Anthropic Bought $3.4 Billion of Gas Generators to Power Its AI Data Centers

AI labs have been described as buying electricity. Anthropic just bought the generators. The Financial Times reports the company paid roughly $3.4 billion for commitments to 1.7 gigawatts of heavy-duty gas-fired engines from Caterpillar -- ordered two years ago by a San Antonio equipment supplier that needed cash to finish paying for them. Anthropic did not acquire that supplier, and neither company has commented.

What the deal actually was

The structure matters more than the number. OnePWR Solutions had placed Caterpillar generator orders about two years earlier, at prices set before the surge. Anthropic took over those orders in late September -- along with the engineering and construction capability needed to install them. Anthropic did not buy OnePWR, and none of its employees moved to Anthropic. Piper Sandler lists a $3.4 billion sale of OnePWR's power generation platform to an unnamed frontier AI lab, dated September 2026.

Big generators now come with multi-year waits. Taking over an existing order is the equivalent of buying out someone else's restaurant reservation rather than joining the waitlist. The FT calls it the largest known single purchase of power-generation equipment by an AI lab.

Divide $3.4 billion by 1.7 gigawatts and you get about $2 billion per gigawatt, or roughly $2,000 per kilowatt of generator capacity. That is arithmetic from the two reported figures, not a clean equipment price -- the deal also bought installation capability, and the split was not disclosed.

Why gas at all

Anthropic's own IPO filing answers this. The amended S-1 states that "certain of our planned data center facilities rely on dedicated on-site natural gas-fired generation." In other words, some sites plan to generate their own electricity on site rather than buying it off a grid. The company also notes its arrangements "generally provide for fuel and related costs to be borne by us on a pass-through basis."

That is the tell. Grid interconnection queues are long, transmission is not keeping pace with demand, and a lab that wants capacity on its own schedule cannot wait for a utility. The realistic read is that on-site generation is a bridge -- a way to get megawatts sooner than the grid can deliver them.

Scale check

Against Anthropic's own disclosure, the deal is small. The amended S-1 lists roughly $895.4 billion in purchase and lease commitments, mostly compute. $3.4 billion is about 0.38 percent of that. Reuters reported in September that Anthropic expects to spend more than $500 billion over the next ten years building AI infrastructure.

For scale on the physical side: a gigawatt of continuous power is roughly the output of a large nuclear station, and one source estimates about 800,000 homes' worth of electricity per gigawatt. At 1.7 gigawatts, this single purchase is on the order of powering 1.4 million homes -- though a data centre load differs from a residential one in shape as well as magnitude.

How to read the report

Two honest caveats. The FT sourced this to a single person familiar with the matter, Reuters could not verify it, and neither Anthropic nor OnePWR responded to requests for comment. So the deal is reported, not confirmed -- though the Piper Sandler deal record is consistent with it.

The second caveat is the one that matters most: what the deal says about the next phase of AI infrastructure. For two years the story of the data centre buildout was that the constraint had moved from chips to power, with utilities scrambling to sign bespoke nuclear and gas contracts with the labs themselves. This is that future arriving -- the labs buying the machines themselves, taking the grid out of the loop, and quietly deciding the shape of the power mix for a whole industry.

Gas is the near-term answer because it is dispatchable and available. It is also the answer that ages badly, both politically and commercially, in a market where data-centre demand is colliding with decarbonisation targets and local resistance. Expect this deal to be studied as much for what it signals about the next three years of AI infrastructure as for the 1.7 gigawatts themselves.

Sources: Financial Times via CNBC, Bloomberg, Reuters via ETEnterpriseAI, aistockwire, CellCog. Compiled November 9, 2026.

Want more AI news and information? nov.link/skoolAI