In 2024, Constellation Energy agreed to spend roughly $1.6 billion restarting the undamaged reactor at Three Mile Island, idle since 2019, under a twenty-year deal to sell its output directly to Microsoft for AI data centers. A power plant that closed for being economically unviable is reopening because a single customer's compute demand made the math work again.
Every inference — every answer a model produces — draws real electricity through real silicon in a real building with real cooling requirements. As usage scales from occasional queries to constant, ambient use across products, the aggregate draw stops being a rounding error and starts being a planning input, which is why utility planners in regions with new data center buildout are now forecasting grid demand around AI training and inference schedules.
The industry faces a genuine efficiency race — smaller models, better chips, smarter caching — running alongside a demand race, where each efficiency gain gets partly absorbed by more usage rather than translating fully into lower total consumption, a pattern familiar from other domains of computing history.
Communities near new data center construction are experiencing this most directly, through competition for grid capacity, water for cooling, and land — tradeoffs that used to be abstract policy debates and are now local zoning fights with AI's name attached to them for the first time.
Microsoft's Three Mile Island deal is not an isolated bet. Amazon has invested in X-energy's small modular reactor design and Google has signed an agreement with Kairos Power, both aimed at the same underlying problem — securing dedicated, carbon-free power at a scale the existing grid was not built to absorb on demand.
For years, the cost of thought was measured in human time. It now has a second meter running alongside it, legible to anyone who reads a utility filing — and Constellation's restarted reactor is a twenty-year bet that meter keeps climbing, not a hedge against it falling.