Executive Summary

Key Numbers

Of the four numbers below, the last two come from outside this deal. The global turbine backlog explains why the line for a new order grew so long, and the $895.4 billion is the size of the commitment book this $3.4 billion sits inside.

Source: TechCrunch (citing FT reporting, November 9, 2026) · Wood Mackenzie · amended IPO filing (November 2, 2026)

The deal itself is simple. Anthropic secured 1.7 gigawatts of heavy-duty gas generators built by Caterpillar for about $3.4 billion. The Financial Times reported it citing a person familiar with the matter, and TechCrunch and Tom's Hardware, among others, picked it up. Anthropic and OnePWR both declined to comment, and Reuters said it could not independently verify the report. Most of what follows rests on that reporting.

One record exists outside the reporting as well. Piper Sandler, the investment bank that advised OnePWR on the sale, lists a transaction on its own deal page dated September 26, 2026. The client is OnePWR Solutions, what was sold is a "power generation platform," the buyer is described only as "a frontier AI lab," and the value is $3.4 billion. The buyer's name is not there. The date, the amount and the seller all match the reporting, but the part that says the buyer is Anthropic still rests on the Financial Times alone.

What stands out is that the direction of the deal changed partway through. Anthropic was originally going to be a OnePWR customer. OnePWR Solutions, founded in 2025 in San Antonio, develops behind-the-meter power plants and microgrids. Instead of buying electricity, Anthropic took the equipment order OnePWR had placed with Caterpillar. The engineering and construction capability needed to erect those generators came with it, and OnePWR is helping with the installation work at Anthropic's data centers.

The company itself was not acquired. OnePWR remains in place and none of its staff moved to Anthropic. What changed hands was a production slot locked in roughly two years ago and the hands to turn that slot into working generation. Another of the terms is that the order predates the recent round of price increases.

OnePWR's original plan shows how far this moved. In March 2026 Caterpillar, OnePWR and Vero3 announced a joint structure: natural gas generation bundled with carbon capture, battery storage and permanent underground storage of the captured carbon dioxide, supplied to data centers. The roles were split three ways. Caterpillar handled the generation equipment, the controls and the basic design of the capture plant. Vero3 stored the captured carbon dioxide and managed the related tax credits. OnePWR built, owned and operated the generation and sold power to customers under long-term contracts. A first project of 500 megawatts had been announced.

In that structure the AI company's place is the customer paying the electricity bill. Who owns and runs the generators is the supplier's problem, and the customer takes the contracted power. That was the place Anthropic was going to stand in. Then the financing stalled and the boundary between supplier and customer moved. Anthropic had the money to take hold of the equipment, and OnePWR had the spot inside the line. The March announcement did not name Anthropic, and there is still no confirmation that the generators now changing hands go to that 500-megawatt site.

Holding the power directly brings its own exposure. In the amended IPO filing of November 2, Anthropic wrote that certain of its planned data center facilities "rely on dedicated on-site natural gas-fired generation, which exposes us to fuel price volatility and supply disruptions that are outside of our control and for which we bear the cost on a pass-through basis." The filing adds that it carries those fuel costs without cost caps. A company that buys electricity watches only the bill. A company that holds the generators puts gas prices and the fuel supply chain on its own books.

The numbers explain why the wait for generation equipment grew so long. Wood Mackenzie counts the global turbine order backlog at about 110 gigawatts against annual manufacturing capacity of 60 to 70 gigawatts. Orders in the second quarter of 2026 alone reached 38 gigawatts, up 71% from a year earlier and a record. GE Vernova's contracted volume added 11 gigawatts in the third quarter, going from 116 to 122 gigawatts, and the company plans to lift output capacity from 20 gigawatts a year to 24 in 2028 and 30 in 2030. The volume already on the books has sold several years of production. That is where the Wood Mackenzie analyst's figure of six years for a new turbine in the United States comes from.

There is a counterview. Scott Strazik, GE Vernova's chief executive, says turbines are not the only bottleneck in data center construction. Building a single gas plant usually takes three years, and 80% of the company's backlog came from traditional customers rather than data centers. Delivery estimates differ by source as well. The Electric Power Research Institute puts large turbines at five years and small ones at 18 to 36 months. Six years is an average for the US market, not a number attached to every machine.

So the industry looks outside the line. Used turbines in good condition trade at $1.5 million to $1.7 million per megawatt, against $2 million to $2.5 million new. In one brokered sale two turbines totaling 29 megawatts went for $12.5 million, where the same capacity new would have cost about $35 million and a five-year wait, by the comparison the report of that sale drew. That sale prices out at $430,000 per megawatt, far below the quoted range, because the class of machine is different. The $1.5 million range is for the large turbines data centers want, while the two units on the brokerage market were 14.5 megawatts each. Converting retired jet engines is another route. Crusoe, an AI infrastructure company, ordered 13 units of 50 megawatts in April, 650 megawatts in all. Anthropic's path is neither of these. It is a third route: buying someone else's place inside the line outright.

So is the price of that place high? Anthropic paid about $2,000 per kilowatt. The table below gathers other figures written in the same unit. The four rows measure different things, so they give a range rather than a single yardstick.

The third row and the first row land in a similar band, which is the striking part. Combined-cycle plants being filed now for 2030 and 2031 startup carry capital costs in the $2,000-per-kilowatt range, and Anthropic paid money at that level for equipment it can use within two years. What each figure measures is different, so they do not convert into one another. Still, all four rows point the same way: what money buys in this market is not equipment alone.

Anthropic is a company that has mostly avoided buying physical equipment outright. Even its compute is rented. According to the IPO filing SpaceX made in May, Anthropic pays $1.25 billion a month through May 2029 for the compute capacity of Colossus 1, the facility SpaceX's subsidiary xAI built in Memphis. It is a 300-megawatt site holding more than 200,000 Nvidia GPUs, and Anthropic uses all of it. SpaceX wrote in the same filing that it expects to enter into additional similar services contracts for compute capacity with third parties.

The site it rents runs on 300 megawatts, and the generation capacity it has just bought is 1.7 gigawatts, more than five times that. The two numbers are written in the same unit but held in different ways. The rented side ends on 90 days' notice from either party, and the bought side stays on the books for years. Renting the compute while buying the power equipment is the shape of this deal.

The figures disclosed in the November 2 amended filing give a sense of the weight of that choice. Purchase commitments and leases together total $895.4 billion through 2045, most of it tied to computing and infrastructure. Of that, $447.5 billion is expected before the end of 2030 and $69.8 billion within a year of September 30. Cash and short-term investments in the same filing are about $132 billion, and operating cash flow for the nine months was $6.6 billion. Third-quarter revenue was $17 billion.

The risk factors in the filing make the case for this deal from the other side. Turbines and generators appear on the list of equipment facing extended delivery timelines and allocation constraints. Behind-the-meter arrangements are described as feasible only in certain jurisdictions under current regulation. Another passage says opposition to data centers is organizing in several regions and that the company's visibility could draw attention to its projects. TechCrunch carried alongside it an analyst's forecast that natural gas prices could more than triple in some markets by the end of the decade.

The decision to make a model bigger was for a long time a question of how to split a budget across data, algorithms and compute. This deal shows what that decision finally touches. Gigawatts and lead times. Neither is a unit the people who work with data usually count.

Designing to use training data sparingly has normally been filed under cost reduction. Stripping out duplicates and dropping the low-quality stretches buys the same performance on less compute, the argument goes. But once a unit of compute is tied to a power plan and an equipment lead time, that saving changes character. Compute you never needed does not simply come off the cost column. It lowers the pressure to secure today the equipment that arrives years from now. The same money, written into a different ledger.

What this article notes is that Anthropic handled its resources in two classes. It rented the compute and held the power. The test it applied was procurement time. The same question can be put to data quality.

• Is the training and evaluation data in use now something you can buy more of when you need it, or something that only accumulates over quarters? Attaching labels and provenance to data coming in from the field mostly belongs to the second kind.
• Who is working out what a data design that reduces compute is worth? In an organization where that saving converts into electricity and equipment, the calculation cannot stay on the data team's ledger alone.
• Do you know which resource takes longest to procure? Anthropic answered power and bought the equipment. If the answer at your own organization turns out to be data, the moment to invest is before it is needed rather than after.

This is also the place the phrase AI-Ready Data points at. Treated as a consumable you can always buy more of, data turns quality investment into an item that can be put off indefinitely. Seen as a resource you have to wait in line for, the order changes.

Thank you for reading this far. The original coverage is at TechCrunch and The Next Web. If you have started putting power or lead time into a column when you plan a model, we would be glad to hear what you are writing in it.

Primary Reporting

• 1.TechCrunch. (2026). "Anthropic joins AI's natural gas frenzy, buys $3.4B worth of turbines." Nov. 9, 2026.
• 2.Tom's Hardware. (2026). "Anthropic skips six-year waitlist with massive $3.4 billion gas turbine generator deal." Nov. 9, 2026.
• 3.The Next Web. (2026). "Anthropic buys $3.4bn of gas generators for its data centres, FT reports."

Industry & Market Data

• 4.Wood Mackenzie. (2026). "Gas turbine prices soar 195% as market faces supply-demand crisis." Press release.
• 5.Decrypt. (2026). "AI Data Centers Are Buying Used Jet Engines as New Power Turbines Take Six Years."
• 6.Kern, M. (2026). "The Gas Turbine Shortage Just Became AI's Biggest Constraint." OilPrice.com.

Official Filings & Financials

• 7.CNBC. (2026). "Anthropic reports $17 billion in Q3 revenue, purchase commitments of $895.4 billion in revised S-1." Nov. 2, 2026.