Nuclear power for AI data centers: what the deals actually promise, and when

Two piles of megawatts

Since late 2024, the largest US technology companies have announced a run of nuclear power agreements tied to their artificial intelligence data centers. The headline numbers are large, and the press releases blur together. It helps to sort the deals into two piles. One pile is electricity that can reach the grid this decade because the reactors already exist. The other is electricity promised from reactors that have not been built, and in most cases not yet fully licensed. The distinction matters, because the two piles carry very different cost, risk, and timing.

The near-term power is mostly old reactors

The firmest commitments involve restarting or relicensing existing light-water plants, the same technology that has run for decades. In September 2024, Constellation Energy said it would restart Three Mile Island Unit 1, rebranded the Crane Clean Energy Center, under a 20-year power purchase agreement with Microsoft. The 835 megawatt reactor was shut for economic reasons in 2019, and Constellation now targets a 2027 restart, a year ahead of its original plan. In 2025 it secured a $1 billion Department of Energy loan to support the work.

Meta followed the same template. In June 2025 it signed a 20-year agreement for the entire output of Constellation's Clinton Clean Energy Center in Illinois, about 1,121 megawatts, with power flowing from June 2027. The deal keeps alive a plant that had faced closure, and adds roughly 30 megawatts through an uprate. Meta later agreed to take more than 2.1 gigawatts from Vistra's existing plants in Pennsylvania and Ohio.

These are real near-term megawatts. They are also, mostly, not new power. Restarting a retired reactor or relicensing an operating one keeps supply on the grid that might otherwise have left it. That is useful for a data center operator trying to match its consumption with carbon-free generation, but it is different from adding capacity the grid did not already have. It also leans on plants whose economics the grid had already judged marginal, which is why state subsidies or new corporate buyers were needed to keep several of them open.

The headline gigawatts come from reactors that do not exist

The second pile is where the large forward numbers live, and where the uncertainty concentrates. In October 2024 Google signed a master agreement with Kairos Power for a fleet of small modular reactors, targeting a first unit by 2030 and 500 megawatts by 2035. Amazon took an equity stake in X-energy and backed an Energy Northwest project in Washington for about 320 megawatts in the early 2030s, with an option to expand, plus an agreement to explore reactors with Dominion in Virginia. Amazon and X-energy have described a goal of more than 5 gigawatts by 2039.

Meta's request for proposals produced the biggest set of numbers. Its June 2025 agreements with Oklo, Vistra, and TerraPower were framed as unlocking up to 6.6 gigawatts by 2035. But a large share of that total depends on reactors still on the drawing board. TerraPower's two 345 megawatt Natrium units, Meta said, will not supply power until at least 2032, and Oklo's planned 1.2 gigawatt campus in Ohio is a development plan, not a running plant.

The pattern across these deals is a wide gap between a signed contract and a reactor feeding a server hall. A power purchase agreement is a commitment to buy electricity, not a finished plant, and the dates that matter sit years out.

Why the SMR timelines keep slipping

Small modular reactors are the technology these forward deals rely on, and they have a thin track record. Only a handful operate anywhere, in China and Russia, and as of early 2026 no SMR had received a construction license in the United States. Kairos has made visible progress: the Nuclear Regulatory Commission issued construction permits for its Hermes 2 demonstration plant in Oak Ridge in November 2024, and the company broke ground in 2026. Kairos has even adjusted the plant's plans, moving Hermes 2 to a 50 megawatt electric output after a supply agreement with the Tennessee Valley Authority linked to Google. But Hermes 2 is a demonstration unit, not a commercial fleet, and commercial operation is still a 2030 target.

The reasons for slippage are well documented. The Institute for Energy Economics and Financial Analysis argued in 2024 that SMRs remain too expensive, too slow, and too risky to count on. The clearest cautionary case is NuScale, whose Idaho project was canceled in 2023 after its estimated cost rose from about $5 billion to $9 billion. Advanced designs also need high-assay low-enriched uranium fuel, which has no commercial-scale US supply until at least the end of the decade, and novel reactor licensing has historically taken five to eight years.

What the contracts actually commit to

The demand pressure is real. The International Energy Agency estimated that data centers used about 415 terawatt hours of electricity in 2024, around 1.5 percent of the world total, and projected that to roughly double by 2030. The agency also noted that conditional offtake agreements between data center operators and SMR projects nearly doubled in a year, to about 45 gigawatts. Nuclear, the agency said, starts to play a larger role only toward the end of this decade and beyond. Conditional is the operative word.

Most of these announcements are power purchase or development agreements, not guarantees that a reactor will exist. They let a buyer commit to purchase output if and when a plant comes online, which helps the developer raise financing but transfers little construction risk onto the headline gigawatt figure. Read that way, the honest summary is narrow. The nuclear electricity AI data centers can count on before 2030 comes almost entirely from old light-water reactors being restarted or kept open. The small modular reactors that dominate the press releases are, for now, promises with long dates attached.