The AI Boom Just Dragged Three Mile Island Back Into Politics
Three Mile Island is back in the political conversation for a reason that would have sounded bizarre a decade ago: artificial intelligence needs power.
Not metaphorical power. Not narrative power. Actual electricity.
Constellation Energy is working to restart Unit 1 at the Three Mile Island site in Pennsylvania, now renamed the Crane Clean Energy Center. Microsoft has agreed to buy the plant’s output under a 20-year power purchase agreement to support its cloud and AI data center needs. The unit being restarted is separate from Unit 2, the reactor involved in the 1979 partial meltdown, but the symbolism is impossible to miss. The most loaded nuclear address in America is being repositioned as part of the AI infrastructure stack.
That tells you where the next phase of the AI boom is headed. The first phase was software. The second phase was chips. The third phase is going to look like a utility hearing.
AI companies can announce models instantly. They can raise capital quickly. They can buy GPUs in huge waves if they can get supply. But they cannot summon reliable baseload electricity out of a press release. They need land, transmission, substations, cooling, permits, interconnection queues, financing, and public tolerance. The hard constraint is moving from the screen back into the physical world.
That is healthy in one sense. It punctures some of the fantasy around AI. This is not magic floating above society. It is an industrial buildout. It consumes power, water, steel, concrete, chips, land, and political attention. If the technology is going to be as important as its boosters claim, then it has to be judged like other major infrastructure shifts: who pays, who benefits, who absorbs the risk, and who gets a say.
The Three Mile Island deal is a useful test case because it strips away the abstractions. A hyperscaler wants dependable electricity for AI. A nuclear operator has an old asset that was uneconomic in the previous market structure. A community has memories, legitimate concerns, and a need for jobs and tax base. Regulators have to decide whether safety, reliability, and economics line up. Politicians get to posture as either pro-innovation or pro-safety. And ordinary citizens are left trying to figure out whether this helps their grid or just reserves scarce reliability for the richest corporate buyers.
That last question is the one that matters most.
The public sales pitch for AI infrastructure is usually national competitiveness. America needs to beat China. America needs more data centers. America needs more energy. America needs to move faster. There is truth in that argument, but it is incomplete. Whenever the country is told that a private buildout is a national imperative, citizens should immediately ask how the bill is being allocated.
If Microsoft signs a long-term power deal that helps restart a nuclear plant, that may be good for reliability. It may preserve or create skilled jobs. It may bring a serious zero-carbon baseload asset back onto the grid. It may be better than forcing more gas plants onto the system to satisfy data-center demand. Those are real arguments.
But there is another possibility. The largest technology companies may be able to lock up reliable power on terms that ordinary households, small businesses, and local governments could never get. They can sign 20-year deals. They can finance around bottlenecks. They can hire lawyers and consultants to move projects through the maze. They can turn electricity into a strategic moat.
Meanwhile, the public is told to accept higher rates, slower interconnections, more transmission fights, and greater strain on the grid because the future requires it.
That is where the politics begins.
The AI boom has been described as a productivity revolution, but the infrastructure layer has distributional consequences. If data centers increase local demand for electricity, somebody has to build capacity. Somebody has to upgrade transmission. Somebody has to manage congestion. Somebody has to carry financing risk. Somebody has to explain why residential bills are rising even as corporations are promising a glorious automated future.
The country should not be reflexively anti-data-center or anti-nuclear. That would be lazy. The United States needs more serious infrastructure, not less. It needs a grid that can handle electrification, manufacturing, AI, and national-security requirements. It needs nuclear power in the conversation if the goal is reliable, low-carbon electricity at scale. Pretending we can run an advanced industrial economy on slogans is not a plan.
But being pro-infrastructure does not mean giving every politically connected project a blank check. It means demanding a better bargain.
If AI companies need enormous amounts of power, then the public should expect enormous transparency. How much capacity is being reserved? What happens to local rates? Who pays for transmission upgrades? Are costs being socialized while benefits are privatized? Are utilities protecting ordinary customers from data-center-driven load growth, or are they quietly turning households into the backstop for hyperscaler expansion? What safety obligations and emergency planning standards apply when a politically symbolic nuclear site comes back online? What happens if AI demand projections prove exaggerated after billions are committed?
Those are not anti-technology questions. They are pro-accountability questions.
The most revealing part of this story is that AI is forcing the technology sector to negotiate with reality. For years, Silicon Valley’s preferred story was that software eats the world. Now software is eating so much electricity that it has to make peace with nuclear plants, utilities, state regulators, local zoning boards, and public memory.
That is a different political economy than the app era.
Apps could scale globally while appearing local and weightless. AI infrastructure cannot. It lands somewhere. It plugs into a grid. It competes for power. It changes the economics of generation assets. It creates winners and losers before the first chatbot answers a question. The politics will not be confined to Washington hearings about model safety. It will show up in state utility commissions, municipal tax deals, water permits, rate cases, and energy-market rules.
Three Mile Island also highlights the trust problem. The restarted unit is not the reactor that melted down in 1979. That distinction matters and should be stated clearly. But public trust is not rebuilt by technical distinctions alone. Communities remember institutions that told them risks were under control. They remember experts who spoke confidently and later had to revise the story. They remember being asked to accept costs for benefits that seemed to flow elsewhere.
If the AI industry wants nuclear power as part of its growth plan, it should understand that it is not just buying electrons. It is buying into decades of institutional memory. The public will not judge these projects only by engineering diagrams. It will judge them by whether officials are honest about risk, whether emergency planning is credible, whether economic benefits are local as well as corporate, and whether ratepayers are protected.
The worst possible version of the AI buildout would combine technological hype with old-fashioned political insulation: private companies capturing reliable power, utilities shifting costs into opaque rate structures, politicians celebrating “innovation,” and citizens discovering the tradeoffs only after the bills arrive.
The better version is possible, but it requires discipline. If restarting nuclear assets is the least-bad way to power an AI-heavy economy, then say that plainly. If data centers are going to drive load growth, make the cost allocation visible. If corporate power purchase agreements help finance reliable generation, show whether they reduce pressure on ordinary customers or simply let the biggest buyers jump the line. If communities are being asked to host the infrastructure of the future, they deserve more than a ribbon-cutting and a promise that everything will work out.
This is why the Three Mile Island story matters beyond one plant in Pennsylvania. It is a preview of the next argument.
The public has been told AI will transform medicine, education, logistics, finance, defense, software, media, and work itself. Fine. Then the public should also be told what that transformation requires physically. The servers have to sit somewhere. The power has to come from somewhere. The grid upgrades have to be financed somehow. The political risks have to be owned by someone.
The accountability test is simple: does the AI infrastructure boom leave citizens with a stronger, cheaper, more reliable energy system, or does it leave them underwriting a private arms race among hyperscalers?
That is the question politicians would rather avoid, because it cuts across the normal partisan script. One side wants to celebrate innovation. The other side wants to denounce corporate power. Both can miss the boring machinery where the real deal is made: rate design, permitting, tax incentives, interconnection rules, utility planning, and long-term contracts.
But that boring machinery is exactly where the public interest either survives or disappears.
Three Mile Island coming back as an AI power asset is not just a nuclear story. It is not just a Microsoft story. It is not even just an AI story. It is a reminder that every supposedly digital revolution eventually runs into physical scarcity and political allocation.
The machine needs electricity. The electricity needs infrastructure. The infrastructure needs public consent. And public consent requires a bargain citizens can actually see.
If the AI boom is strong enough to drag Three Mile Island back into American politics, it is strong enough to answer basic questions about cost, risk, and control.