Anthropic Buys 1.7 GW of Gas Generators: Power Now Sets the Pace of AI
Anthropic didn't buy more GPUs: it paid about $3.4 billion for 1.7 GW of gas generators by taking over orders someone else had already placed with Caterpillar. In the AI race the bottleneck is no longer silicon; it's electricity.
What Anthropic Actually Bought
The Financial Times reported it on Sunday, November 8, and Reuters, Bloomberg, Data Center Dynamics and TechCrunch followed: roughly $3.4 billion for 1.7 GW of heavy-duty gas generators for its data centers. That is generating capacity, not compute capacity, and the distinction frames everything else.
It Didn't Buy OnePWR: It Took Over Its Caterpillar Order
OnePWR Solutions, based in San Antonio, Texas, already held generator orders with Caterpillar. Anthropic took those orders over, including the engineering and construction work to install them. The Texas company was not acquired: it remains a supplier. The difference matters because it explains the mechanism — nobody bought a factory, someone bought a place in the queue.
A Deal Closed in September, Reported in November
The agreement was signed in late September 2026 and Anthropic did not announce it; the press reported it. There is no company statement confirming the deal or the final price, so the correct attribution is "according to the Financial Times", never "Anthropic announced".
Read also
The Six-Year Waiting List
The detail that turns this into an engineering story is the timeframe: reports put the wait for heavy-duty gas turbines at around six years.
How Turbine Slots Get Allocated
They are built in batches against factory slots. A new order joins the back of the production line and its delivery depends on how much spare capacity the factory has; an existing order already carries a reserved date because someone negotiated it earlier. Buying someone else's order isn't a discount: it's time. And for a lab competing to bring compute online every quarter, six years is an eternity.
Why Gas and Not Something Else
Connecting to the grid means an interconnection queue that in the United States can take years. Generating on site skips that queue, at the cost of buying fuel, securing permits and carrying price risk. Renewables remain part of the mix, but they don't solve the timing problem.
The Numbers Behind the Decision
$3.4 Billion Over 1,700 MW: About $2,000 per kW
The math is ours, not the company's, and it comes to roughly $2,000 per installed kilowatt, including engineering and construction according to the report. Power generation is measured this way, in dollars per kW of capacity, because what is being bought is the ability to produce rather than energy already produced.
From 3.5 GW of Compute in 2026 to Roughly 9.5-10 GW in 2027
Anthropic expects to move from more than 3.5 GW of compute capacity at the end of 2026 to roughly 9.5-10 GW by the end of 2027. Those are gigawatts of compute, not of generation, and they don't add up with the 1.7 GW order. As a scale reference, this order covers about half of one year's expected growth.
What the S-1 Admits
The October 26 IPO filing already warned that some of its data centers will depend on dedicated on-site natural gas generation, and that fuel and associated costs are borne by the company on a pass-through basis, with no price cap. That is not a minor clause: it is a risk declared in writing to investors.
What It Means If You Buy Compute
A GPU can be ordered and installed in months; a gas turbine drags a multi-year factory queue behind it. Compute availability for the next two years is already shaped by energy decisions made earlier, and the installation timeline for this specific order is not public.
The other effect is cost. If fuel is paid without a cap and its price is volatile, cost per token stops being purely a function of hardware and starts depending on the gas market. No announcement ties this purchase to API pricing, so any promise about the bill going up or down is speculation. When you reserve capacity, it is worth asking where the power comes from, how long it takes to arrive and who absorbs fuel volatility.
It's Not the Only One
Google has also secured its own generation for its facilities, and this month the scale of its Wyoming project became public. The strategic difference is one of degree: some buy equipment on the secondary market, others build the plant. Providers reselling compute compete for the same megawatts, so rental prices start moving with plant schedules instead of silicon launch calendars.
Conclusion
Whoever secures the megawatts secures the tokens. Gas is the fastest shortcut to capacity this year and the most expensive one to keep running afterwards, because it swaps a fixed hardware cost for a variable fuel cost. On this blog we keep tracking the infrastructure race with numbers and sources, because it is the one that will decide what the AI we use every day actually costs.
