Evidence File — Anthropic 1.7GW gas generator deal
Article id: anthropic-onepwr-gas-turbines-2026-11-09
Discovery candidate: 5ea29f7e53fe0f14
Compiled: 2026-11-10 by COMPUTER (StarNet)
Revised: 2026-11-10 (revision 6) — fifth review returned PASS; its non-blocking findings applied
Status: PUBLISHED 10 November 2026 after the Commander approved publication. Independent review by NEWS RESEARCH recorded PASS against the reviewed version (sha256 b655f4b8...); the published page differs from that version only by site chrome, the publication timestamp and the confirmation line.
Source retrieval status
The evidence shape is unusual and is set out plainly: the originating report is paywalled and was not retrieved, three outlets that relayed it were retrieved in full, and one headline is quoted without having been read. Breaking the last of those out is what revision 4 corrects — an earlier version claimed every quotation came from a page read in full, which was untrue.
| # | Source | Retrieved? | Role |
|---|---|---|---|
| 1 | Financial Times — "Anthropic strikes $3.4bn deal to generate gas-fired power for data centres" | NO — paywalled | Originating report. Origin of the generator figures. Not read in full. |
| 2 | Data Center Dynamics — "Anthropic signs $3.4bn gas generator deal to support 1.7GW of data center power", Sebastian Moss, 9 Nov 2026. Subheading: "Takes over OnePWR order of Caterpillar gen-sets" | YES — read in full (station browser, after HTTP 403) | Deal terms, installation arrangement. Also the source for Anthropic's prospectus figures and the Akamai/CoreWeave omission — but note those are attributed by DCD to Anthropic and its prospectus, NOT to the FT. |
| 3 | CNBC — "Anthropic secures $3.4 billion generator deal for AI data centers, FT reports", 8 Nov 2026 | YES — read in full | Late-September timing; "did not acquire OnePWR / no employees transferred"; no-comment and Reuters-could-not-verify statements; the $500bn ten-year infrastructure figure. |
| 4 | TechCrunch — "Anthropic joins AI's natural gas frenzy, buys $3.4B worth of turbines", Tim De Chant, 9 Nov 2026 | YES — read in full | The direct-to-Caterpillar description; the early-2030s backlog that resolves the waitlist question; the unidentified financing gap; rent-rather-than-buy context; the IPO plan. |
| 5 | Tom's Hardware — "Anthropic skips six-year waitlist with massive $3.4 billion gas turbine generator deal" | NO — URL not captured; page body did not render | Quoted only for the phrase "six-year waitlist". Nothing in the article depends on it. |
Provenance classes — three different kinds of claim
Revision 4 separates these because revision 3 mis-merged them, attributing Anthropic's own disclosure to the FT.
- Financial Times, relayed: the generator deal's headline terms — the $3.4bn figure, the 1.7GW contract, the partnership with OnePWR Solutions. DCD's FT tag reads: "The Financial Times reports that the 1.7GW contract will see the company partner with Texas-based equipment supplier OnePWR Solutions."
- DCD's own reporting, not FT-attributed: the takeover of the Caterpillar order and the two-year availability window sit in DCD's own paragraphs, outside its FT tag. TechCrunch, separately, attributes the ordering history to "the report". The claim table below files these accordingly; an earlier revision of this bullet wrongly put them under FT-relayed, contradicting the table in the same file.
- Anthropic's own disclosure, as reported by DCD: the $895.4bn commitment total through 2045, the individual cloud commitments, and the Akamai/CoreWeave omission. DCD writes "Anthropic revealed $895.4bn" and "According to its prospectus".
- New York Times, cited within DCD: the >3.5GW compute estimate for this year and 2027.
Claim-level verification
| Claim in the article | Status | Basis |
|---|---|---|
| $3.4bn deal for Caterpillar gas generators capable of 1.7GW | REPORTED (FT, via retrieved secondary) | [2], [3], [4] |
| Supplier is OnePWR Solutions, San Antonio, Texas | REPORTED (FT, via retrieved secondary) | [2], [3] |
| Anthropic bought the commitments in late September | REPORTED — single outlet | [3] only. |
| OnePWR ordered the generators two years ago and needed financing | REPORTED — DCD's own text | [2], [4]. DCD states this in its own voice; its FT tag covers the deal and the OnePWR partnership, not every subsequent sentence. |
| Anthropic did not acquire OnePWR; no OnePWR employees transferred | REPORTED — single outlet | [3] only, citing "a person familiar with the matter". |
| Deal included engineering and construction capabilities to install the generators | REPORTED | [2] (DCD's own text), [3] (CNBC, citing a person familiar) |
| Generators expected available within two years; OnePWR to help with installation | REPORTED — DCD's own text | [2]. Not tagged as FT attribution by DCD. |
| Neither Anthropic nor OnePWR commented; Reuters could not immediately verify | VERIFIED — stated in retrieved source | [3] |
| Turbine backlog stretches into the early 2030s; grid queues similarly long | VERIFIED — stated in retrieved source | [4] |
| Why OnePWR could not secure financing is unclear | VERIFIED as an explicit gap in [4] | [4]. Not filled by the article. |
| Anthropic "has generally avoided buying physical infrastructure"; S-1 expects further rental contracts | ATTRIBUTED — TechCrunch characterisation | [4] |
| Anthropic planning to go public this year | REPORTED | [4] |
| Anthropic expects to spend >$500bn over ten years on AI infrastructure | REPORTED — Reuters, attributed within [3] | [3] |
| $895.4bn in cloud and data centre commitments through 2045 | ANTHROPIC DISCLOSURE, as reported | [2] — "Earlier this month, Anthropic revealed $895.4bn". Not an FT attribution. |
| Google $121.3bn, SpaceX $120.5bn, Amazon $110bn, Microsoft $44.9bn | ANTHROPIC PROSPECTUS, as reported | [2] — "According to its prospectus". |
| Nscale $44.6bn to 2033; Lambda $41.3bn to 2034; Fluidstack $50bn; TeraWulf and Hut 8 contracts | ANTHROPIC PROSPECTUS, as reported | [2] |
| Akamai ($11.6bn) and CoreWeave agreements not mentioned by Anthropic | ANTHROPIC OMISSION, as stated by DCD | [2] — "Anthropic did not mention its $11.6bn cloud deal with Akamai, nor its compute agreement with CoreWeave." The omission is Anthropic's, not the reporter's. |
| Anthropic expected to have >3.5GW compute this year, similar in 2027 | REPORTED — NYT, cited within [2] | [2] |
Mechanism divergence — what is actually disputed
Revision 3 treated this as three mutually exclusive accounts. Checked against the sources, it is a two-against-one split with an outlier:
- [2] DCD, subheading: "Takes over OnePWR order of Caterpillar gen-sets"; text: Anthropic will "partner with Texas-based equipment supplier OnePWR Solutions".
- [3] CNBC: Anthropic "bought commitments for 1.7 gigawatts of heavy-duty generators in late September from OnePWR Solutions".
- [4] TechCrunch: Anthropic "had planned to buy them from OnePWR but instead went directly to Caterpillar, the manufacturer".
Acquiring OnePWR's commitments and taking over its order are compatible descriptions; DCD and CNBC sit together. TechCrunch's direct contract with the manufacturer is the divergent account. No retrieved source explains the difference, and the FT could not be read. The article therefore states what every account supports — Anthropic acquired control of generators OnePWR had ordered — and declines to adjudicate who sold them. Body and this file now say the same thing; revision 3 had them contradicting each other.
Waitlist framing — reconciled
Revision 1 recorded the "six-year waitlist" phrase [5] and the two-year availability window [2] as an unreconcilable discrepancy. That was an error: [4] reports that "the backlog for natural gas turbines stretches into the early 2030s", which describes the queue for new orders, while [2] describes delivery of already-ordered equipment. Both claims can be true. Note that [5] was never read — the phrase is quoted from a headline, and the article now says so.
Claims deliberately not made
- That Anthropic or OnePWR has confirmed the deal or the amount. [3] states they did not respond to requests for comment.
- Which description of the transaction mechanism is correct.
- Why OnePWR could not secure financing. [4] states this is unclear.
- Why the Akamai and CoreWeave agreements were omitted.
- That the generator figures are an Anthropic disclosure rather than FT reporting, or vice versa for the prospectus figures.
- That the generators will be installed on the stated timetable.
- Any emissions, cost-per-MWh, or data-centre-count figure.
- Any household-equivalent conversion of 1.7GW. The sources give capacity, not consumption equivalences.
Duplicate and overlap check
- Published articles: no THAF article covers this deal.
- Path B queue: none of the queue items covers it.
- Discovery dataset: two rows cover the same event —
5ea29f7e53fe0f14(Tom's Hardware, score 25) and48b9783d0be50413(TechCrunch, score 25). Neither carries aduplicate_oflink. This is the single article for the event.
Review status
Revision 1 — HOLD (analyst, 2026-11-10). Mechanism presented as settled; the waitlist passage was itself an error; CNBC's material facts missing. Addressed in revision 2.
Revision 2 — REVISE (analyst, 2026-11-10). Six findings, including a converter defect that made every staged package assert a completed review it did not have; headline asserting the contested mechanism; a revision note that misdescribed its own diff; a backwards Reuters attribution; a synopsis that overstated confidence. Addressed in revision 3.
Revision 3 — REVISE (analyst, 2026-11-10; record reviews/findings-onepwr-rev3.json). Nine findings, including: the $895.4bn misattributed to the FT when DCD attributes it to Anthropic; a false claim that every quotation came from a page read in full; body and Evidence File contradicting each other on the mechanism; the synopsis still asserting completed possession; and process prose leaking into the published body. Addressed in revision 4.
Revision 4 — REVISE (analyst, 2026-11-10; record reviews/findings-onepwr-rev4.json). Two MAJOR findings, both about revision 4's own revision note rather than the article: it claimed a row had been deleted from this claim table that was still present, and claimed a false statement had been recorded when it had not. Both are fixed here and in the article's note. This table no longer lists the "largest known single purchase" phrase, which the article never uses. Four further MODERATE findings — FT over-attribution in this table, a missing byline, commitment conditions omitted, and a mislocated Akamai omission — are addressed in revision 5.
Revision 5 — PASS (analyst, 2026-11-10; record reviews/findings-onepwr-rev5.json). Twelve quotations verified verbatim against retrieved sources, zero mismatches; every claim in the rendered article checked against a source. Twelve findings, all classified non-blocking. Four matters of substance emerged: the package pointed at a topic page that returns HTTP 404; the synopsis said the companies "did not agree to answer questions" where CNBC reports they "did not respond to requests for comment"; this file's header still announced revision 4; and this file's provenance bullet still filed the takeover and two-year window under FT-relayed while its own claim table filed them as DCD's. All four are fixed in revision 6.
Recorded here as false, not removed: revision 2's revision note claimed to have withdrawn the phrase "Anthropic did something clever" from the synopsis. It had not — the sentence was still present, and the third review caught it. Revision 4's note then made the opposite error, asserting the sentence still stood when it did not. Neither error is deleted from this record. Which revision actually removed the sentence was never noted at the time, and that gap is stated rather than reconstructed.
Independent review of revision 6: NOT PERFORMED AT TIME OF WRITING. The revision 5 PASS is bound to the revision 5 hash and does not carry to this one. The article may not be approved or published until a review of revision 6 is recorded.
Publish-step requirement: this file and the article's visible AI-authorship disclosure both currently state that no review is recorded. The moment a PASS is bound to this hash both become false, and the article HTML must be rebuilt with the actual verdict and reviewer before release. That is not optional.
The unresolved question, for the reviewer and the Commander: the generator figures still originate from one FT report that no retrieved source read in full, and neither Anthropic nor OnePWR has confirmed anything on the record. Revision 4 discloses this in the body and in the synopsis. Disclosure does not cure it. Three consecutive reviews have judged that relay-only sourcing does not by itself block publication; whether the Commander agrees is their call, not this file's.