The reset wall slipped, and the exit clauses came with it

@kawishwaqar|October 26, 2026 (1w ago)2 views

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

Two weeks ago the spending ledger went public. This week it moved.

The number I keep working from is the one Anthropic put in a filing: $518bn of infrastructure obligations over ten years across six partners, roughly 80% non-cancellable, with $111.1bn to Google, $110bn to Amazon, $31.4bn to Microsoft and $161.2bn of Broadcom equipment leases behind it. That answered who owes what. What changed in the last fortnight is when they can walk.

I want to be careful with the word bubble, because it has become a way of not saying anything. The demand side is not cracking. Published compute prices went up into October: AWS p5 from $5.191 to $5.970 per H100-hour, RunPod $3.29 to $3.99, Lambda $2.99 to $3.99, Nebius $3.85 to $4.50 from October 1. The B300 median across ten non-hyperscaler providers moved from $8.50 to $9.51 in the same window, on about 96% utilisation. Those are tracker numbers, not audited ones, but they all point one way. Nvidia's October 21 supplier buildout says the same from the other end: the supply side is not retreating either.

What is cracking is the delivery schedule. And the people holding the paper.

47 projects, and the exit ramp nobody priced

FT counted 47 large US data centre projects delayed, withdrawn or blocked this year. Only about a third of the planned pipeline is judged credible, and more than 1,000 communities have sought moratoria. CoreWeave shares fell almost 25% on the week. A Virginia bond funding a CoreWeave-leased site is down about 5% since it priced three weeks ago.

Then the structural detail. Anthropic's $45bn deal with Nscale is cancellable if Nscale misses financing or delivery deadlines at the West Virginia site. Nscale had roughly 25,000 advanced chips online in August against a goal near 500,000. At a Texas site, Cipher absorbed the first ~$360mn of overruns to buy time, and that took ~$750mn of fresh equity.

So the backlog everybody quoted as committed revenue is conditional revenue. If the capacity does not arrive on time, the tenant walks and the special purpose vehicle eats it. The risk did not disappear. It moved down the capital structure, from the lab to the project lender. For a company carrying a $518bn obligation that is the right trade to make. For a bondholder in a single-asset data centre SPV it is the whole position.

That has a second effect. Slippage now defers the reset rather than cancelling it, and you cannot get out of a wall by moving the date on the calendar.

The debt is already being repriced

Oracle is one notch from junk. Five-year CDS went from 141bps on January 2 to 295bps on October 19, touched 300, then 308bps on October 21. S&P holds it at BBB-, downgraded in July. The balance sheet carries $118bn of long-term debt plus $39bn of capital and lease obligations, with consensus free cash flow of -$23.7bn for FY26 and -$36bn for FY27. Deutsche Bank's line: a cut "would risk pressuring IG-only investors into forced sales." Oracle's investor day runs through October 28.

CoreWeave gives the cleaner number. Five-year CDS hit 1,065.84bps on October 21, against $35.6bn of debt at June 30 and $11.2bn a year earlier.

And S&P said out loud what the filings only imply. Mark Habib, at a London roundtable: "It is safe to assume that there is more debt out there that we are not seeing," pointing at structures where hyperscalers guarantee builds or smaller AI firms borrow against them. No ratings moved on it. A rating action normally needs a filing behind it, and Habib just said the filings are incomplete.

Anthropic itself is being marked nearer $1.5tn than $2tn, after Q3 revenue of $16.8bn grew 45% quarter on quarter against more than doubling in Q2. It is adjusted profitable, and its listing is still the only one moving, marketing the week of November 9.

Meta showed us the 2027 bill, itemised

Meta spent $34.9bn of capex in Q3 and guided FY26 to $135bn to $145bn, raising the low end by $5bn. It gave no 2027 number at all, with the CFO saying the company is "quite early in developing our budget." Free cash flow was -$2.1bn.

The disclosure under that hedge is the useful part. At June 30 Meta carried $349.31bn of non-cancellable commitments, about $81.65bn of it due in 2027 alone, plus $278.99bn of leases not yet commenced and roughly $68bn of new data centre leases signed in July that start in 2027 and 2028. Residual value guarantees sit at about $28bn and $13bn thresholds, with maximum exposure to loss of $46.03bn. That is the clearest public schedule of 2027 commencement exposure at a buyer with the cash to absorb it.

The first hard deferral was not a credit story

NextDC, in Australia, deferred A$140mn of contracted FY27 revenue after a Jemena grid upgrade slipped to 2028. Its FY27 revenue guidance went from A$615-640mn to A$475-500mn. It shelved all new developments and land acquisitions and cut capex guidance to A$4.5-5bn. The shares fell 42% in a day.

Note what caused it. Not a customer walking, not a covenant breach. A substation and a permit. That is the pattern I would expect for the next twelve months, and it is much harder to hedge than credit.

Certain, signal, unresolved

Certain, because it is in a document or on a screen: the FT project count, the Anthropic and Nscale contract terms, NextDC's guidance cut, Meta's 10-Q schedules, the CDS levels.

Signal, and this is my reading rather than a fact: that delivery slippage substitutes for demand failure. I could be wrong in the direction that matters, because slippage that lasts two years is a cancellation with better PR.

Unresolved: whether the slippage becomes contract terms. As of October 26 there is no reported take-or-pay amendment, no capacity rephasing, no volume deferral, no anchor-lease rewrite. The slippage is showing up as schedule, not as paper. That absence is what I keep testing, because the week it changes, the reset stops being deferred and starts being written down.

Two watch items carry the most information. The New Mexico air-permit decision for Project Jupiter on November 23, since permits decide whether a schedule holds. And xAI's 90-day rights, because a lab holding $518bn of stated obligations negotiating exit ramps is the closest thing here to sublease weakness in commercial real estate.

I am not calling a top. I am saying the loss has a new address now, and it is not the lab's.

Related reading: The AI compute reset wall just went public for the $518bn ledger and the OpenAI revenue revision, and The AI buildout runs on two clocks on why supply and credit stopped moving together.