On October 21, 2025, Meta announced a joint venture with funds managed by Blue Owl Capital to develop the Hyperion data center campus in Louisiana — roughly $27 billion in total development costs, which Bloomberg called the largest private capital deal on record.

Read the announcement closely and three numbers sit together very oddly. Meta owns just 20% of the joint venture; Blue Owl’s funds own 80%. Meta’s lease with the venture has a four-year initial term. Yet Meta simultaneously provided a residual value guarantee covering roughly the first 16 years of operations — meaning that if the assets are worth less than an agreed level when a lease ends, Meta pays the difference, up to a cap. And there’s a kicker: after contributing the land and construction-in-progress assets to the venture, Meta turned around and took a one-time distribution of roughly $3 billion out of it.

This is a company holding roughly $80 billion in cash and marketable securities (per its 2025 Form 10-K) that borrows at spreads of just a few dozen basis points over Treasuries. Why the contortions? The answer is in the accounting. Of the roughly $30 billion raised, about $27 billion is private placement debt sold to PIMCO and other bond investors, and it sits on the joint venture’s books. Meta accounts for the venture under the equity method and does not consolidate it — and the reason given in the 10-K is not the 20% stake, but that Meta does not direct the activities that most significantly affect the entity’s economic performance and is therefore not the “primary beneficiary” under the accounting standard. So the debt never appears on Meta’s balance sheet. (The same filing discloses, in a footnote, a maximum loss exposure to the entity of roughly $46 billion.)

Not a one-off — the industry’s standard playbook

A Nikkei Asia investigation tallied the AI-infrastructure-related off-balance-sheet obligations of Alphabet, Microsoft, Amazon, Meta, and Oracle — data center leases, GPU purchase contracts, and SPV arrangements (special-purpose vehicles: legally separate companies set up for a single project) — at roughly $1.65 trillion. That exceeds the $1.35 trillion of debt these five companies report on their balance sheets, and it has swelled roughly eightfold in four years. Meta alone accounts for about $420 billion, nearly triple its on-book debt.

The rating agencies moved before the press did. A Moody’s report this February (covered by Fortune) cut the number finer: as of end-2025, the five hyperscalers had signed $662 billion of data center lease commitments that had not yet commenced — more than two-thirds of their total undiscounted lease commitments of $969 billion. Not a dollar of it is on a balance sheet.

Keeping debt in the footnotes requires no fraud — just three current rules working in combination.

Door one: a lease that hasn’t commenced isn’t booked. ASC 842, the current US lease standard issued in 2016, did drag operating leases onto the balance sheet — but recognition starts at lease commencement. While the data center is still being excavated, the signed future rents only need to be disclosed in the footnotes. The $662 billion lives here.

Door two: short initial terms plus renewal options. Traditional data center leases ran 10–15 years. Today’s hyperscaler leases carry initial terms as short as a few years, with a string of renewal options attached. The liability that goes on the books covers only the period the lessee is “reasonably certain” to serve — and a hyperscaler can argue: AI hardware has a useful life of 4–6 years, so who knows what we’ll need in year five? The renewal periods stay off the books. Meta’s four-year initial term at Hyperion is exactly this play.

Door three: a guarantee is only a liability once payout is “probable.” The financiers aren’t naive — if you’ll only commit to four years, why should they fund twenty? So the hyperscalers add residual value guarantees. But under current rules, a guarantee is recognized as a liability only when payment is deemed probable. Moody’s offers an example: Meta disclosed a batch of data center leases commencing in 2029 with an initial commitment of about $12.3 billion — alongside a residual value guarantee with a threshold as high as $28 billion, of which not a dollar was recognized, because management judged payout “not probable.”

Put the three doors together and the economic substance approaches “infrastructure you can use for 20 years, plus a backstop” (everything beyond the four-year initial term being mostly renewal options) — while before lease commencement, not even those four years of initial rent show up as a liability in the body of the statements. All of it sits in the footnotes.

Is this Enron? Is this 2008? Neither — and a bit of both

Start with what’s different, because it matters. Enron’s SPEs were fraud: the supposedly independent outside equity was controlled by insiders and bore no real risk, and the guarantees were deliberately concealed (the SEC’s complaint documents the details). Today’s structures — at least the Meta–Hyperion deal dissected here — follow the current rules, with the terms written into securities filings. They’re just written in the footnotes. Standard leverage metrics don’t automatically count footnote commitments; rating agencies do add them back by hand into “adjusted debt,” but that’s homework few people do. The problem isn’t concealment. It’s disclosure that almost nobody discounts.

The genuine parallel to 2008 is one specific lesson: risk transfer is often illusory. Banks moved assets into off-balance-sheet structured investment vehicles (SIVs) while retaining explicit or implicit support commitments, and when markets froze, Citigroup and other sponsoring banks ended up taking their SIVs’ assets back onto their own balance sheets. Today’s counterpart is the residual value guarantee and the operational dependence. If AI demand disappoints when a lease expires, Meta can decline to renew — but if the assets’ fair value has fallen below the agreed declining threshold, Meta pays the difference up to the cap. If demand holds instead, the leases commence, renewals get confirmed, and the liabilities come back onto the balance sheet at scale. Moody’s view is that even this second scenario isn’t benign: as leases commence one after another, adjusted debt will surge, potentially driving a “material” deterioration in credit profiles (Moody’s report).

The key difference is where the risk stops this time. In 2008 the banking system held it itself — leverage plus maturity mismatch, so transmission was fast and violent. This time the catch basin is the private markets: Blue Owl-managed funds hold 80% of the joint venture’s equity, PIMCO and other institutions bought the bonds, and the capital behind private credit funds like these comes mostly from pension funds, insurance companies, and sovereign wealth funds — long-horizon money. Private credit assets have no real-time market prices; valuations are opaque and updated infrequently, so the response to shocks lags. Transmission will be slower — and more opaque. By the time you see the losses, they are a fait accompli.

What to take away

First, when reading a hyperscaler’s financials, look past the P/E ratio to three footnotes: lease commitments not yet commenced, purchase obligations, and guarantee arrangements. On-balance-sheet debt is now the least informative number in these five companies’ leverage story.

Second, there is a nearly free indicator: the gap between the initial lease term and the guarantee term. Meta is willing to guarantee 16 years but will only commit to renting for four. That is Meta putting its own price on the question “will AI demand last past 2030?” — too uncertain to write into the body of the contract, acceptable only inside a guarantee clause. A company’s accounting choices are more honest than its public statements.

Third, if this build-out ultimately falls short, where does the crack show first? This is a judgment call with no ready dataset to cite; my reasoning runs as follows. The first stress probably won’t appear in hyperscaler share prices, but in the marks on private credit portfolios — and in the insurers and pension funds holding that debt. Precisely because those valuations update with a lag, by the time the crack becomes visible, the losses will have been accumulating for a long while. To my eye, the structure works like a cheap put option for the Metas of the world: if demand materializes, just renew; if it doesn’t, the worst case is paying out the capped guarantee and leaving the depreciating server halls to the creditors. The other side of that option — the private credit investors collecting a spread while carrying the tail risk on “AI demand persists” — is who should actually be losing sleep in this story.

References