Small Modular Reactors: EIA Counts 30 Designs as AI Load Climbs

Thirty reactor designs and very little new power

The U.S. Energy Information Administration has published a review of the American small-reactor pipeline. The main finding is that the country has many reactor designs in development but almost no new nuclear power to show for them yet. U.S. utilities run about 98 gigawatts of nuclear capacity, and very little of it was added in recent decades because building reactors has been expensive and licensing has been slow.1 The agency surveyed commercial small modular reactor (SMR) and microreactor designs as of February 2026.1 Trade coverage of the review counted about thirty advanced reactor designs being developed for roughly 36 possible projects across the country.34

The review uses two size classes. SMRs usually produce about 300 megawatts or less per unit. They are built from factory-made modules that are shipped to a site and assembled there. Microreactors are smaller still, at 20 megawatts or less, and can run on their own or as part of a microgrid.4 The EIA says plainly that SMRs are being considered for AI, data centers and other industrial uses where developers may not want or need a grid connection.1 That explains why a statistical survey of reactor designs is being read as part of the AI infrastructure story.

Why the timing matters: demand records

The review comes as U.S. electricity demand grows faster than it has in a long time. In its October Short-Term Energy Outlook, the EIA projected consumption of about 4,288 billion kilowatt-hours in 2026 and 4,356 billion in 2027. The 2025 record was 4,195 billion.3338 The agency links the growth to data centers serving AI and cryptocurrency, along with homes and businesses switching from fossil fuels to electricity.38

The EIA's monthly forecasts have moved around this year. The July outlook put 2027 demand at 4,399 billion kWh.34 The October number is lower, but the trend is the same in every edition: demand keeps rising. The more important change is which part of the economy is driving it. Data centers are counted in the commercial sector. The EIA has forecast that commercial electricity sales will overtake residential sales for the first time on record. Earlier outlooks put that crossover in 2027 and later ones in 2026.3934 In September the agency said the commercial sector would account for 63% of the growth in electricity sales in 2026 and 56% in 2027.36

The effects are uneven across regions, and that is where grid limits start to bite. Coverage of the October outlook cited a 41% jump in wholesale prices in PJM, the data-center-heavy grid serving the mid-Atlantic, while Mid-Columbia prices in the Northwest fell 23%.33 The EIA expects PJM to supply almost 45% of the country's generation growth.36 Household bills are rising as well. Residential prices were forecast to climb nearly 5% in 2026, and utilities cite spending to strengthen and expand the grid as one reason.39

The EIA's own forecast also shows that nuclear will not help much in the near term. Its share of generation stays at about 18% through 2027. Natural gas stays near 40%, and renewables rise from 24% to 27%.3240 For the next two years, AI demand will be met by gas, solar, wind and the existing reactor fleet. Small reactors will not contribute.

The hyperscaler bet

Tech companies are still signing deals for nuclear power in the 2030s. Trackers that follow these deals mostly agree on one number: the big hyperscalers have committed about 9.8 GW of nuclear capacity across 13 announced deals.1820 Meta has made the largest commitment, up to 6.6 GW, through agreements with TerraPower, Oklo, Vistra and Constellation.1918 Google's deal with Kairos Power targets 500 MW of SMRs starting around 2030.1318 Amazon backed X-energy and agreed with Energy Northwest to fund four Xe-100 reactors producing 320 MW, with room to expand to 960 MW.15

Some reported figures differ by source. Amazon's investment in X-energy appears as $500 million in some accounts and as $700 million for rights to up to 12 units in others.1417 One outlet put the value of Microsoft's Three Mile Island restart agreement at $1.6 billion. Another estimated about $16 billion over the life of the contract.1915 The headline totals are also measured in different ways. The 9.8 GW figure counts signed corporate deals. A separate count attributed to the International Energy Agency puts conditional offtake agreements between data centers and SMR projects at 45 GW, up from 25 GW at the end of 2024.15 The difference is between firm deals and announced intentions, and the larger number should be treated as a ceiling, not a forecast.

The most useful figure is how little of this power is actually being delivered. One analysis found that as of July 2026 only about a fifth of the committed capacity was flowing. Nearly all of that came from Amazon's agreement for output from the existing Susquehanna plant. None of it came from a small reactor.18 The same analysis expects the first SMR power for a data center from a Kairos demonstration unit around 2030. It estimates first-of-a-kind SMR power at $80–150 per megawatt-hour, well above the $60–80 that vendors say later units will reach.18 So far, the AI-nuclear story mostly means buying power from reactors that already exist. New-build SMRs remain an option on capacity in the next decade.

A permit, but not a plant

The most concrete step since the EIA review came on September 29. The Nuclear Regulatory Commission gave the Tennessee Valley Authority a construction permit for a 300 MW GE Vernova Hitachi BWRX-300 at the Clinch River site near Oak Ridge.2123 The NRC finished the review in 14 months, four months ahead of schedule. TVA's earlier Early Site Permit helped speed it up.22

Outlets describe the permit's place in history differently. One policy group called it the first construction authorization for a commercial SMR in the U.S.22 Others called it the second advanced-reactor construction permit, after TerraPower's Natrium project at Kemmerer, Wyoming, was approved in March.2328 The difference is mostly about labels. Natrium is a sodium-cooled design, not a light-water SMR. The underlying fact is the same: two permits for new kinds of reactors in one year, after decades with almost none.

The more skeptical coverage points out what the permit does not cover. TVA has not announced a construction start date, a project cost or a cost-sharing partner.25 It still needs a separate NRC operating license before it can load fuel.2226 One estimate puts the earliest operation in 2032.25 Federal money is involved: the Department of Energy committed $400 million to the project, and TVA committed about $150 million of its own.26 Still, the last large U.S. reactors, Vogtle Units 3 and 4, came in years late and billions over budget. Only actual construction will show whether smaller, simpler designs can avoid the same problems.26

The microreactor track

The EIA review gives a lot of space to a part of the market that has little to do with data centers: the military. In June 2025 the Energy Department launched a pilot program to speed testing of advanced reactors outside the national laboratories. It selected companies including Aalo, Oklo, Radiant, Valar and Last Energy, and the companies pay for their own pilots.1 The Army's Janus program, launched in October 2025, builds on Project Pele, a transportable reactor program.1 The Air Force is working with Oklo on a microreactor at Eielson Air Force Base in Alaska, aiming for 1 to 5 MW by 2027.2 The Army later said it would award up to $2.2 billion to five companies at five bases, with a goal of having at least one microreactor producing usable power by September 30, 2028.8

This could end up mattering more for AI than the hyperscaler announcements. Military projects have deadlines, public funding and sites that are already secured. If they reach operation, they will produce the cost and performance data that commercial buyers do not yet have.

Reading the signal

The EIA review is best read as a list of options, not a supply forecast. The design pipeline is broad, federal support has increased, and licensing is getting faster: TVA's review finished early.22 But the agency's own demand forecasts show the AI-driven increase in electricity use happening now, in places like PJM, while the first commercial SMRs are on schedules that stretch into the next decade.3325 Data center growth will strain grids long before new reactors arrive. For roughly the next five years, nuclear's contribution to AI will come from restarted plants and long contracts with existing ones. Small reactors will start to matter only if the first few units are actually built on schedule, at a known cost and with a paying customer behind them.