The AI race is really a power grab
The bottleneck for artificial intelligence was never chips or models. It is electricity. The biggest tech companies already know it, and they are quietly buying reactors. Dated, sourced, no hype.
Every headline about AI is about the model. The real constraint is the wall socket. A frontier data center draws more power than a small city, the grid cannot add capacity fast enough, and the companies at the front of the race have figured out that whoever secures the electrons wins. So they stopped waiting for the utilities and started signing directly for nuclear power, including restarting a reactor that had been shut down. Here is the dated record, and the honest math on whether it is enough.
The receipts
A Carnegie Endowment analysis catalogued roughly 13 gigawatts of hyperscaler nuclear commitments: Microsoft contracting 835 MW from the restarted Three Mile Island Unit 1 (returning 2027), Amazon taking 1.9 GW from Talen’s Susquehanna plant plus up to 5 GW from X-energy by 2039, Meta signing Constellation, TerraPower, and Oklo, and Google backing Kairos Power and Elementl. The biggest names in software are now buying reactors.
Source: Carnegie EndowmentThe honest catch is in the same paper: “If all the hyperscaler-backed nuclear projects described in the previous section come to fruition, they would generate roughly 102 terawatt-hours-per-year. However, this would not be enough to meet even the lowest demand case let alone the 560 TWh/yr mid-range linear expansion estimate.” All of it together covers less than a fifth of the projected need.
Source: Carnegie EndowmentThe clearest single tell: Three Mile Island Unit 1, a reactor that was shut down as uneconomic, is being brought back online specifically to feed AI compute, with its output contracted to Microsoft. When a company reopens a mothballed nuclear plant to run chatbots, the scarce resource is not intelligence. It is power.
Source: Carnegie EndowmentStrip the AI story down to physics and it becomes an energy story. Models are copyable, chips can be fabbed, but a gigawatt of firm, always-on power takes years to build and cannot be conjured by a funding round. The hyperscalers understand this better than anyone, which is why they are signing twenty-year power contracts and reopening reactors while everyone else argues about model benchmarks. The constraint moved from the lab to the grid, and the companies that lock up generation and interconnection first will set the ceiling on what everyone downstream can build.
The front-run is to stop tracking parameter counts and start tracking megawatts, interconnection queues, and who controls baseload power near the data-center corridors. Even the giants’ combined nuclear buildout covers less than a fifth of the need, which means the fight for the remaining electrons is only starting. Follow the power, not the model. The winners of the AI economy will be decided at the substation, not the leaderboard.
