Nuclear-Powered Data Centers Lean on Old Reactors as SMRs Lag

The power bottleneck has become the AI bottleneck

The nuclear industry's trade press now reads partly like data center coverage. Over the past year, Nuclear Newswire, the news service of the American Nuclear Society, has run a steady flow of AI-related items. They include reactor startups building toward on-site computing, Department of Energy land being offered for data center campuses, and European governments arguing over whether nuclear power counts as "sustainable" for AI. Taken together, they point to one conclusion. AI is no longer one demand driver among many for nuclear power. It is the main commercial argument for new reactors. Our reading, though, is that existing reactors will supply most of the power data centers actually receive this decade. Small modular reactors (SMRs) remain largely a bet for the 2030s.

Demand is growing fast. Argonne National Laboratory's Nicolas Stauff, writing for the ANS, cites projections that data centers will grow from about 4 percent to 9 percent of U.S. electricity use by 2030. He says this growth is already changing utility planning and interconnection queues, the waiting lists for new grid connections.1 Other estimates disagree on the exact figures but agree on the direction. One tally puts global data center use at 485 TWh in 2025, citing IEA data, and expects it to roughly double by 2030. Gartner starts from a lower 2025 baseline of 447 TWh.15 Estimates of AI-specific load vary much more. The Uptime Institute expects global AI power load to reach about 10 GW by the end of 2026, while Deloitte projects roughly 96 GW.15 Numbers that far apart show how uncertain the forecasts are. Still, no major forecaster expects demand to flatten.

The grid, not generation, is the immediate constraint

Across the coverage, the scarce resource is often grid delivery rather than generating capacity. Of roughly 140 large U.S. data center projects tracked in one analysis, only about a third, about 12 GW, are actually under construction. The rest are waiting on interconnection, transmission upgrades or permits.15 Gartner expects power shortages to limit 40 percent of AI data centers by 2027.11 Sightline Climate reports that about half of global projects face delays tied to power limits and shortages of grid equipment.15

The strain shows up in reliability warnings and in prices. In May 2026, NERC, the body that sets grid reliability standards, issued its highest-urgency Level 3 alert after gigawatt-scale data center loads dropped off the grid within seconds. Its long-term assessment expects North American summer peak demand to grow by more than 224 GW over the next decade.19 In the PJM region, the largest U.S. grid operator, the independent market monitor blamed roughly 40 percent of record capacity costs on data center demand.19 The same pressure is visible outside the U.S. The UN Economic Commission for Europe warned in September that data centers are being built faster than the grids that must supply them. It also flagged that sudden swings in AI load are hard for renewable-heavy systems to absorb.16

This is why nuclear power appeals to data center builders. Stauff argues it offers round-the-clock output, stable long-term costs and less exposure to transmission limits. He describes three ways to buy it: grid-connected power purchase agreements, "behind-the-meter" supply from a plant to an adjacent campus that keeps a grid backup, and fully islanded, off-grid setups.1 He also notes that existing nuclear sites and retired coal plants may offer land, grid connections and water that new sites lack.1 Deloitte points to a cheaper, non-nuclear lever as well. Its research cites a study finding that if data centers cut just 1 percent of their annual load at peak times, grids could connect about 126 GW of new load with little expansion.20

Restarts and uprates deliver first

The deal trackers largely agree on one point: the oldest assets deliver soonest. Microsoft's 20-year agreement for the restarted Three Mile Island Unit 1, now called the Crane Clean Energy Center, got a FERC transmission waiver in June 2026. One analysis says that moved delivery up to the second half of 2027.22 Amazon's 1.9 GW agreement with Talen Energy's Susquehanna plant runs through 2042.22 Meta's January deal with Vistra covers more than 2.1 GW from existing Ohio plants plus 433 MW of uprates, meaning output increases at existing reactors.24 Duane Arnold in Iowa is another candidate. Local officials there have approved the rezoning needed to support NextEra's restart effort.10

Stauff's view from Argonne fits this pattern. He names uprates, license extensions, contracts with the current fleet and restarts as the near-term sources of power. New large reactors, SMRs and microreactors come later, in the 2030s.1 Stauff's ANS essay and the independent trackers reach the same conclusion: restarts are years ahead of any new-build SMR.122

The SMR pipeline: big headlines, softer contracts

The coverage diverges most on how firm the new-reactor commitments are. Headlines describe Meta as having "secured" 6.6 GW of nuclear energy. That total includes two TerraPower Natrium units expected by 2032 and rights to six more by 2035.27 One analysis that read the deal structures more closely found that Meta's Oklo agreement is a prepayment, with no power purchase agreement described. It also found that the TerraPower agreement is development funding plus rights to energy.24 The same analysis counts only one power purchase agreement with a new advanced reactor: up to 50 MW from Kairos sold through the Tennessee Valley Authority (TVA).24 Industry trackers add up big-tech nuclear commitments at more than 10 GW.25 That figure is better read as options on future capacity than as power that has been bought.

Licensing status supports that caution. As of early October, the only SMR-class designs with permission to build were GE Vernova Hitachi's BWRX-300 in Canada and two U.S. designs: TerraPower's Natrium and Kairos's Hermes 2. TerraPower expects Natrium to be completed around 2030.24 Oklo has no Nuclear Regulatory Commission (NRC) application pending. X-energy's Long Mott project has no permit yet, and the final safety evaluation is targeted for November 2026.24 Cost is another open question. First-of-a-kind SMRs are estimated at $4,000 to more than $12,000 per kilowatt. Factory-style savings are expected only after six to ten units of a single design are built.22

There is real progress at the small end of the market. Aalo Atomics reached criticality, a sustained chain reaction, on July 4 with a zero-power test version of its 10-MWe reactor at Idaho National Laboratory. It was the fourth reactor authorized by the DOE to reach that milestone. Aalo says it plans a commercial-scale unit to power an on-site data center in 2027.8 The University of Utah plans to generate electricity from its TRIGA research reactor for the first time to run a "mini AI data center."4 Terrestrial Energy and Riot Platforms are exploring co-located data center and molten-salt reactor projects in Texas and Kentucky.2 Federal land is also coming into play. The DOE has announced partnerships to build AI data centers combined with energy infrastructure at the Paducah and Savannah River sites.4

Project Beehive and the limits of ambition

Valar Atomics' proposal in Utah shows both the scale of current ambitions and the friction they create. NPR reported that the company wants to place 456 small reactors on more than 9,000 acres of Bureau of Land Management land near Price. The site would also include data centers, fuel production and waste storage.23 Nuclear Newswire pointed out problems in the reporting. The project's reported total of 9.6 GW implies at least 21 MWe per reactor, higher than any rating previously given for Valar's design. It also noted that the DOE has described its reactor authorizations as meant for research and test reactors, a description that hardly fits a 456-unit commercial campus.5 Local acceptance is another hurdle. Canary Media reports pushback in Carbon County, where a large gas plant is also planned to serve data centers.26 Stauff identifies local acceptance as a basic requirement for any nuclear-powered campus.1

What it means

The accurate reading is two-tiered. Through about 2030, data center demand will mainly raise the value of the existing reactor fleet, through restarts, uprates and long-term contracts. It will also push grid operators to find flexibility and transmission capacity faster. Stauff describes the relationship as mutual: data center contracts and investment improve financing conditions for new nuclear projects.1 That benefit is real, but it supports reactors that mostly will not exist before the next decade. Fuel and supply chains are expanding too. Urenco, for example, is adding 2.1 million separative work units of enrichment capacity in New Mexico, with first production scheduled for 2032.9 That date also falls in the 2030s.

The signals that will separate real projects from announcements are concrete: NRC construction permits, signed power purchase agreements in place of prepayments, and first power from Darlington and Natrium. Until those arrive, most AI electricity will come from the grid and from reactors built decades ago.