Nuclear Power for Data Centers: Old Reactors Lead as SMRs Lag Behind

The short version

Two years after Microsoft agreed to restart Three Mile Island, nuclear power for AI data centers has turned into a crowded market of deals. The deals do not all carry the same weight, though. Most of the electricity tied to data centers today comes from reactors built decades ago. The small modular reactors (SMRs) behind the biggest announcements will mostly arrive in the 2030s, if they arrive at all.

One tally from October puts the output that Amazon, Google, Meta and Microsoft have contracted from operating, restarting and uprated US reactors at about 8.5 GW. The International Energy Agency counts 45 GW of conditional agreements between data centers and SMR projects, up from 25 GW at the end of 202434. Those two numbers measure different things. The first is power that mostly exists already. The second is a list of options. Only one new reactor built for a data-center buyer is actually under construction: Kairos Power's 50 MW Hermes 2 in Oak Ridge, Tennessee34.

Why the buyers are in a hurry

The deals respond to fast-growing demand. Gartner expects global data-center electricity use to reach 565 TWh in 2026, up 26% from 447 TWh in 2025. It estimates that AI-optimized servers will account for 31% of that consumption this year and will use more power than conventional servers in 20271. TrendForce, which measures capacity rather than energy, puts global data-center demand at 161 GW in 2026, up about 31%. It expects grid supply to start falling behind this year, with a 268 GW gap by 2030 and more than 170 GW of that in the US8.

Forecasters disagree on the baseline. The IEA puts 2025 global data-center consumption at 485 TWh, Gartner at 447 TWh, and Deloitte at a much higher 681 TWh for all data centers3. In the US, Goldman Sachs projects demand rising from 31 GW in 2025 to 66 GW in 20279. Bloom Energy's estimate starts from a base of about 80 GW and reaches 150 GW by 20289. The numbers disagree, but the direction does not. Every major forecast shows demand outpacing what utilities can connect.

The grid itself is the binding constraint. Analysts report interconnection waits of four to ten years. Lead times for high-voltage transformers have stretched from roughly 24–30 months before 2020 to three to five years9. EPRI estimates data centers could use 9% to 17% of US electricity generation by 2030, and between 39% and 57% of Virginia's9. In the PJM market, the independent market monitor blamed roughly 12 GW of extra data-center demand in 2026/27 for a doubling of capacity costs3. Seen this way, nuclear is attractive less as a climate story than as firm, around-the-clock power that hyperscalers can lock in for 20 years.

What is actually being bought

The deals fall into three categories, and they should be kept apart.

Existing plants. Amazon's agreement with Talen for up to 1,920 MW from Pennsylvania's Susquehanna plant is the largest single nuclear contract for data centers. It runs through 2042 and moved to a front-of-the-meter arrangement in April 202621. Meta has a 20-year deal for 1,121 MW from Constellation's Clinton plant in Illinois, with deliveries starting in June 202734. In January, Meta added 2,176 MW from Vistra's Perry and Davis-Besse plants, plus 433 MW of uprates21. Amazon has 1,200 MW from Vistra's Comanche Peak in Texas starting in late 2027. On September 30 it added a 690 MW Calvert Cliffs deal with Constellation, including a 190 MW uprate21. Google has signed for up to half of the output of Fortum's Loviisa plant in Finland from 2030 to 2049, which is part of a broader push into Europe2125.

Restarts. Microsoft's 20-year deal covers all 835 MW of Three Mile Island Unit 1, now called the Crane Clean Energy Center33. In June, FERC approved a transmission waiver that moves grid-connection rights to the site, which removed a major obstacle33. Constellation says the restart is on track for 202734. Some coverage still cites 2028 for full service, so the exact date is uncertain13. Google is backing the restart of NextEra's Duane Arnold plant in Iowa. A $1.9 billion Department of Energy loan closed in September, and the target is the first quarter of 202921.

New reactors. Google's 2024 deal with Kairos covers up to 500 MW from a fleet of reactors cooled by molten fluoride salt. The first unit is targeted for 2030 and the rest by 203517. Amazon wants up to 5 GW of X-energy's Xe-100 reactors by 2039, starting with a 320 MW, four-unit plant in Washington state that could grow to 960 MW11. Meta is funding two of TerraPower's 345 MW Natrium units, which could deliver as early as 2032, and has rights to six more by 2035. Meta's deal with Oklo supports a campus of up to 1.2 GW in Pike County, Ohio, that could start delivering power in 203031.

Where the numbers diverge

The headline totals depend on who is counting. A widely quoted tracker says 13 projects add up to more than 9.8 GW. It adds firm contracts with existing plants to framework agreements for reactors that have not been designed for specific sites yet22. The October tally avoids that by keeping existing-plant output separate from new-build plans. It also lists Meta's TerraPower agreement as a framework with no site announced34.

Meta's 6.6 GW figure shows the problem. Meta says its combined agreements will "support" up to that much current and new generation by 203531. Some aggregators report it as committed capacity24. Even the parts are described inconsistently. One account splits the TerraPower package into 2.8 GW of reactor capacity plus 1.2 GW of molten-salt storage33. TechCrunch says the storage adds 100 to 500 MW per reactor for more than five hours35. Amazon's X-energy target has the same issue. X-energy's own April 2026 filing calls the 5 GW figure a target that assumes Amazon exercises every option, and says no final investment decision has been made on any reactor11.

The takeaway: firm contracted nuclear capacity for data centers is close to the 8.5 GW existing-plant figure. Most numbers above that are options.

The additionality question

A quieter part of the story is whether these deals add anything to the grid. A contract with an operating reactor mostly moves existing output from one buyer to another. It adds capacity only when it pays for an uprate, a restart or a license extension that would not otherwise happen. That is why recent contracts increasingly include uprates or relicensing34. Vistra's chief executive said the extensions and uprates covered by the Meta deal would not have been financially feasible without Meta's money, and all three Vistra plants had been slated for closure a few years ago.

This matters for grid capacity. A data center that takes a reactor's existing output still adds load, and other customers have to be served from other sources. Deals that fund uprates and restarts are the ones that actually expand supply. Analytically, the move toward uprate-heavy contracts is the most constructive trend in this market.

The SMR reality check

The SMR timeline is still mostly a set of targets. As of this month, none of the US designs in hyperscaler deals is producing power. The earliest first-power dates cluster around 2030, and the only SMRs operating anywhere are in China and Russia11. The regulatory record is moving: the NRC has approved construction at TerraPower's Kemmerer site, Kairos's Hermes reactors and TVA's Clinch River project11. Even so, X-energy and Energy Northwest had not filed a construction permit application in the materials reviewed11.

Cost and fuel are the weakest parts. Announcements for Kairos, X-energy, TerraPower and Oklo include no firm construction budgets11. TerraPower targets $50–60 per megawatt-hour and Oklo targets $80–130, but those figures apply to later plants, and the first units will probably cost more35. Several leading designs, including Natrium, the Xe-100 and Oklo's Aurora, need high-assay low-enriched uranium (HALEU), a fuel whose supply outside Russia and China is still being built up12. X-energy's filing warns that a shortfall would materially harm its business11.

The reading

Restarts and long contracts with existing plants are the only nuclear power that can reach data centers within this decade, and new existing-plant deals can start delivering within about two years of signing34. That is why Microsoft, which picked a restart, gets nuclear power first, while Meta's larger portfolio mostly lands between 2032 and 203533.

The SMR deals are better read as venture-style bets on industrial learning. Hyperscalers are prepaying, funding development and buying options so that startups can raise money and work down a cost curve. Meta has said this directly, arguing the deals give Oklo and TerraPower more business certainty to finance their projects31. That is a reasonable bet, but it does not ease the grid crunch forecasters expect around 20288.

The next year has clear checkpoints: Crane's return to service, the first deliveries under Meta's Clinton deal and Amazon's Comanche Peak deal in 2027, site announcements for Meta's Natrium units, and NRC review of Blue Energy's permit application for a Crusoe campus in Texas34. Each one that slips will add to the pressure to bridge the gap with gas, which some nuclear-for-AI projects, such as the Crusoe–Blue Energy plan, already include in their own plans21.