The companies racing to build artificial intelligence have quietly become some of the most aggressive buyers of nuclear electricity on the planet. Over roughly the past 18 months, Microsoft, Meta, Amazon, and Google have collectively committed to more than 9.8 gigawatts of nuclear capacity, a figure large enough to power millions of homes and a signal that the AI boom now runs on atoms as much as algorithms. The scramble has revived aging reactors, resurrected a plant famous for the wrong reasons, and turned a technology many had written off into a strategic asset.
Why AI turned into an electricity problem
The trigger is raw demand. Training and running large AI models requires dense clusters of chips that draw enormous, around-the-clock power, and the data centers housing them cannot tolerate the intermittency of wind or solar without expensive backup. Global data-center electricity consumption reached hundreds of terawatt-hours in 2025 and is forecast to keep climbing steeply, leaving hyperscalers hunting for firm, carbon-free supply. Nuclear fits that profile almost perfectly, delivering steady baseload output regardless of weather or time of day, and its appeal to lenders and utilities has grown as the contracts pile up. That shift in financing psychology, in which long-term corporate offtake makes reactors bankable in a way they had not been for decades, is at the center of how the AI build-out is rehabilitating nuclear economics.
The single largest pledge belongs to Meta
Among the four, Meta has staked out the biggest position. In early 2026 the company announced arrangements to secure up to 6.6 gigawatts of nuclear power over the coming decade, the largest single commitment by any of the hyperscalers to date. The deals involve output from existing plants, including more than two gigawatts tied to Vistra’s Beaver Valley station in Pennsylvania and its Perry and Davis-Besse plants in Ohio, alongside support for new capacity further out. The scale of the pledge, and its structure as a mix of relicensing support and long-term purchases, was detailed when Meta signed multi-gigawatt nuclear agreements aimed squarely at its AI data centers. The arrangement underscores a broader tactic: rather than wait years for new reactors, the tech giants are paying to keep existing ones running and to extend their operating lives.
Restarting reactors that were already switched off
The most symbolically loaded piece of the story is Microsoft’s involvement in restarting Three Mile Island, the Pennsylvania site of the 1979 partial meltdown that helped freeze American nuclear construction for a generation. The undamaged Unit 1 reactor there had been shut for economic reasons, and under a 20-year power agreement the plant is being brought back to feed Microsoft’s growing electricity needs, with the company effectively contracting for the full output once the unit returns to service. Amazon has taken a different but equally striking route, investing heavily to expand a data-center campus tied to the Susquehanna nuclear plant, positioning reactors as the anchor tenant’s dedicated power source. The pattern of these individual arrangements, from plant restarts to co-located campuses, is catalogued in running tallies of the nuclear-powered data-center deals signed by Google, Amazon, Meta, and Microsoft, which show how quickly a handful of firms have reshaped the market.
Betting on reactors that do not exist yet
Not every commitment involves existing steel and concrete. A large share of the announced capacity is pinned to small modular reactors, or SMRs, a class of compact designs that promise faster construction and factory-built components but that remain, for the most part, unbuilt in the West. Google has partnered with Kairos Power on what has been described as the first corporate deal to develop a fleet of SMRs for the U.S. grid, with meaningful output not expected until the end of the decade at the earliest. That timing gap is the central caveat behind the headline number. Much of the 9.8 gigawatts is a commitment for the 2030s, not power flowing today, and nuclear projects have a long history of delays and cost overruns that could erode the promised totals. The full sweep of the trend, including the mix of near-term restarts and longer-dated SMR bets, is laid out in industry accounting of the year nuclear power reclaimed relevance amid AI demand.
What the buying spree means for everyone else
The consequences reach well beyond corporate balance sheets. When a few enormous buyers lock up firm, carbon-free generation on multi-decade contracts, they change the calculus for utilities, regulators, and ordinary ratepayers who share the same grid. Supporters argue the AI companies are underwriting exactly the kind of clean baseload that decarbonization plans have struggled to finance, effectively acting as anchor customers that make new nuclear viable again. Skeptics counter that much of the announced capacity is aspirational, that SMRs remain unproven at commercial scale, and that concentrating so much power demand in data centers could strain grids and push costs onto communities that never signed up for an AI arms race. Both readings can be true at once. The 9.8-gigawatt figure is real as a sum of commitments, but whether it materializes as delivered electricity, and who ultimately pays for it, will be decided over the next decade rather than announced in a press release.
This article was produced with the assistance of AI and reviewed by the Morning Overview editorial team.
More from Morning Overview
- A California supervolcano has bulged upward about two and a half feet since 1978
- The FTC is warning about a scam quietly draining thousands from victims
- The NSA is again telling phone owners to switch off one location setting
- A handful of car transmissions are so tough mechanics say they almost never fail
