Anthropic Seals $3.4B Gas Generator Deal, Says FT

Anthropic has committed to pay approximately $3.4 billion for 1.7 gigawatts of large-scale gas generators to power its data centers, according to a Financial Times report published on Sunday (8). Both Anthropic and supplier OnePWR Solutions have not commented on this matter, and Reuters has reported an inability to independently verify the article. Anthropic had not publicly commented by the time this report was published.

How the Deal Was Structured

According to the FT, Anthropic took on orders for Caterpillar generators that OnePWR, based in San Antonio, Texas, had placed around two years prior and was unable to finance. The AI company did not purchase OnePWR, and no employees from the supplier were transferred. The package, according to the newspaper, includes the engineering and construction capabilities needed to install the equipment. The report does not specify where the generators will be used.

The implicit calculation is around $2,000 per kilowatt, already including engineering and construction. What Anthropic bought, however, is primarily a position in the queue. Orders placed two years ago mean machines that arrive before any new order, in a market where the wait for grid connection can exceed the timeline for constructing the data center itself. In Northern Virginia, the largest data center hub in the world, connecting 100 megawatts to the grid takes about seven years, according to data from the International Energy Agency cited by TechCrunch.

Gas alternative engines, such as those from Caterpillar, have become the route for those who cannot wait for turbines, whose orders are committed for years at the major manufacturers. The choice has a cost: local gas generation requires environmental permits, and American operators are already facing state resistance to this model. In July, Virginia Mercury reported that operators want to build their own turbines in the state, which has opened a dispute over whether local generation could circumvent state renewable energy laws.

The Timing: Roadshow Week

The report comes out in the week when Anthropic is expected to begin presenting its initial public offering to investors, according to The Information and Reuters. Reuters reported, based on four people familiar with the discussions, that investors are pushing for a valuation below the $1.5 trillion recently discussed, with figures between $1.2 trillion and $1.3 trillion circulating. The revised prospectus, filed on November 2, still leaves the price range blank. The company plans to list the shares on Nasdaq under the ticker ANT.

The contract was signed in September, before the recent drop in AI stock prices, and has come to light now. For an investor evaluating the offering, a commitment of $3.4 billion in energy demonstrates demand for contracted computing, but also capital tied up in physical assets that a software company would typically lease. The competition for the same investor base is fierce: excluding Anthropic's offering, AI companies have sold or priced over $300 billion in shares in 2026, with about $100 billion on the way, according to a survey by The Information.

Reading Across Markets

In the United States, the deal places an AI laboratory within the electric sector of Texas and any state where the generators will be installed. A block of 1.7 GW needs an address, and the decision of where to connect it involves state permits, gas pipeline networks, and utility commissions, not just a cloud contract.

In Brazil and India, the channel is the manufacturer's queue. Data center operators in both countries rely on generator sets for backup power, often from the same global suppliers. When a single buyer absorbs 1.7 GW of already placed orders, the delivery timeframe for those ordering afterward tends to extend, and colocation projects in São Paulo or Mumbai enter into the same competition for engines that an American laboratory has just won with raised capital.

The detail that speaks most about the strategy is what Anthropic left out. They did not acquire OnePWR, did not absorb its team, and retained only the order, engineering, and construction. It’s a design from one who wants control over a scarce input without setting up an energy company, and indicates that the next frontier of negotiations between AI labs and suppliers is in the equipment yard, not in the pricing table per token.