Three Mile Island Comes Back for Microsoft

The most infamous name in American nuclear power is coming back to life to run a chatbot.

Three Mile Island is where the United States lost its nerve on the atom: a partial meltdown in 1979 that never killed anyone but froze new construction for a generation. The reactor that melted, Unit 2, is gone for good. Its undamaged twin, Unit 1, ran cleanly for decades and was shut in 2019 for the dullest of reasons, it couldn’t compete on price. Now Constellation Energy is spending 1.6 billion dollars to restart it by 2028, and has given it a fresh name, the Crane Clean Energy Center. Every megawatt it produces is already sold, on a twenty-year contract, to Microsoft, to feed the data centres behind its artificial intelligence.

The plant whose accident helped kill American nuclear is being raised from the dead, and not by a climate law or a change of heart. By a power bill.

That one deal holds the whole story people are telling about an AI-driven nuclear renaissance, and also the three things the story quietly gets wrong.

The first is that almost none of it is new. The revival everyone is cheering is overwhelmingly old reactors switched back on or kept alive, not new ones rising. Three Mile Island is a restart. Palisades in Michigan is a restart, propped up by a federal loan. Diablo Canyon in California is a stay of execution for a plant greens spent years trying to close. The genuinely new idea, the small modular reactor, the sleek unit meant to make nuclear cheap and quick, has essentially nothing running on the grid in the West. America’s flagship attempt, NuScale’s project in Idaho, was cancelled in 2023 when its price jumped roughly three-quarters before anyone broke ground. And when the industry does build big and new, it builds Vogtle: two reactors finished in Georgia in 2024, seven years late, at around thirty-five billion dollars, more than double the estimate and the costliest power plant in American history, with the bulk of the tab handed to ordinary Georgia customers whose disconnections climbed as their rates did.

So the “boom” is mostly recapturing power that already existed, wrapped around a pile of announcements for reactors that will exist, if they ever do, sometime in the 2030s.

Which is the second problem: timing. An AI cluster needs firm power in two to four years. New nuclear, at best, shows up in the next decade. That gap does not sit politely empty. The same companies posing with nuclear deals are quietly signing for natural gas and for solar backed by batteries to bridge it, and by some estimates most new data-centre capacity through 2027 will lean on fossil generation. The reactor is the halo in the press release. Gas is a good part of the fuel in the meantime.

The third thing is the one that matters if you are not training a model, and it is the one nobody puts on the slide. Restarting Three Mile Island does not add a single clean electron to your grid. It takes the entire output of an existing plant and routes it to one buyer. The same is true of the other marquee deals, the contracts on Susquehanna, on Diablo, on the rest: a fixed block of scarce, round-the-clock, carbon-free power is claimed privately, at a locked-in price, frequently wired straight into the data centre behind the meter so it barely touches the public network. But the megawatts a data centre takes are megawatts the rest of the grid does not get. That tightens supply, and tighter supply lifts the price everyone else pays.

And the cost of the grid itself, the new transmission lines, the spare capacity that keeps the lights on when everyone pulls at once, gets spread across all of us. It is already landing on bills. In the PJM region that runs the mid-Atlantic, the 2025 capacity auction blew out, and the market’s own independent monitor named data-centre demand as the main reason. Households there are looking at bills a few percent higher now and, stretched out over the years of build-out, something like seventy dollars a month more. The tech firm walks away with a clean, fixed-price supply and a green headline. The public gets the wiring invoice and the tighter market. Privatise the electrons, socialise the grid.

The fourth thing is smaller but worth saying, because the framing is everywhere: that tech money has converted the environmental movement, flipped the greens who spent fifty years fighting nuclear. It hasn’t. The movement didn’t reverse, it split, mostly along age. Younger, technology-friendly environmentalists were already coming around, polling on the atom runs well above half among the under-thirties, while the old guard, Greenpeace, Friends of the Earth, the Sierra Club, still sue over new reactors and have not moved an inch on waste or proliferation. The tech billions didn’t win the argument. They poured into the side that was already winning it.

Let me be fair, because some of this is genuinely good. Keeping Diablo Canyon open and restarting a sound reactor like Three Mile Island Unit 1 are, on the climate merits, exactly the right calls. Shutting working nuclear plants early was one of the worst own-goals of the old anti-nuclear politics, and AI demand is, almost by accident, rescuing clean generators that would otherwise have been scrapped. And if all this money finally drags the small modular reactor across the line into something that actually works and is actually cheap, that would be a real prize. I doubt it, the record is a graveyard of cost overruns and cancellations, but I would be happy to be wrong.

So the honest version is neither “AI is bringing nuclear back” nor “it’s all hype.” It’s that a handful of companies have suddenly decided they want gigawatts of firm, clean power, and the fight that actually matters is not green versus atom. That one is the sideshow everyone’s watching. The real contest is over who gets first claim on scarce clean electrons, and who pays for the grid that moves them. For now the answer is plain enough: the electrons go to whoever can sign a twenty-year contract, and the wires go on everyone’s bill. Three Mile Island is being resurrected for Microsoft. The next time a reactor is brought back in the name of clean energy, the question worth asking out loud is clean energy for whom, and paid for by whom.