Why Is AI Restarting Nuclear Reactors?

On an island in the Susquehanna River stands the most famous failed machine in America. Three Mile Island Unit 2 melted down in 1979 and ended nuclear power's first age in the United States. Its twin, Unit 1, ran quietly for four more decades before shutting down in 2019 — uneconomical, unwanted. That should have been the end of the story.

It wasn't. In September 2024, Constellation Energy announced it would bring Unit 1 back from retirement — the first restart of a shuttered American nuclear plant in history — with every megawatt for twenty years already sold to a single customer: Microsoft, to power artificial intelligence [3]. The revival is well underway. A federal loan guarantee of one billion dollars was secured this spring; a regulatory waiver cleared one of the last grid hurdles on the first of June; the target date has moved a year earlier, to 2027 [4][5]. Eight hundred thirty-five megawatts — enough for roughly 800,000 homes — reserved for software.

This is the strangest sentence in the modern energy business, so it is worth saying plainly: a chatbot is un-retiring a nuclear reactor. The question this investigation pursues is why — and what it reveals about a collision that will shape the next decade: the world's fastest-moving technology has run into its slowest-moving infrastructure.

Every technology that changed civilization eventually presented an electricity bill — or a coal bill. In 1865, the economist William Stanley Jevons studied Britain's steam engines and found a paradox that now carries his name: as engines became more efficient, Britain burned more coal, not less, because efficiency made steam power cheap enough to use everywhere [1]. Efficiency, he argued, is not a brake on consumption. It is often the accelerator.

A century later the pattern repeated with electricity itself. The economic historian Paul David showed that factory electrification took roughly forty years to transform productivity — not because the dynamo was weak, but because factories, cities, and grids had to be physically rebuilt around it [2]. The lesson from both episodes: information technologies move at the speed of software; their consequences move at the speed of concrete.

Then, for one remarkable decade, computing seemed to escape the pattern. Between 2010 and 2018, the world's data-center computing output grew more than five-fold — while data-center electricity use stayed nearly flat, because efficiency gains kept pace with demand [6]. That peer-reviewed finding quietly underwrote a belief the industry internalized: digital growth was energetically free. So what? The entire surprise of the present moment — the scramble for reactors, the strained grids — is the sound of that decade ending. Jevons is back.

The International Energy Agency's first full analysis of AI and energy puts measured numbers under the anecdotes. Data centers consumed roughly 415 terawatt-hours in 2024 — about 1.5 percent of global electricity. In the IEA's base case that roughly doubles by 2030, to around 945 TWh — near 3 percent of the world total, with the United States and China accounting for nearly 80 percent of the growth [7]. In the United States, data centers are projected to drive almost half of all electricity demand growth between now and 2030 [7] — this in a country where demand had been essentially flat for nearly two decades.

Two honest context notes belong beside those numbers. First, even doubled, data centers remain a few percent of global electricity — the strain is not planetary but intensely local: the grid regions and interconnection queues where gigawatt campuses cluster [7]. Second, the corporate response is now measurable rather than rhetorical: the Crane restart is contracted, financed, and regulator-approved [3][4][5], and it is the leading edge of a broader turn by computing companies toward buying, restarting, and commissioning firm power. The significant fact is not the percentage. It is the reversal of roles: technology companies — for decades pure consumers of infrastructure — have started acting like energy developers.

Competing hypothesis How big this actually gets is where qualified analysts diverge. Demand forecasts span a wide range, and the field carries a humbling memory: in the early 2000s, widely repeated claims that the internet would devour half of US electricity proved wrong by an order of magnitude, undone by exactly the efficiency gains later measured in the flat decade [6]. One camp reads today's forecasts the same way — model efficiency is improving rapidly, and cheaper computation per task could bend the curve again. The other camp answers with Jevons [1]: every past efficiency gain in computing was converted into more computing, and contracted power purchases suggest the buyers themselves expect demand, not moderation. The IEA holds the middle: its scenarios span a wide range, and it says so [7].

Competing hypothesis Whether AI load accelerates clean energy (by financing nuclear restarts and new firm capacity, as at Crane) or delays it (by keeping fossil plants running and pushing utilities toward new gas) is genuinely unresolved — early evidence exists on both sides, and the answer likely differs by region. So is the question of whether data-center growth raises electricity rates for households near the clusters; regulators in several states are litigating exactly that now. These disputes share a structure: the technology's benefits are global and diffuse, its energy footprint is local and concentrated. That mismatch — not the total terawatt-hours — is where the political friction of the Age of AI will be felt first.

This is not purely an energy story — it is what happens when three systems that used to move independently start moving together. Capital markets now underwrite reactor restarts the way they once underwrote data centers alone: the billion-dollar federal loan guarantee and the twenty-year Microsoft power contract are financial instruments as much as energy ones, and they only exist because a software company's demand forecast was solid enough to de-risk a decade-long infrastructure bet [3][4]. Regulatory systems are the second constraint: the FERC waiver that cleared a grid hurdle this spring shows permitting speed, not physics, is now a binding limit on how fast compute can get power [5]. And the same demographic pressure that is straining the nuclear workforce more broadly — an aging cohort of specialists with few successors — sits underneath every restart timeline like Crane's, since a reactor cannot come back online faster than the licensed staff needed to run it can be assembled.

A reactor takes a decade to plan and two years to restart; a data center takes eighteen months to build. That mismatch alone produces more than one plausible ending. None of what follows is a forecast — each path traces where the evidence already points, if it keeps pointing that way.

Intelligence turned out to have a body: thinking, done at scale, is now an industrial process with an electricity bill, a water bill, and a construction schedule — which is why the following indicators are as concrete as any in energy policy.

• Reactor & regulatory milestones — Crane Clean Energy Center milestones: fuel load, NRC sign-offs, staffing — does the 2027 restart date hold or slip? [4][5] Whether Crane is actually delivering power, and whether any small modular reactor ordered for data centers breaks ground — the test of the anchor-tenant thesis.
• Corporate power contracts — New hyperscaler power contracts: firm clean power, or gas? Whether new deals keep favoring firm clean capacity over gas as the buildout continues.
• Measured demand versus forecast — Reported US data-center consumption versus the IEA trajectory [7] — the first check on whether 945 TWh by 2030 is on path. Whether model-efficiency gains show up as slower demand growth (the 2010s pattern [6]) or more consumption (the Jevons pattern [1]).
• Regulatory and rate outcomes — State regulatory rulings on who pays for data-center grid upgrades.
• Longer-run trajectory — The IEA's 2030 base case tested against measured reality [7]. The carbon intensity of AI load: did the buildout leave behind a cleaner grid or a new gas fleet? US electricity demand: structural growth era, or another plateau?

A data center may be coming to a grid near you — most readers are within one utility territory of a proposed campus. If it does: what would you want your regulator to demand in exchange — and what evidence would convince you the deal was good for the people who live there? Tell us. We read every message.