Small Modular Reactors Win First US Permit as AI Power Gap Widens

A regulatory first arrives during a power crunch

On September 29, 2026, the U.S. Nuclear Regulatory Commission approved a construction permit for the Tennessee Valley Authority to build a 300-megawatt GE Vernova Hitachi BWRX-300 at its Clinch River site in Oak Ridge, Tennessee.23 Industry coverage calls it the first construction permit the agency has issued for a commercial small modular reactor.21 The review officially began in July 2025 and took about 14 months, finishing four months ahead of schedule.21 TVA has not set a construction start date. It says it is still negotiating with industry partners over cost sharing, and its interim CEO said the priority is making sure Tennessee Valley ratepayers do not pay for first-of-a-kind technology.30

The timing matters. For two years, AI companies have said nuclear power is how they will meet their electricity needs. The Clinch River permit is the clearest sign yet that the regulatory side is moving. Read alongside the demand forecasts and the cost figures from comparable projects, though, it also shows how far apart the AI build-out and the reactor build-out are in time. Approval to build is real progress. It does not mean reactors will power data centers this decade.

What an SMR is, and what it isn't

Small modular reactors are fission reactors rated at up to 300 megawatts of electricity per module, compared with 550 to 1,500 megawatts for conventional plants. Microreactors generally produce 20 megawatts or less.19 The idea is to build modules in factories, ship them to a site, and add more as demand grows.13 Designs range from light-water units like the BWRX-300 and NuScale's 77-megawatt module to high-temperature gas, molten salt and sodium-cooled reactors.1924 Many of the advanced designs need high-assay low-enriched uranium (HALEU), a fuel with almost no commercial supply today. Centrus is targeting first production from new enrichment equipment in 2029, and TerraPower's Natrium, X-energy's Xe-100 and Kairos all depend on it.23

For data center operators, the appeal is steady, carbon-free power around the clock, with some proposals putting reactors right next to server halls. Supporters argue that these "behind-the-meter" setups could skip the years-long waits for grid upgrades.20 That pitch speaks directly to the grid problem described below.

Demand forecasts disagree on size, not direction

Forecasters agree that data center power demand is rising sharply. Their numbers differ a lot, mostly because they measure different things. Gartner expects worldwide data center power demand to grow about 27% in 2026, to 132 gigawatts, and to reach 290 gigawatts by 2030. It also expects consumption to rise from 447 terawatt-hours in 2025 to 565 terawatt-hours in 2026.1 TrendForce, using a broader "demand capacity" measure, puts 2026 at 161 gigawatts, up about 31% from 122.9 gigawatts in 2025.2 The IEA estimates data center electricity use grew 17% in 2025, with AI-focused sites growing 50%, and projects demand rising from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030.13

The U.S. figures vary even more. Goldman Sachs Research expects U.S. data center demand to rise from 31 gigawatts in 2025 to 41 gigawatts in 2026 and 66 gigawatts in 2027.5 Another analysis puts U.S. demand at 75.8 gigawatts in 2026.9 Gaps that large come from definitions and methods. Readers should treat any single number as an estimate, not a measurement.

What all of these forecasts share is the claim that supply is the limiting factor. Gartner says AI capacity is now constrained by power availability and that grid supply will not keep up with planned construction.1 TrendForce gives the most specific version: data center demand capacity of 490.7 gigawatts by 2030, against roughly 222.6 gigawatts of grid capacity available to data centers, a gap of 268 gigawatts. More than 170 gigawatts of that gap is in the United States.2 TrendForce notes that the gap leaves out on-site generation, so it is not a forecast of actual shortages. It does expect grid delivery delays to become more common after 2028.27

The grid is filling the gap with solar and storage

The U.S. grid is adding capacity quickly, just not nuclear capacity. S&P Global Market Intelligence expects more than 90 gigawatts of new U.S. capacity in 2026: about 51.2 gigawatts of solar, 25.7 gigawatts of storage and 13.1 gigawatts of wind. Gas, geothermal, hydro and nuclear together add less than 9 gigawatts.6 In PJM, the largest U.S. grid, 74% of projects studied in the interconnection queue since 2020 have withdrawn.6 Reported PJM capacity prices rose 833% between the 2024–25 and 2025–26 delivery years, which pushed regulators to make data centers pay for grid upgrades.3

That is the environment SMRs are trying to enter, and it explains why the near-term nuclear deals are mostly about existing reactors. Microsoft's 20-year agreement to restart Three Mile Island Unit 1 as the Crane Clean Energy Center involves an 835-megawatt conventional reactor, not a modular one.13 Amazon's Pennsylvania campus next to the Susquehanna plant uses an operating plant's output.14 Meta has signed 20-year power deals with three Vistra nuclear plants while also supporting Oklo and TerraPower's SMR plans.17

The deals point to the 2030s

The advanced reactor agreements tell the same story about timing. Google's deal with Kairos Power covers up to 500 megawatts by 2035, starting with a 50-megawatt unit around 2030.22 Amazon anchored a roughly $500 million investment round in X-energy and is targeting more than 5 gigawatts by 2039. Its Cascade project in Washington state starts with four units totaling 320 megawatts.24 As of early October, no license application had been filed for Cascade. X-energy's separate Long Mott project with Dow had no permit, with the NRC's final safety evaluation targeted for November 2026.24

Coverage disagrees on how much physical work is underway, and some of the disagreement comes down to definitions. The Bulletin of the Atomic Scientists argues that no SMR projects are under construction in the United States and that new reactors are at least a decade away.11 Other reporting says Kairos broke ground on its Hermes 2 site in April 2026 under permits issued in 2024.26 TerraPower's Natrium received a construction permit in March 2026 for its Kemmerer, Wyoming site, with completion expected around 2030. Sources give the permit date as either March 4 or March 9.1324 Natrium's 345-megawatt base rating also puts it just above the usual SMR size limit.24 Either way, the overall point holds: none of the new Western designs is producing commercial power, and the earliest credible dates are around 2030.13

Cost is the real constraint

The most useful figures in this story are cost figures, and they are high. Ontario Power Generation's approved budget for four BWRX-300 units at Darlington, the same design TVA will build, is C$20.9 billion, with C$7.7 billion for the first unit.23 NuScale's cancelled Carbon Free Power Project reportedly reached a final estimate of $9.2 billion for 462 megawatts, about $19,900 per kilowatt. Its target power price rose from $58 to $89 per megawatt-hour.24 Those numbers contradict marketing claims of $1 billion to $3 billion per SMR and three-to-five-year construction schedules.20 They also support the Bulletin's argument that the money tech companies have announced so far does not come close to the cost of building reactors at scale.11

The stock market has noticed. NuScale's market value fell from about $5.3 billion in March to roughly $2.95 billion by October 9, and Oklo's fell from about $12.9 billion to $6.45 billion.22 NuScale hit a 52-week low on October 14, after a quarter with only $0.07 million in revenue.25 That same day, the 10-year Treasury yield reached 5.26% and nuclear developer stocks fell. Higher rates reduce the present value of distant cash flows and make capital-heavy construction more expensive.27 NuScale holds the first NRC design approval, but analysts say it has no firm customer for a U.S. build, and its proposed TVA deployment through ENTRA1 still needs a power purchase agreement.21

The takeaway

The coverage supports a clear conclusion. Clinch River shows that the NRC can license an SMR faster than its critics expected, and TVA says the permit gives future BWRX-300 applicants a repeatable path.30 That reduces a major source of uncertainty. Regulatory approval was never the hardest problem, though. Cost, fuel supply and timing are, and on all three the reactor schedule is out of step with the AI demand curve. The steepest growth and the widest grid shortfalls are forecast for 2026 through 203012, while the first new modular reactors are expected around 2030 or later.1322

For the rest of this decade, AI's electricity will come mainly from existing nuclear plants, restarts, gas, and the solar and storage now being added to the grid. SMRs are a bet on the 2030s, and whether that bet pays off will depend on first-unit costs at Darlington and Clinch River more than on any announcement from a tech company. TVA's eventual cost estimate, the first U.S. figure for a BWRX-300, will be the number to watch.23