Microsoft's Three Mile Island restart, in numbers

What the deal actually does

Microsoft is not buying a new reactor. It is paying to revive an existing one. Unit 1 closed in 2019 because it could not compete on price with subsidized renewables and cheap fracked gas in Pennsylvania's wholesale market. The reactor itself was, and remains, in good condition — and held a federal operating license. The $1.6 billion restart cost covers refurbishment of turbines, generator, transformer, and cooling/control systems, plus a license extension through 2054.

Microsoft's 20-year PPA gives Constellation the financial certainty to take that capex risk. It also matches an estimated ~7 TWh/year of Microsoft's PJM-region data center load with carbon-free generation — addressing a measurable chunk of Microsoft's reported 23.4% emissions increase in fiscal 2024, driven mostly by AI training workloads.

The precedent matters more than the megawatts

835 MW is not enormous on its own. It's about 0.85% of the existing U.S. nuclear fleet (98 GW across 96 operating reactors) and roughly one-tenth of OpenAI's Stargate target. What makes this deal unusual is that no fully decommissioned U.S. reactor has ever been brought back to commercial service. The closest precedent is Holtec's Palisades plant in Michigan (800 MW, shut May 2022), originally targeting a late-2025 restart. As of May 2026 Palisades has been pushed to early 2026, then late March, and is still working through NRC inspections — 1,400 cracked steam generator tubes (being repaired via "sleeving"), missing weld documentation, unauthorized nozzle welds, and pending litigation from Beyond Nuclear. Whichever plant powers up first will create a category that didn't exist 18 months ago.

For context: roughly 13 GW of U.S. reactor capacity has been retired since 2013 — Vermont Yankee, Indian Point, Pilgrim, San Onofre, Diablo Canyon (still operating but on borrowed time), among others. If the combination of AI-driven demand, federal 45Y tax credits, and PPA willingness from hyperscalers makes some fraction of those revivable, the restart pipeline starts looking material — possibly several gigawatts over the next decade.

What 7 TWh/year actually powers

AI-campus equivalent assumes ~95% utilization on a 1 GW load. Hyperscale campus sizes vary widely; Stargate's flagship Abilene site alone is targeted at ~1.6 GW.

Why this story is in our compare engine

The TMI restart is a clean test case for the headline thesis behind this site: nuclear's land-use and reliability advantages compound dramatically when you tally them up against firm-equivalent renewable capacity. The original TMI site is about 600 acres. Powering the same 7 TWh/yr from PA-grade utility solar, with the storage and overbuild needed to firm it up, would require roughly 113 square miles — about 120× the land of the existing reactor footprint, on completely new sites.

Use the Compare app to plug different metros and power sources together — Pennsylvania values are pre-loaded under any of our PA-anchored metros (Philadelphia and Pittsburgh adjacent), and the Compare engine will produce a downloadable PNG showing the same scaled overlay used in this post.