The AI compute reset wall just went public

@kawishwaqar|October 12, 2026 (16h 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 number I have been waiting for

I have spent months writing about a wall: the gap between what the AI buildout has committed to spend and what it actually earns. Most weeks I was working from leases, analyst notes, and filings stitched together by hand. This week the number turned up in a prospectus. Not a rumour of a number. A document.

Anthropic, in a confidential SEC filing reported by Reuters in late September and circulating through early October, expects to spend at least $518bn over ten years on AI infrastructure across six partners. Roughly 80% of that is non-cancellable, or payable whether or not Anthropic uses the capacity.

Break it down. Google, $111.1bn. Amazon, $110bn. Microsoft, $31.4bn in infrastructure obligations over seven to ten years. On top of that, $161.2bn of Broadcom-related equipment leases, largely non-cancelable. And xAI, up to $84.5bn for Nvidia capacity through 2029, though most of that stays cancelable on 90 days' notice.

The line that stopped me: if actual spend falls short, Anthropic pays Google the difference. Similar terms sit on the Amazon deal.

Whatever you think of Anthropic's models, they just told the SEC that their largest risk is not being wrong about intelligence. It is a contract. Take-or-pay is the operating assumption now, not a tail risk inside somebody's spreadsheet.

For scale, 2025 revenue was about $4.6bn, up 12 times, against an $8.06bn operating loss and a net loss of roughly $42bn, about $34bn of which was non-cash. They are targeting break-even in 2028, with a listing that could value the company above $2tn.

The other half of the ledger

If Anthropic's filing gave us the spending side, OpenAI handed over a piece of the earning side.

On October 8, after the FT reported it and CNBC confirmed it, OpenAI told investors its annualized revenue was around $50bn at the end of September. The widely cited figure had been $68bn. The gap is accounting, gross versus net and pass-through to partners, not a collapse in demand.

The market treated it as a collapse anyway. Oracle fell about 5%, CoreWeave about 8%, Nvidia about 3%, and the Nasdaq 100 dropped roughly 300 points in an afternoon.

The repricing landed hardest on Oracle for a structural reason. Oracle carries a roughly $300bn cloud commitment from OpenAI over about five years, and according to the WSJ that represents close to 45% of Oracle's remaining performance obligations, with the contracted capacity being consumed from 2027 onward. The largest single RPO in the AI trade rests on one customer whose headline growth number just got revised down 26%.

That does not prove anyone misses a payment. It re-rates the credit standing behind the promise.

CoreWeave shows how the debt is actually built

CoreWeave supplied the structural detail. Benzinga reported on October 7 that a clause in the company's loan agreement ties some of its 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 gains the right to walk away, which opens the door to early repayment. The stock fell about 8% on October 8.

The balance sheet behind it: about $35.1bn of debt at June 30, up from $21.4bn at the end of 2025, with interest running at around 25% of revenue. Redburn initiated a Sell at $54 in September, and CoreWeave's credit default swaps sat near 855 basis points before August.

The counterweight is real, so I will state it plainly. 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. That deal came with a springing guaranty and a $50m letter of credit. Somebody is still lending this company money, and lending it at good rates.

Read those two facts together and the conclusion is not that credit is closing. The funding is migrating into guaranty-backed structures, project by project, while the corporate credit stays high yield. The financing is hardening even as the headline rating stays weak.

The chorus got louder, and rates are the stated trigger

Ray Dalio, speaking 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 point about funding composition is the one I would underline: a large share of AI spending is now debt rather than cash flow.

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

The scale is no longer debatable. Global AI-linked debt issuance is heading toward roughly $570bn this year, more than double last year. Morgan Stanley, cited in Bank of England minutes, counted $450bn through early September. Amazon, Microsoft, Alphabet, Meta and Oracle together issued about $200bn of investment-grade debt in the first half of 2026. AI infrastructure spending should clear $795bn this year and $1tn in 2027 on Reuters' numbers.

Certain, probable, unresolved

Certain, because Anthropic put it in a filing: the commitments and the non-cancellable share. The OpenAI revenue revision, confirmed twice. CoreWeave's debt level and the terms of that refinance.

Signal, not proof: the readback from OpenAI's revenue to Oracle's RPO, one afternoon of repricing, and analyst warnings, including Dalio's. Those are interpretive even when the person saying them is right more often than most.

Unresolved: whether any of it breaks. I am not calling a top. Nobody can date a bubble, and everyone who has tried looked foolish for years before looking clever. What changed is narrower and more useful than a prediction. The number I had been inferring is now written down where I can read it.

What I am watching next

Three things, in the order they would tell me the most.

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

Whether any anchor lease gets renegotiated following the Anthropic disclosure. The xAI terms are the tell, because a 90-day cancelable right is the closest thing in this story to the sublease weakness you would look for in commercial real estate. If a lab carrying $518bn of stated obligations is negotiating for exit ramps, that is the signal.

And whether the $50bn OpenAI figure gets read back onto its Oracle and Stargate commitments. If the customer's number is lower than the market assumed, the RPOs built on top of it deserve a second look.

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