Five years ago US nuclear was an industry managing its own decline. Then AI demand met its exact product — huge, constant, carbon-free power — and the order book reversed. The deals on our Power page tell the story properly; here is the shape of it.
Act one: stop the closures
The cheapest clean gigawatt is the one already built. Diablo Canyon was un-cancelled to at least 2029–30; Clinton got a 20-year Meta deal the month its subsidies were to expire. Every planned retirement in the US fleet is now being re-examined through an AI-demand lens.
Act two: restart the closed
Unthinkable until 2023, then routine: Palisades returned to service in October 2025 — the first restart in US history — with a $1.52B DOE loan. Three Mile Island Unit 1 follows for Microsoft (835 MW, 20 years, all of it), and Duane Arnold targets ~2029. The playbook now exists: fuel, staff, NRC re-licensing, hyperscaler offtake.
Act three: buy the output of the living
Amazon's Susquehanna arrangement — rejected by FERC as a behind-the-meter special deal, rebuilt front-of-meter, scaling toward 1,920 MW — wrote the regulatory case law: reactor-adjacent AI is allowed, but the grid must be kept whole.
Act four: order new ones
SMRs promise factory-built 15–345 MW modules sited at demand. Google signed for a Kairos fleet (500 MW by 2035); Amazon backs X-energy; Switch signed a 12 GW Oklo framework; Natrium is furthest along in actual construction. Honesty requires the caveat every record here carries: no Western SMR yet operates commercially; every date is a target, and the industry's cost history (Vogtle: years late, ~2x budget — then finished) is the base rate to beat.
How to read it
Nuclear cannot power this decade's buildout — the timelines don't fit; gas and renewables will carry the 2020s. What the AI deals actually bought is the 2030s: financing certainty that restarts pencil, that first-of-a-kind SMRs get built, and that the US nuclear supply chain (BWXT, Westinghouse) has demand worth retooling for. That is the trade both sides signed.
