📰 Key Takeaways

Anthropic has always avoided buying physical infrastructure itself, but now it’s spending $3.4 billion on natural gas turbines, a sign of how serious AI’s power gap has become. According to the Financial Times, these Caterpillar generators were originally ordered years ago by equipment supplier OnePWR, with total capacity of up to 1.7 gigawatts, but OnePWR could never find the funding to complete the purchase, so Anthropic bypassed OnePWR and bought directly from the manufacturer, Caterpillar. The original article speculates that Anthropic may also not have wanted to give competitors a chance to snap up this generating capacity.

In the past, Anthropic preferred renting to buying. According to its S-1 filing in May, the company pays xAI, now part of SpaceX, $1.25 billion a month for compute through May 2029, and says it expects to sign more service contracts like it. But the backlog for natural gas turbines stretches into the early 2030s, and the wait to connect to the grid is just as long, which makes this purchase look like a last resort.

This trend isn’t unique to Anthropic. Google, Amazon, Microsoft, and Meta have all announced plans to build large natural gas power plants that supply data centers directly on site, instead of relying on the public grid. That’s also giving new life to hydraulic fracturing, the technique used to extract natural gas. If construction goes as planned, US data centers could use more natural gas by 2035 than Germany and Japan combined. On top of that, relying on natural gas exposes operators to price volatility, and some analysts expect gas prices in certain markets to more than triple by the end of this decade, at which point Anthropic may well regret the deal.

💬 JudyAI Lab Take

Anthropic used to prefer renting compute and avoiding buying physical facilities, but now it’s spending $3.4 billion on natural gas turbines with total capacity of up to 1.7 gigawatts. This shift shows that AI’s power gap has become so severe that even the rental route isn’t enough.

What this story reflects is that the center of gravity in AI competition is moving down from the models themselves to energy and the supply chain. The original article notes that the turbine order backlog runs into the early 2030s and the wait to connect to the grid is just as long, which is why Anthropic bypassed OnePWR and ordered directly from Caterpillar, and it may also not have wanted rivals to grab this generating capacity. Google, Amazon, Microsoft, and Meta have likewise announced plans to build natural gas plants on site at their data centers. For AI builders, this means compute isn’t just a line item on a bill. It’s a constraint shaped by scarce physical resources. At the same time, relying on natural gas brings price risk too, and some analysts expect gas prices in certain markets to more than triple by the end of this decade, so today’s fix may not be a good long-term one.

I’d suggest taking stock of how sensitive your product is to compute prices and supply, and leaving yourself room to switch models or providers.

📅 Source Info

🔗 Further Reading