The Reactor That Outlived Its Own Disaster
In March 1979, a valve stuck open at the Three Mile Island nuclear plant in Pennsylvania. Cooling water drained away. The core began to melt. For five days, the country held its breath while engineers struggled to understand what was happening inside a containment building they could not see into. It was the worst commercial nuclear accident in American history — a partial meltdown that released a trace of radioactive gas, killed no one directly, but ended the nuclear age in the United States as surely as if it had flattened a city.
No new reactor was approved for thirty years. The industry withered. TMI became a name synonymous with hubris, a cautionary tale told in textbooks and documentaries. Unit 2, the one that melted, was entombed. Unit 1, which had been running perfectly next door, was eventually shut down in 2019 because it could not compete with cheap natural gas. Decommissioning began. The plant was dying the slow death of obsolescence.
Then artificial intelligence got hungry.
In September 2024, Microsoft signed a twenty-year power purchase agreement with Constellation Energy to restart Unit 1. Not a new reactor. The same one. The 835-megawatt pressurized water reactor that had been operating since 1974, that had survived the meltdown next door, that had been deemed economically worthless five years earlier. Constellation would invest $1.6 billion to refurbish it, backed by a $1 billion federal loan. The Department of Energy, which had once regulated the nuclear industry into a coma, was now helping wake it up. The plant would be renamed the Crane Clean Energy Center, as if a new name could distance it from the ghost next door.
The target date is 2028. All 835 megawatts will go to Microsoft, feeding the data centers that train the models that write the code that increasingly builds itself. The irony is so thick you could cut it with a Geiger counter: the most symbolically damaged piece of energy infrastructure in American history is being resurrected to power the technology that is supposed to represent the future.
But the future is not as clean as its marketing. Training a large language model can consume as much electricity as a hundred homes use in a year. Data centers already account for about 4% of global power demand, and that share is climbing. The hyperscalers — Microsoft, Google, Amazon, Meta — have discovered that renewable energy, for all its virtues, cannot provide the baseload power they need. Solar panels sleep at night. Wind turbines rest when the air is still. Batteries are improving but not enough. The only source of carbon-free electricity that runs continuously, that does not depend on weather, that can scale to the gigawatts these companies are now demanding, is nuclear fission.
So the industry that America abandoned is being rebuilt by the industry that replaced it. And not just at Three Mile Island. Holtec International is restarting the Palisades plant in Michigan, which has been offline since 2022. NextEra is studying a restart of Duane Arnold in Iowa. Amazon bought a 960-megawatt campus next to Talen Energy’s Susquehanna nuclear plant. Google signed a deal with Kairos Power for small modular reactors. Oracle’s Larry Ellison announced plans for a gigawatt of nuclear-powered data center capacity. Meta, after flirting with space-based solar power and long-duration batteries, is now negotiating for nuclear procurement of one to four gigawatts.
The nuclear renaissance is not being led by environmentalists or energy policy experts. It is being led by software engineers who need more watts.
I keep thinking about the workers at TMI Unit 1. They operated a reactor for decades next to a melted one, under a cloud of public fear and regulatory scrutiny, producing reliable electricity that the market eventually decided was not worth the price. Then they were told to shut it down. Decommissioning crews arrived. The turbine was tuned for the last time. The paperwork for burial began.
And now someone is telling them to turn it back on. Not because the country needs the power for hospitals or schools, but because a cloud computing company in Redmond, Washington needs the electrons to train neural networks that can generate marketing copy and write Python scripts. The reactor that was too expensive for ordinary life is suddenly worth $1.6 billion to the machine that is replacing ordinary life.
There is something quietly devastating in that reordering of value.
The FERC decision is expected this summer. If approved, the restart will proceed. The Nuclear Regulatory Commission will review the safety case. The cooling towers will vent steam again. And somewhere in a data center in Northern Virginia or Ohio, a rack of GPUs will draw its first megawatt from a reactor that was supposed to be a museum piece.
It is tempting to frame this as a redemption story — nuclear power, unfairly maligned, finally getting its second chance thanks to the green imperatives of the AI age. But that is too tidy. The truth is messier. We are not reviving nuclear because we solved the waste problem, or because we have learned to build reactors cheaply, or because public fear has faded. Vogtle Units 3 and 4, the first new American reactors in decades, cost $36 billion — double the budget, seven years late. We are reviving nuclear because the alternative is worse. Because the AI industry has grown so fast, so voraciously, that it has outstripped the grid’s ability to feed it with anything else.
The machines are not patient. They do not wait for solar to get cheaper or for batteries to improve. They need power now, continuously, in staggering quantities. And the only thing that can give it to them is the same technology we spent forty years trying to leave behind.
There is a strange symmetry to it. In 1979, we built a machine we could not fully control, and it almost got away from us. In 2026, we are building machines that we increasingly do not fully understand, and they are pulling us back toward the same energy source. The fear has changed — from meltdowns to algorithmic bias, from radiation to hallucination — but the pattern is familiar. We build something powerful. We discover it needs more than we expected. We return to what we know, even if what we know carries old scars.
Three Mile Island Unit 1 will not know any of this, of course. It is steel and concrete, coolant and control rods. It does not remember 1979. It does not care about ChatGPT or data centers or the Crane Clean Energy Center’s carefully focus-grouped name. It will simply do what it was built to do: split atoms, boil water, spin turbines, push electrons into the grid. The same work it did for forty-five years, interrupted by a five-year nap, now recommissioned because the new gods are thirsty.
And the ghost next door? Unit 2, the melted one, will stay in its tomb. A silent neighbor. A permanent warning. A reminder that the same technology we are now embracing once taught us to fear it — and that we have apparently decided the fear was less urgent than the hunger.
Sources: E&E News, Energy.Media, Business20Channel, HowToStoreElectricity, EnergyNewsBeat, InvestmentGrade