The U.S. is undertaking an unusually large effort to restart retired nuclear capacity, pulling reactors out of decommissioning to help power artificial intelligence. Three plants across Michigan, Pennsylvania, and Iowa carry a combined projected output of approximately 2.25 GW, backed by federal loans and long-term energy agreements with Microsoft and Google.
Electricity has become a serious constraint on AI expansion. Data centers running accelerator clusters operate at high density around the clock, and new generation or transmission projects can take years to permit and build.
Three Plants, One Driving Forc
Holtec is working to restart the Palisades plant in Michigan, which had already entered active decommissioning. Holtec received financing support from the U.S. Department of Energy and is reported to be targeting a 2027 restart, subject to NRC approvals and ongoing project work, with an expected output of approximately 800 MW.
Constellation Energy is refurbishing Three Mile Island Unit 1 in Pennsylvania, now rebranded the Crane Clean Energy Center. Unit 1 is entirely separate from Unit 2, whose partial meltdown in 1979 made the site one of the most recognized addresses in American energy history.
Crane is projected to deliver 835 MW of carbon-free electricity demand at a refurbishment cost of approximately $1.6 billion. Constellation also received a $1 billion Department of Energy loan, as reported by the Philadelphia Inquirer.
Microsoft has signed a 20-year power-purchase agreement tied to the plant, structured to match the company’s data-center consumption on the PJM regional grid with carbon-free generation. Constellation identifies 2028 as its documented in-service estimate, subject to regulatory and construction progress, with the Nuclear Regulatory Commission still required to authorize key steps before the reactor can load fuel.
In Iowa, NextEra Energy Resources is partnering with Google to restart the 615 MW Duane Arnold Energy Center, which shut down in 2020. The project targets 2029 and has been linked to a Department of Energy loan of up to $1.9 billion, according to Gizmodo reporting.
NextEra CEO John Ketchum stated that Iowa residential customers would not bear the project’s costs. That reflects the company’s position rather than an independently verified assessment of all possible rate impacts.
Why AI Changed Nuclear’s Math
Plants that were retired because operating costs undercut their economics are now attracting billions in refurbishment investment, driven by hyperscalers that cannot wait for new generation to be built from scratch.
Reliable, low-carbon electricity at scale is exactly what hyperscalers need, and transmission corridors can take years to permit and build. Restarting an existing reactor, while still expensive and time-consuming, can sidestep some of those delays.
A power-purchase agreement gives a developer a commercial foundation for an expensive restart. Signing one does not mean electrons flow immediately to a specific facility; the electricity still moves through regional grid infrastructure like any other generation source.
Projected data-center electricity demand in Michigan, Pennsylvania, and Iowa runs into the gigawatts, but those figures are planning scenarios, not guaranteed loads. Projects in those forecasts will inevitably be delayed, downsized, or canceled.
The Number That Still Needs Earning
Projected capacity and available capacity are two very different things, and all three projects have significant milestones ahead.
The 2.25 GW figure is projected nameplate capacity, not power currently available to any data center. Technical, regulatory, and financing hurdles remain across all three projects before generation reaches the grid.
If these restarts succeed, retired nuclear plants could offer a replicable model for meeting AI-driven electricity demand, though success at three projects would not by itself guarantee the economics or regulatory conditions hold elsewhere. If schedules slip or costs escalate, hyperscalers may turn to natural gas, renewables paired with storage, or workload distribution across regions where power is cheaper and more accessible.
