AI's bottleneck is a power plant
Microsoft signed a 20-year deal to restart a reactor at Three Mile Island to power its data centers. The constraint on AI is moving from chips to electricity, and baseload nuclear is suddenly back.
On Friday, Constellation Energy announced a 20-year agreement to sell Microsoft the power from a restarted reactor at Three Mile Island. It’s Unit 1, not the one that partially melted down in 1979. It was shut down in 2019 for economic reasons, and Constellation plans to bring it back by 2028, renamed the Crane Clean Energy Center, with about 835 megawatts going to Microsoft’s data centers. Constellation’s stock jumped more than 20% that day.
A few years ago the idea of restarting a shut-down American nuclear plant, and at that site of all places, for a software company, would have sounded like a joke. It’s a clear sign of where the constraint on AI is moving.
For the last two years the bottleneck was chips. Everyone wanted Nvidia GPUs and couldn’t get enough. That’s easing. The next bottleneck is electricity, and it’s much slower to fix. A large training cluster can draw hundreds of megawatts, and the big companies are talking about gigawatt-scale campuses. The IEA estimated data centers used about 460 terawatt-hours in 2022, a bit under 2% of global electricity, and projected that could roughly double by 2026, with AI a major driver. Utilities in the US are revising their demand forecasts upward for the first time in decades.
As a mechanical engineer, I find this refreshing, because it forces people to think about physics. You can write software fast. You can’t build a power plant fast. A new gas plant takes years, and grid interconnection queues in many US regions are years long. Transmission lines take a decade. Nuclear plants traditionally take even longer.
Why nuclear specifically: data centers want power that’s always on. Training runs go 24 hours a day for months, and inference serving doesn’t stop at sunset. Solar and wind are cheap per unit of energy but intermittent, and batteries to smooth them at this scale are expensive. Nuclear produces steady baseload with no carbon emissions, which the tech companies care about because of their climate pledges. So existing reactors, and plans for small modular reactors, have suddenly become strategic assets for companies that had never thought about them.
There’s a supply chain underneath this that I’ve started reading about. Reactors need enriched uranium. After Fukushima in 2011, prices collapsed and stayed low for a decade, mines closed, and investment in new supply dried up. A lot of the world’s enrichment capacity is in Russia. If nuclear demand really rises because of AI, on top of what was already rising for climate reasons, the fuel cycle is a place where supply can’t respond quickly.
I’m not in the energy business and I’m not giving investment advice. But my expectation is that over the next decade, the companies that lead in AI will be partly energy companies, whether they like it or not, and that the real limit on how much intelligence we can run will be measured in gigawatts.