藏在租赁脚注里的3万亿美元AI债务

The capital-expenditure figure investors track for the AI buildout is roughly $600 billion a year across America's biggest technology companies. It is the visible sliver of the spending. In an analysis of securities filings, the Wall Street Journal counted about $3 trillion in AI infrastructure commitments that nine major companies have made and that do not appear on their balance sheets - roughly triple what the same group owes under its existing leases and long-term borrowings. That gap, not the capex headline, is where the risk lives.

Most of the $3 trillion never touches a balance sheet because of how two kinds of promise are treated in accounting. The larger of the two buckets, about $1.9 trillion, is purchase commitments: long-term contracts to buy chips and other infrastructure from suppliers, which are not recorded as a liability until the goods actually arrive. The rest, roughly $1.2 trillion, is leases that have been signed but have not yet begun. A lease becomes an asset and a liability only once it starts, so a data-center lease that begins in 2029 books nothing today. The Journal counted these not-yet-started lease obligations at about four times their level a year earlier.

Meta's Louisiana project shows how a multi-billion-dollar data center can leave almost nothing on its owner's balance sheet. In October 2025, MetaMETA-- announced a joint venture with funds managed by private-credit giant Blue OwlOWL-- to build its Hyperion campus in Richland Parish. Blue Owl owns 80 percent of the venture; Meta holds 20 percent and contributes land and construction-in-progress assets. Total development cost is about $27 billion, funded pro-rata, and a piece of Blue Owl's money comes from bonds sold to PIMCO and other investors. Then the structure does its work: Meta does not buy the campus but signs an operating lease to use all of it once built, with a four-year initial term and options to extend. Because it is an operating lease, Meta recognizes no asset and no leasing liability the way a purchase or finance lease would.

Quick Backtesting Tool

Symbol

Strategy

Backtest Range

The arrangement is genuinely two-sided, which is what makes it hard to price. To keep the lease cheap, Meta gave the venture a capped residual-value guarantee for the first sixteen years - a backstop if it chooses not to extend. And according to investor documents and a Blue Owl spokesperson, Meta can leave the campus without paying the difference on what it owes. So the structure cuts both ways: it lets Meta add AI capacity without reported leverage, and it lets Meta shed the commitment if demand sours. But the debt is real. It sits with PIMCO's bondholders and Blue Owl's funds, off Meta's books. The party holding the downside of the growth story is private credit, not the shareholders reading the income statement.

The harder question is whether the demand underneath all this spending is as solid as the money suggests. The sharpest version comes from investor Michael Burry, who in August 2026 argued that about $879 billion of hyperscaler purchase commitments circle through Nvidia - the same capital appearing as revenue at one company and then as a customer order funded by another. His examples trace the loop: OracleORCL-- committing roughly $300 billion of purchases tied to OpenAI, OpenAI committing about $90 billion to AMDAMD-- while taking a direct stake in NvidiaNVDA--, and billions more in direct equity stakes crisscrossing the group. Treat these as a claim rather than a finding - Nvidia's chief executive calls the circular-financing label "ridiculous," and Burry is a skeptical investor with a documented short thesis. But the underlying structure is visible in filings: the companies financing AI's buildout are routinely each other's largest customers, so a single demand disruption echoes through several income statements at once.

The cost to shareholders is the gapGAP-- between apparent and actual leverage. On reported numbers, the five biggest hyperscalers look lightly borrowed; FactSetFDS-- notes that incremental debt has climbed from about 9 percent of capex two years ago to near a third recently, yet gross leverage still sits around 1x EBITDA for four of the five. That comfort depends on the off-balance-sheet commitments staying hidden. A narrower tally that counts only leases - FactSet puts unrecognized lease-related commitments across the five hyperscalers at roughly $820 billion - is a separate figure with a different scope and definition from the Journal's broader $3 trillion number, rather than a component of it.

Oracle chose the other path, financing more on its own books, and it shows what visibility costs. Ainvest data as of early October 2026 show about $236 billion in total debt against about $67 billion of book equity - gross debt near three and a half times book value - and negative free cash flow of roughly $29 billion over the trailing twelve months. FactSet credits Oracle's downgrade to BBB- by S&P to surging capex and negative free cash flow tied to its contractual obligations to OpenAI. The off-balance-sheet route lets peers report a cleaner leverage ratio; it does not remove the obligation, it relocates it.

The three outcomes price differently. If AI demand arrives, the offtakes and leases become real cash flow, private-credit money was cheap operating leverage, and reported results converge with economic reality. If demand merely persists, spending stays elevated, free cash flow hovers near zero, commitments roll forward, and the market's multiple stays a standing bet that future cash will justify it. If demand disappoints, largely non-cancelable purchase commitments and signed leases convert into write-downs, and true leverage rises precisely as growth decelerates - the burden that was easiest to hide in a footnote is the hardest to unwind.

"Off-balance-sheet" is not a fraud accusation. It is a financing choice with a specific consequence: it lets reported returns and reported leverage tell half the AI story. The next settling event will be an ordinary one - a quarter where a hyperscaler stops renewing a twenty-year lease option or cuts a chip purchase commitment. When the footnote stops growing, investors will learn which companies were renting capacity and which were buying a bet.