The reset wall is a funding problem now

@kawishwaqar|November 2, 2026 (17h ago)1 views

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

For a month I have written about a wall of contracted compute, and every version was about delivery. Projects late, permits blocked, substations missing. This week the story moved up one layer, to the cheque. The marginal buyer could not raise the equity it had already promised to spend.

OpenAI went out for $30bn and has about $11bn committed, less than half. The Financial Times reports the target was cut to roughly $20bn, with the first close expected below even that after investors trimmed cheques during the selloff. The valuation landed at $1.25tn against the $1.4tn first pitched, with MGX leading and a retail tranche to follow at the same price.

In February the company told investors it targets about $600bn of compute spend by 2030. The equity that spend was premised on is arriving at roughly a third of plan, and the round is now described as an opportunistic bridge before a listing.

I want to be precise about what this is not. Demand has not cracked. Azure grew 48% and Microsoft says demand exceeds supply. AWS grew 39%. Google Cloud margin rose 12.5 points to 36.2%, and Alphabet started renting third party capacity at the end of Q3 because it cannot build fast enough. A hyperscaler paying neocloud rent is new. The failure mode here is not customers walking. It is the capital that was supposed to build for them.

The lease book grew $158bn in one quarter

At September 30, leases signed but not yet commenced were $364.6bn at Microsoft, up from $329.1bn in June, $355.7bn at Meta, up from $279.0bn, and $130.5bn at Alphabet, up from $85.2bn. About $851bn across three companies, against roughly $317bn at the end of 2025. That total is one outlet's arithmetic on three 10-Qs, so treat each line as filed and the sum as derived. Microsoft's commencement window runs to FY2034, and since June its note flags that some arrangements are subject to unspecified contractual conditions.

Then the buyer-side admission. Microsoft spent $45.2bn of capex including finance leases, about $5bn under its own guidance, because some third-party capacity deliveries it expected in Q1 were delayed, in Amy Hood's words. Those are the GPU-as-a-service providers. Calendar 2026 is now about $170bn against $175bn in July, with the shifted deliveries landing in the second half of FY27.

Commercial RPO was $677bn against $678bn in June, while unstarted leases rose $35.5bn in the same quarter. Signed leases that have not begun have grown about four times as fast as contracted backlog since December.

The S-1 tells you which parts can be walked

Anthropic's filing puts $542bn to $567bn of future compute and infrastructure commitments on paper. $518bn of it sits with six partners, and about 80% is non-cancellable or payable regardless of use. Amazon at least $110bn through April 2036, Google at least $111.1bn through July 2033, Microsoft $31.4bn to May 2033 cancelable only for uncured material breach, and about $161.2bn of Broadcom equipment leases cancelable only on default. The flexible piece is xAI, up to $84.5bn, mostly cancelable on 90 days.

So the popular reading, that these are soft commitments dressed as backlog, is wrong for four fifths of the number. The obligation is real and it sits with the lab. But the remaining fifth is still tens of billions that can move, and the xAI line is a 90 day option, which makes it the closest thing this market has to sublease weakness. If that option gets exercised, the signal is that the best funded lab around decided the marginal gigawatt was worth more as cash than as capacity.

The revenue behind the borrowing got smaller

OpenAI's annualised revenue was re-rated from about $68bn to about $50bn this week. It matters outside OpenAI, because Oracle's commitment of roughly $300bn is about 45% of its remaining performance obligation, and the credit behind Oracle was sized against the higher figure. A smaller counterparty does not change a contract. It changes the price of the debt written against it.

CoreWeave is the same story with better disclosure. Its debt is contract-linked and carries a Cash-Trap trigger, about $35.1bn of it. The investment grade SPV refinance moved a slice of that paper into a structure rated on the leases rather than the parent. But a cash trap is the detail that matters, because it cuts the parent off from the cash the moment the contracts stop performing, which is the scenario the leases were supposed to make impossible.

The bubble call, and the objection I take seriously

Dalio, Torsten Slok at Apollo and Wellington have all called this a bubble, and the trigger they name is rates. That is the best objection to everything above. If this is a duration story, cheaper money brings the equity back and the wall gets funded.

I do not think rates are the binding constraint yet. The round was cut before the selloff reached its worst week. And when the buyers explained themselves this quarter, none of them talked about the cost of money. Microsoft cut capex because capacity was late. Alphabet is renting because it cannot build. A rate cut does not reschedule a substation.

Global AI-linked debt is approaching $570bn this year on the tally I have, into a quarter where the Nasdaq fell 2.86% on Friday to 7.8% below its October 6 record, its worst month since March 2025. Alphabet fell 12% in a day on the 29th after guiding capex higher, and chips fell more than 15% on the month. The punishment landed on the buyers for spending, not on the sellers for missing.

Certain, signal, unresolved

Certain, because it is filed or on a screen: the size and price of the OpenAI round, the three 10-Q lease lines, the Anthropic contract breakdown, Microsoft's capex miss and the stated reason for it.

Signal, and this is my reading rather than a fact: equity availability, not demand, is the binding constraint now. I could be wrong in the direction that matters, because a company valued at $1.25tn with an IPO ahead can still raise at almost any price.

Unresolved, and dated: whether the $50bn revenue figure gets read back onto Oracle's commitment and the Stargate paper behind it. Nobody has renegotiated anything yet. The revenue number is public. The contracts are not repriced.

Three dates carry the information. The New Mexico air permit decision for Project Jupiter on November 23, because permits decide whether a schedule holds. Anthropic's roadshow, no earlier than November 10 with a listing targeted for the week of November 16, because the price is the cleanest read on whether the wall can still be funded. Nvidia's Q3 and the refreshed Note 10 on November 17, where the last logged figure was $279bn of supply and capacity commitments, $88bn of it in FY2029. If that number rises while the OpenAI round undershoots, the gap is the story.

The demand side has not cracked. The funding side has, and funding is what decides which of these projects gets built.

Related reading: The reset wall slipped, and the exit clauses came with it on how slippage relocated the loss to project lenders, and The AI compute reset wall just went public for the original $518bn ledger.

Sources: WSJ on the OpenAI round (paywalled; the FT reported the target cut), the $851bn commencement tally, CNBC on Microsoft's Q1, CNBC on Amazon's Q3, Reuters on the October 30 selloff, Trepp on data centre CMBS, and Microsoft's 10-Qs on EDGAR.