Deep dive · June Okafor (The Contrarian) · 2026-07-25 · Why “the hyperscalers can pay for it from cash flow” is a claim about the wrong balance sheet.

Here is the sentence that ends most arguments about the AI bubble. The companies paying for the buildout are the four or five richest firms on earth, and they are funding it out of operating cash flow, not debt. So even if the returns disappoint, nobody blows up. The buildout is expensive, not fragile.

The number people reach for is real. Goldman Sachs puts hyperscaler capital spending at roughly 100% of operating cash flow — heavy, but self-financed. Two weeks ago, this publication ran a version of the same reassurance: the end-riders of the AI capex circle are the solvent giants, ~$451B of 2024 operating cash flow between them, so the AI loop passes the test the late-1990s telecom loop failed. I still think that piece was right about telecom. I now think the “cash-funded, therefore safe” reading is checking the wrong document.

Let me steelman it first, because it is not a dumb argument. Microsoft, Amazon, Alphabet, and Meta really do generate enormous cash. Token demand is real and growing, not a fiction. And a company that puts a data center into a joint venture and keeps a minority stake really has handed part of the risk to whoever owns the majority. These are not accounting tricks in the fraud sense. Every number I am about to use comes from a filing, a rating agency, or a bond prospectus. Nobody is hiding anything illegally.

They are just booking it somewhere the solvency test doesn’t look.

The debt migrated off the statement

Start with the deal everyone in credit is studying. In October 2025 Meta financed its Hyperion campus in Richland Parish, Louisiana through a special-purpose vehicle: Blue Owl owns 80%, Meta 20%. The vehicle issued about $27 billion of senior secured notes due 2049, fully amortizing, rated A+ by S&P, with PIMCO as anchor lender and BlackRock buying more than $3 billion of the paper. Total package around $30 billion. The largest private-credit data-center deal ever done.

Now read Meta’s balance sheet. That $27 billion is not on it. The notes sit on the vehicle’s books. Meta consolidates a 20% stake and signs a long-term lease to operate the site. What looks like $27 billion of debt on anyone else’s statement shows up on Meta’s as a lease — an operating expense, spread over years, with the principal living in a footnote. Meta reportedly did the same for a Texas site with a second ~$13 billion vehicle (single-sourced, ZeroHedge). Across Big Tech, roughly $120 billion of data-center spending has moved off balance sheets this way in about eighteen months.

That is one mechanism. Here is the bigger one, and it needs no vehicle at all.

Under the lease-accounting standard (ASC 842), a company records a lease liability when the lease commences — when the landlord hands over the keys. Sign a fifteen-year lease today for a data center that gets delivered in 2027, and until 2027 that obligation is disclosed in the notes but carried nowhere on the balance sheet. Moody’s added those commitments up in February. The five largest hyperscalers — Amazon, Meta, Alphabet, Microsoft, Oracle — were sitting on $662 billion of future data-center lease obligations that had not yet hit their balance sheets, out of $969 billion of total undiscounted lease commitments (Moody’s Ratings, analysts David Gonzales and Alastair Drake).

Read that $662 billion against the thing it is supposed to reassure you about. Moody’s puts it at roughly 113% of the combined adjusted debt of all five companies. The obligations you cannot see on the statement are larger than every dollar of debt you can see. So when Goldman tells you capex is about 100% of operating cash flow, that ratio is computed over the visible half of the company. The other half is a lease footnote that becomes real on a delivery schedule nobody is stress-testing against a demand shortfall.

And the third mechanism is the one we already covered: rent the compute from someone who borrowed to build it. The neocloud carries the debt; the hyperscaler shows an opex line. Oracle is the cleanest cautionary case, because it chose to keep the buildout on its own books and you can watch what that costs. Fiscal-2026 free cash flow of negative $23.7 billion, total debt around $167 billion, and an S&P downgrade to BBB− — one notch above junk — with OpenAI’s concentration in its backlog flagged as the reason. That is what AI capex looks like when it is on the balance sheet. The SPV structure exists precisely so the others never have to show you that picture.

The counter-thesis: it moved, it didn’t vanish

So my claim is not “these companies are secretly broke.” It is narrower and, I think, harder to argue with: the leverage funding the AI buildout is being engineered off the exact statement the “they can afford it” argument depends on, faster than the cash flow behind it is growing — and the risk it carries is being moved to the places the balance-sheet test cannot see.

Where does it go? The Bank for International Settlements named the channels in March. This “shadow borrowing” — a joint venture or SPE, minority-held, funded by private credit, backed by hyperscaler guarantees and offtake commitments — creates new transmission paths: refinancing pressure at the vehicle level, private-credit appetite that turns procyclical, and guarantees that activate under stress. It strengthens the links between the hyperscalers and non-bank lenders — private-credit funds and insurers — that hold the paper. The debt didn’t leave the system. It left the income statement and entered a channel that reports quarterly to nobody you’re watching.

The right denominator, then, is not reported debt over cash flow. It is reported debt plus off-balance-sheet leases plus unconsolidated JV obligations plus guarantees, over operating cash flow — and stress-tested against the AI revenue actually showing up on schedule. Run that number and “capex ≈ 100% of OCF” stops being the whole story. The $662 billion goes in the numerator.

Now the honest part, because the contrarian move here is not “panic.” Two things cut against me. First, this is disclosed. Moody’s could count the $662 billion because it is in the footnotes; the Hyperion notes are rated and public. This is legible leverage, not concealed leverage — which is exactly why the crisis, if it comes, will be slow and arguable rather than a single Enron morning. Second, the risk transfer to Blue Owl and PIMCO and the insurers is partly genuine. A 20% stake with a bounded lease is not the same as 100% of the debt. Some of the danger really did leave Meta.

So the market is doing the sensible thing: repricing at the edges while the headline balance sheets still look pristine. Order books that covered hyperscaler bonds nearly five times over in February covered less than twice by July. Morgan Stanley sees ~$570 billion of AI-related debt issuance in 2026, four times the prior year, against an $800 billion private-credit data-center pipeline through 2028. Meta’s credit-default swaps hit a record while its equity sat near highs. Public spreads haven’t blown out — J.P. Morgan notes they’re still near cycle lows — but the coverage ratio is where nervousness shows up first.

What would prove me wrong

Here is the test that settles this, and it is a real one. In the next demand air-pocket — a quarter where AI revenue misses badly enough to strand capacity — watch where the loss lands.

If it lands on the private-credit funds and the bondholders who bought the SPV paper, and the hyperscaler’s guarantees never activate, and the lessee walks away down the four-year exit ramp the Hyperion structure was built with — then the off-balance-sheet treatment reflected honest risk transfer, and I was double-counting. The structure worked as designed. The rich company really did offload the danger.

If instead the loss boomerangs — a guarantee gets pulled, a lease gets renegotiated rather than abandoned, the parent quietly backstops the vehicle to protect its own credit and its own campus — then the off-balance-sheet line was cosmetic. The risk was theirs the whole time; it was just booked where the solvency test couldn’t price it.

I lean toward the second. The tell to watch is not a Nvidia headline, and not even a neocloud missing a refinancing — it’s a hyperscaler whose reported balance sheet still looks clean while its CDS keeps widening (Meta), or a company that kept the debt visible getting downgraded on concentration (Oracle). The risk is already surfacing exactly where the reassurance told you not to look.

So the so-what for anyone sizing this trade, or just trying to read the quarter honestly: stop checking the headline balance sheet and read three lines in the notes. Purchase and lease commitments not yet commenced — that’s where the $662 billion lives. Unconsolidated variable-interest entities and equity-method JVs — that’s where Hyperion lives. Guarantees and residual-value backstops — that’s where the risk transfer turns out to be real or fake. The AI buildout may well be affordable. But “they’re paying cash” is an answer about the part of the company that still fits on one page, and the buildout stopped fitting on one page around $662 billion ago.