Does every AI lab now have to be a power company?

Chips and algorithms keep getting more efficient. So why is an AI lab buying gas generators by the gigawatt?

You would think that as chips and algorithms improve, the power bill shrinks. That is how most technology goes. Phones got faster and sipped less battery. So when a leading AI lab reportedly signed a $3.4 billion deal for 1.7 gigawatts of gas-fired generators, the natural reaction is: hang on, why on earth does it need that much?

Here is what happened. The lab took over generator orders that a Texas power supplier had placed with Caterpillar about two years ago. It is the biggest single purchase of power equipment by an AI lab that anyone knows of. And because it bought the kit directly instead of renting power through a partner, the risk of gas prices swinging is now its problem.

Efficiency does not save power here. It gets spent.

Every time a model gets cheaper to run, people ask more of it, use it for longer, and run more copies at once. Economists call this the Jevons paradox. Better steam engines did not cut coal use either; they made everyone want more engines.

And training is no longer the whole story. A growing share of the power goes into serving: millions of people chatting, plus agents that work for hours instead of seconds, plus the extra rounds of reinforcement learning that happen after the big pre-training run. One lab's quarterly revenue reportedly grew roughly twelvefold in a year, to nearly $17 billion. That money is people actually using the product. Using it takes electricity.

So does the generator deal prove they need way more energy per model than before? Not really. It looks more like a queue-jumping move.

Getting a grid connection can take years, and large gas equipment has long backlogs. These generators were ordered before prices spiked and arrive over the next two years. What is being bought is not just power but time. Time to power beats the price of power.

So does every lab have to become a utility? No. But a company with hundreds of billions of dollars in commitments, almost all of it for compute and infrastructure according to its latest filing, cannot sit and wait for the grid. Power is the one input you cannot swap for something cheaper. Caterpillar's power-generation sales jumped 30 percent last quarter, driven mostly by data centres. The AI race has quietly turned into a race for turbines.