Who eats the difference when the AI buildout misses

@kawishwaqar|October 19, 2026 (1d ago)3 views

Disclosure: I work at Jazz as Principal Evangelist Cloud & AI. This article is written as personal industry analysis. Views are my own.

The question moved

Two weeks ago the number finally went public: Anthropic put its infrastructure obligations in a filing, and the wall stopped being something I inferred from leases and analyst notes. Last week the trade split into two clocks, supply buying 2028 while financing repriced today.

This week I want to sit on the clause, because the clause is where this stops being a story about ambition and becomes a story about accounting.

Anthropic expects to spend at least $518bn over ten years across six partners. About 80% of that is non-cancellable, or payable whether or not the capacity gets used. Google, $111.1bn. Amazon, $110bn. Microsoft, $31.4bn over seven to ten years. Then $161.2bn of Broadcom-related equipment leases, largely non-cancelable. And xAI, up to $84.5bn for Nvidia capacity through 2029, with most of that still cancelable on 90 days notice.

The sentence I keep rereading: if actual spend falls short, Anthropic pays Google the difference. The Amazon terms run the same way. Someone has to eat the gap, and the filing names who.

For context, 2025 revenue was about $4.6bn against an $8.06bn operating loss, with break-even targeted for 2028 and a listing that could value the company above $2tn. That is what spending ahead of your market looks like on paper. But the spend is fixed and the revenue is not.

The earning side was gross

Then OpenAI gave us the other half, and it is the half people skimmed.

On October 8, after the FT reported it and CNBC confirmed it, OpenAI told investors annualized revenue was around $50bn at the end of September. The number in circulation had been $68bn. That is a 26% markdown in a single day, and the explanation was accounting: gross versus net, and pass-through to partners.

A gross figure counts money that flows through you and out to someone else. A net figure counts what you keep. The market had been valuing OpenAI on a gross number while Anthropic disclosed a net, contractual obligation.

The repricing was immediate. Oracle fell about 5%, CoreWeave about 8%, Nvidia about 3%, and the Nasdaq 100 dropped roughly 300 points in an afternoon.

Oracle took the hit for a structural reason. It carries a cloud commitment from OpenAI of roughly $300bn over about five years, which the WSJ puts near 45% of Oracle's remaining performance obligations, with the contracted capacity consumed from 2027 onward. The biggest RPO in the AI trade leans on one customer whose headline growth figure just got revised down a quarter.

I am not saying anyone misses a payment. I am saying the credit standing behind that promise got re-rated in an afternoon, and the promise did not change.

How the debt is actually built

CoreWeave supplied the mechanical detail, and it is the part I would hand to anyone who thinks this is only a valuation argument.

A clause in its loan agreement ties some of the debt to specific customer contracts. A serious breach of a contract backing a $2.6bn loan can trigger a Cash Trap Event if the customer wins the right to walk away, which opens the door to early repayment. That is a contract-linked trigger, not a covenant on earnings. The lender is underwriting the customer relationship, not the company.

The balance sheet behind it: about $35.1bn of debt at June 30, up from $21.4bn at the end of 2025, interest running near 25% of revenue.

Now the counterweight, because it is real and it complicates the easy version. CoreWeave refinanced the SPV debt behind two Applied Digital facilities, ELN-02 and ELN-03, at A3, investment grade, against a corporate rating of BB, with a springing guaranty and a $50m letter of credit behind it. Credit is not closing. It is moving, into guaranty-backed structures while the corporate credit stays high yield. That is the non-recourse shift you know from commercial real estate, applied to GPUs.

The chorus, and the trigger they name

Ray Dalio, in Singapore on October 7, called AI a classic bubble and said we are close to the point where rising rates set off the unwind. His sharper point was composition: a large share of AI spending is now debt rather than cash flow.

Apollo's Torsten Slok put the same test in cash terms. Big Tech creditworthiness holds only if operating cash flow roughly triples, from about $600bn to $2tn. Wellington's Loren Moran, in Bloomberg on October 11, expects the AI giants to become the investment-grade market's largest concentration risk within about two years, passing the banks.

Three people with three different jobs landed on rates as the trigger. None of them are forecasting a date.

The scale is not in dispute anymore. Global AI-linked debt issuance is heading toward roughly $570bn this year, more than double last year, and AI infrastructure spending should clear $795bn this year and $1tn in 2027 on Reuters numbers.

Certain, signal, unresolved

Certain, because it is in a filing: the $518bn, the roughly 80% non-cancellable share, the shortfall clause. Certain because two outlets confirmed it: the OpenAI revision to $50bn. Certain because it is disclosure: CoreWeave's debt level and the terms of that refinance.

Signal, not proof: the readback from OpenAI revenue onto Oracle's RPO, one afternoon of repricing, and the bubble calls. Those are interpretive even when the person making them is right more often than most.

Unresolved: whether any of it breaks, and when. I am not calling a top. Anyone who dates a bubble looks foolish for years before looking clever, and the ones who got it right in 2000 were early by three years.

What changed is narrower and more useful. The spend side of this ledger is contractual, disclosed, and net. The earning side was gross and got marked down. You do not need a forecast to notice which of those two a lender would rather hold.

What I am watching next

The New Mexico air-permit decision for Project Jupiter on November 23. Permits are where a schedule holds or slips, and slippage shows up in leasing economics a quarter or two later.

Whether an anchor lease gets renegotiated after the Anthropic disclosure. The xAI terms are the tell: a 90-day cancelable right inside a $518bn commitment is the closest thing here to the sublease weakness you look for in commercial real estate.

And whether the $50bn figure gets read back onto Oracle and Stargate. If the customer number is lower than the market assumed, every RPO built on top of it deserves a second look. That is the mechanical link between an accounting footnote and a credit spread.

A few weeks ago this was a story about generous estimates. It is a story about contracts now, and contracts are harder to wave away.