center itself — appears on Meta’s balance sheet.

The structure makes it possible. A majority stake in the complex is held by funds managed by Blue Owl Capital, and a third-party holding company raised the $27 billion through a bond offering. Meta is merely a minority partner and future tenant. Lease payments begin in 2029 and can run for up to 20 years, and Meta guarantees payments to bondholders — but because the lease has not commenced, and because Meta itself classifies the probability of paying under the guarantees as “unlikely,” the $12.3 billion future lease obligation is entirely absent from its current financial statements.

This is not fraud, and the WSJ does not allege it is. It is financial engineering operating exactly as designed: every obligation disclosed, every rule followed, and yet the aggregate picture presented to investors is dramatically lighter than the economic reality. Meta discloses $347 billion in uncommenced leases across its filings; Nvidia carries $182 billion in forward purchase commitments.

Why the gap matters now

The scale of the discrepancy — off-balance-sheet commitments now reportedly exceed the nine companies’ official debt from loans and operating leases by a factor of three — matters for two reasons.

First, the commitments are overwhelmingly non-cancellable. These are not options that companies can walk away from if AI demand disappoints; they are binding promises to pay for chips, power, and facilities regardless of whether the revenue materializes. If AI monetization falls short of the curve these companies are underwriting, the liabilities do not evaporate — they become stranded costs that must be serviced out of cash flows from other businesses.

Second, the warning lights are already flickering. In their most recent quarterly reports, Alphabet and Amazon both posted negative free cash flow — capital spending is now outrunning the cash generated by their core operations. That is sustainable for a while for businesses of this resilience, but it converts the off-balance-sheet problem from an accounting curiosity into a live solvency question if the gap persists. Michael Burry, the investor famous for the big short, has been publicly warning that hyperscalers are depreciating chips over five to six years when their economic life is closer to two or three — a related trick of optimism that flatters earnings today at the expense of tomorrow.

Morgan Stanley analysts put the concern plainly in a note cited by the WSJ: as these structures become more common, larger, and more complex, “it is becoming increasingly difficult for investors to assess companies’ aggregate potential debt burden.”

The circular-financing shadow

The WSJ analysis lands on top of an increasingly audible debate about circularity in AI finance. Nvidia invests in OpenAI, which buys Nvidia chips; Broadcom arranges tens of billions in debt so customers can buy Broadcom silicon; Alphabet rents compute to Anthropic while Anthropic commits revenue back to Google Cloud. Each individual transaction has a commercial logic. The aggregate effect is that a meaningful share of the AI industry’s apparent demand is funded by its own suppliers.

Against that backdrop, a $3 trillion stack of footnote-level commitments functions as the debt tail of the same dog. The nine companies in the WSJ sample are not just spending their own cash — they are jointly binding themselves to future payments larger than the GDP of most G20 countries, on the assumption that AI adoption will keep compounding. Senator Warren’s letter to the FSOC in January warned that opaque AI financing could fuel the next financial crisis; the WSJ’s number is the kind of evidence that moves such warnings from the opinion pages toward the regulatory agenda.

What to watch

The report is already reshaping how analysts read filings. Expect three follow-ons. First, rating agencies and sell-side analysts will begin publishing their own sums of disclosed commitments, turning a footnote-scavenging exercise into a standard metric — the way “adjusted EBITDA” or “net debt” became conventions. Second, pressure will grow on accounting standard-setters to tighten when lease and purchase commitments must be recognized, particularly for structures like Hyperion where a single anchor tenant’s guarantees sit behind multiple layers of special-purpose entities. Third, if any one of the nine companies misses revenue expectations badly enough, the market will reprice all of them at once — because the WSJ has now made it impossible to claim the exposure was invisible.

There is a bullish reading, to be fair: commitments this large also signal contracts this locked-in. Suppliers like Nvidia and Broadcom can partially bank on $1.9 trillion of purchase obligations, and energy developers can finance against 2054-dated offtakes. The buildout is real, the chips are real, and the electricity contracts will keep turbines spinning for decades. The $3 trillion is not a hoax — it is a bet, placed at unprecedented scale, with the losses structured to arrive later than the wins.

But that is precisely the asymmetry investors are being asked to underwrite. The wins from AI infrastructure show up in today’s revenue and today’s stock prices. The cost of the bet sits in the footnotes, quietly, until the day it doesn’t.

Figures in this article are drawn from the WSJ’s analysis of the companies’ most recent securities filings as of publication. Lease and purchase commitments change quarterly; totals are approximations the WSJ derived from filing disclosures.