Anthropic Buys $3.4 Billion in Gas Turbines: 1.7 GW to Power AI

Anthropic has always been a company that rents rather than builds. Servers, capacity, even compute from a competitor: it rented all of it. And now it is buying $3.4 billion worth of gas turbines. When a company like this goes shopping for hardware to generate electricity, the energy problem is no longer theoretical.

What exactly happened

According to the Financial Times, as relayed by TechCrunch, Anthropic is buying Caterpillar generators with a combined capacity of up to 1.7 gigawatts. That is the scale of a large power plant. They were originally ordered by equipment supplier OnePWR, a couple of years ago. But it was unable to complete the purchase: it did not find financing. Why this happened with a customer like Anthropic on the books is unclear.

Anthropic had planned to buy the turbines from OnePWR, but ended up going directly to the manufacturer. It likely also wanted to make sure the capacity did not go to a competitor. That is the article author's conjecture, not a confirmed fact.

Why this breaks with its usual strategy

In its S-1, filed in May, Anthropic disclosed that it pays xAI (a SpaceX subsidiary) $1.25 billion a month for compute capacity. The agreement runs through May 2029. That is about $15 billion a year. The company also wrote there that it expects "additional similar service contracts." In other words, the core plan was simple: rent other people's capacity and grow on top of it.

For comparison, $3.4 billion on turbines is less than three months of rent from xAI. Not a gigantic sum against its other expenses. But the move itself shows that renting compute is not enough if there is no electricity behind it.

The bottleneck here is not chips but energy. The queue for gas turbines has stretched into the early 2030s. Grid connection can take just as long. If you did not get in line two years ago, all that is left is buying up other people's orders, and Anthropic seems to have done exactly that.

It is not alone. Google, Amazon, Microsoft and Meta have also announced plans to build large gas power plants next to their data centers so they do not depend on the shared grid. The surge in demand has even revived interest in fracking. Under current plans, US data centers could consume more gas by 2035 than Germany and Japan combined.

Risks and what to do about it in practice

Betting on gas has a downside. The price of fuel becomes a direct line item in the cost of compute. According to one analyst's estimate, by the end of the decade gas in some markets could become more than three times more expensive. In that case, Anthropic may bitterly regret the purchase. This is one expert's forecast, not a guarantee, but the risk is real: the turbines cannot be returned, and the fuel price is out of its control.

What this means for those building products on top of models:

• Do not build a steady decline in token prices into your business model. If energy gets more expensive, cheap compute stops being a given.
• Capacity constraints may show up as rate limits, access queues and priority for large customers. Plan a fallback at the architecture level: a second provider or a lighter model for non-critical requests.
• Inference efficiency is becoming a real competitive advantage. Caching, routing requests by complexity, and smaller models where they get the job done save not only money but also kilowatts.
• If you are a founder looking for infrastructure, consider where your provider physically gets its electricity. An energy shortage is a risk for the supplier, and therefore for you.

One more important detail: Anthropic is preparing for an IPO. For investors, buying turbines is a signal that the company is ready to invest in its own infrastructure and control bottlenecks. On the other hand, these are heavy assets and a dependence on the gas market, which the pure-rental model did not have.

The overall conclusion is simple. The AI race is now being run not only for the best models and chips, but for gigawatts. Whoever secures its own energy first will be able to offer more capacity and better withstand demand spikes. Everyone else will have to rent what is left and pay a price that is under no obligation to fall.