Nuclear Energy Makes a Comeback to Power the AI Boom

Somewhere in Pennsylvania, a reactor that shut down in 2019 because it could not compete on price is being rebuilt from the ground up, and the company footing much of the bill is not a utility. It is Microsoft. That single fact tells you almost everything about what happened to the nuclear power industry in 2026. After four decades on the sidelines, nuclear energy is being pulled back into the center of the grid, not by climate policy or public opinion, but by the sheer electricity appetite of artificial intelligence.

This is not a niche energy story anymore. It touches how fast new AI models can be trained, where the next data center campus gets built, and how much your electricity bill might creep up over the next few years. Here is what is actually driving big tech toward nuclear power, who is signing the checks, and where the plan still has real gaps.

The problem nobody saw coming this fast

AI did not just add a new source of electricity demand, it accelerated one that grid planners assumed they had another decade to prepare for. Global data centre electricity consumption reached roughly 485 terawatt hours in 2025 and is on track to approach 950 terawatt hours by 2030, according to the International Energy Agency, which also found that electricity use from AI focused facilities specifically surged 50 percent in 2025 alone, even faster than data centre demand overall. In the United States, that growth is colliding with a grid that was not built for it. Goldman Sachs Research now flags a structural power shortfall of roughly 9.3 gigawatts in 2026, a gap it expects to widen to around 45 gigawatts by 2028, while interconnection queues, the waiting line new power projects join before they can plug into the grid, now stretch to roughly five years in many US markets.

The strain already shows up in specific places. Data centres consumed about 26 percent of all electricity in Virginia and up to a third of Ireland's national supply, with the International Energy Agency projecting that concentration will only deepen through the rest of the decade as more capacity comes online. Renewables and gas alone cannot fill that gap on the timeline hyperscalers want, which is exactly the opening small modular reactors and restarted plants have stepped into.

Big tech goes nuclear

By mid 2026, every major AI hyperscaler, Microsoft, Google, Amazon and Meta, had signed at least one direct nuclear power agreement, and the combined total across roughly 13 announced projects now tops 9.8 gigawatts of committed capacity. That is a genuinely new posture. These companies are no longer simply buying electricity off the grid, in several cases they are directly financing the power plants themselves, something traditional utilities had mostly stopped doing for new nuclear projects over the past several decades.

Microsoft brings a retired reactor back online

The most visible deal belongs to Microsoft and Constellation Energy. In 2024 the pair signed a 20 year power purchase agreement covering the full output of Three Mile Island Unit 1, a reactor retired in 2019 for economic reasons and now renamed the Crane Clean Energy Center. The project carries a price tag near $1.6 billion, backed in part by a $1 billion federal loan from the Department of Energy, and Constellation has since moved the target restart date up a year to 2027. Once running, the plant's roughly 835 megawatts, enough for hundreds of thousands of homes, will flow entirely to Microsoft's data centres under the agreement, and Pennsylvania officials project the restart alone will add around $16 billion to the state's economy.

Meta spreads its bets across almost every reactor design

Where Microsoft picked one restart, Meta went wide. In January 2026 the company announced agreements with Vistra, TerraPower, Oklo and Constellation covering up to 6.6 gigawatts of nuclear capacity, a combined commitment large enough to power roughly five million homes and one of the largest corporate nuclear purchases in American history. The mix spans both a familiar existing reactor fleet and next generation small modular reactors still working toward commercial deployment, including TerraPower's Natrium design and Oklo's Aurora units planned for a technology campus in Pike County, Ohio.

Google and Amazon chase the small reactor future

Google and Amazon have leaned harder into the newest reactor technology rather than existing plants. Google committed to 500 megawatts of capacity from Kairos Power's molten salt cooled reactor design, while Amazon put $700 million directly into X-energy to help fund development of its Xe-100 high temperature gas reactors, alongside a separate multibillion dollar deal tied to its Susquehanna data centre campus in Pennsylvania. Both bets assume small modular reactors can eventually deliver nuclear power in smaller, faster to build increments than a traditional plant, a proposition still mostly unproven at commercial scale.

Washington clears the runway

None of this would move nearly as fast without a deliberate policy push behind it. The Department of Energy's Reactor Pilot Program, launched in 2025, selected 11 advanced reactor projects and set a target for at least three of them to reach criticality, the point where a reactor sustains a controlled nuclear reaction, by July 4 2026. According to the Department of Energy, four designs hit that milestone, more new reactors reaching criticality in a single year than in the prior six decades combined, though reaching criticality in a pilot program is still a long way from a licensed commercial power plant.

Regulators moved too. In March 2026 the Nuclear Regulatory Commission finalized what is known as Part 53, its first entirely new reactor licensing framework since 1989, built around risk informed, technology neutral rules rather than requirements written specifically for traditional light water reactors. The Nuclear Regulatory Commission describes the goal as making advanced reactor licensing faster and simpler without loosening safety standards, and a companion proposal would let developers lean on prior Energy Department testing to speed up their eventual NRC applications. The push extended beyond US borders too, with the United States and Canada each announcing plans in June 2026 to build ten new reactors, among the largest coordinated nuclear commitments North America has made in decades.

The fine print nobody is advertising

For all the momentum, the honest caveats matter just as much as the headline gigawatt numbers. Most independent energy analysts still expect small modular reactors to only begin operating at meaningful scale in the early 2030s, and first of a kind projects tend to run expensive and behind schedule, a pattern the nuclear industry has repeated more than once over the decades. The Three Mile Island restart carries its own history to manage as well, since the site is still associated in public memory with the 1979 partial meltdown at a separate unit, even though the reactor being restarted was not involved in that accident, and the project has drawn real community pushback alongside its local support.

There is also a harder structural question underneath all of it. Betting on nuclear power to solve AI's electricity crunch assumes these reactors arrive roughly on the ambitious timelines companies have announced, and history suggests that assumption deserves some skepticism even as the grid interconnection queues driving hyperscalers toward direct power deals show no sign of shrinking on their own.

So does nuclear actually solve AI's power problem

Here is the realistic picture. Nuclear power is not going to single handedly meet AI's electricity demand by 2030, gas and renewables will still carry most of the load through the rest of the decade even in the International Energy Agency's own projections. What has genuinely changed is that nuclear energy has gone from a technology most corporate boardrooms treated as too slow and too expensive, to one of the few sources that can promise the tech sector round the clock, carbon free power at the scale AI training clusters actually need.

That shift explains why Microsoft is willing to help resurrect a reactor that once could not turn a profit, and why Meta, Google and Amazon are all placing early, expensive bets on reactor designs that have barely left the drawing board. Whether those bets pay off on schedule is still an open question. But the fact that the biggest companies in technology are now acting more like power utilities than software firms tells you how seriously they are taking the risk of simply running out of electricity.