Can Small Modular Nuclear Reactors Solve the Data-Center Power Problem?

Photo: U.S. Department of Energy, HAER survey, Wikimedia Commons, Public domain

01 The question behind 9.8 gigawatts of deals

The short answer, grounded in the record as of September 2026: small modular reactors are becoming an answer for the 2030s, not the solution to the power problem AI has right now. The numbers that frame the debate come from the SMR Intel deal tracker: as of May 2026, 13 announced hyperscaler nuclear projects commit more than 9.8 gigawatts of capacity to AI data centers, and every major hyperscaler — Meta, Amazon, Microsoft and Google — has signed at least one nuclear deal. Yet as of July 2026, only about 1.9 gigawatts of that committed power is actually flowing, nearly all of it from one power purchase agreement with an existing plant. The rest is schedule.

That is not a dismissal; it is a calendar. U.S. data center electricity demand is projected to surge from roughly 23 gigawatts to 42 gigawatts, with nearly half of AI data centers planned for 2026 already delayed — a roughly 7-gigawatt gap that is bottlenecking hundreds of billions in capital expenditure. The SMR pipeline points at exactly the load growth the forecasts describe. It just arrives years after the shortage begins.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 19, 2026; where evidence is incomplete we say so.

02 What the pipeline actually contains

Strip the press releases into categories and the portfolio is blunt: fast electrons from old reactors, slow electrons from new ones. Microsoft's deal is the simplest and soonest — a 20-year agreement to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center, buying its full 835 megawatts with first power targeted in 2027. Amazon bought time twice: a front-of-meter agreement with Talen's Susquehanna plant that ramps from 840 to 1,200 megawatts in 2029 toward 1,680 to 1,920 megawatts by 2032, plus a stake in X-energy whose Cascade facility starts at 320 megawatts of Xe-100 SMRs, scalable to 960, in the early 2030s. Google signed with Kairos Power for up to 500 megawatts of molten-salt-cooled reactors, the first targeted for 2030. Meta spread bets across Oklo, Vistra and Constellation — up to 6,600 megawatts landing between 2032 and 2035.

The one reactor already in the ground for this market is TerraPower's 345-megawatt Natrium at Kemmerer, Wyoming. In March 2026 the NRC issued its construction permit — the first for a commercial non-light-water reactor in more than 40 years — and Bechtel began earthworks in April. Completion is targeted for February 2031. That is the state of the art in fast: a design that began development in the 2000s, funded by the largest corporate offtakers on earth, delivering power in five years.

03 The gap math: demand doubles before the first SMR lands

Run the two curves against each other and the mismatch is arithmetic, not opinion. Data center demand is projected to roughly double by 2030; the IEA projects nuclear will supply over half of U.S. data center electricity only by 2035. Between those dates sits the entire first generation of SMR deployment: licensing, HALEU fuel supply chains, and construction execution that has no commercial track record. Roughly 80 percent of all committed hyperscaler nuclear capacity has yet to reach commercial operation, with the bulk delivering between 2029 and 2035.

The industry's structural advantages are real — factory fabrication, smaller unit sizes, output that scales in 50-to-300-megawatt increments matching campus expansions — and NuScale, whose 77-megawatt module remains the only NRC-certified SMR design, has joined TVA and ENTRA1 Energy on a 6-gigawatt program. But certification of a design is not operation of a plant. The UAMPS project that would have been NuScale's first deployment collapsed in 2023 on rising cost projections, a reminder that the binding constraint on SMRs has never been physics. It has been first-of-a-kind economics meeting customer tolerance for delay.

04 The economics: who pays for lesson one

First-of-a-kind SMR electricity is estimated at 80 to 150 dollars per megawatt-hour — with early estimates running to 180 — against nth-of-a-kind targets of 60 to 80, existing nuclear at around 60, and renewables-plus-storage bundles near 50. Analysts project costs decline only after 10-plus gigawatts of cumulative deployment. That is precisely what the hyperscaler deals are engineered to solve: guaranteed offtake from the only customers on earth willing to pay a premium for firm, carbon-free, 24/7 power at any price the board will sign.

In effect, Big Tech has volunteered to be the deep pocket that buys lesson one. The equity wave is real — over 1.3 billion dollars flowed into SMR companies in 2025, and DOE-backed deployments plus the first North American approvals landed in 2026. But offtake agreements are not steel in the ground, and every month of schedule slip widens the window in which gas turbines, grid batteries and restarts like Crane absorb the demand SMRs were announced to serve.

05 The fuel and supply-chain fine print

Two constraints receive less attention than cost and usually matter more. The first is HALEU — high-assay low-enriched uranium — which most advanced designs, including Natrium, require and which has no fully domestic commercial supply chain yet; Russia's war economy remains the nominal incumbent supplier. The second is the NRC queue itself: the 2026 Advanced Nuclear Framework under the ADVANCE Act is meant to move licensing from decades toward years, but the agency has never licensed a commercial non-light-water reactor to operation, and TerraPower's permit is for construction, not operation.

Against those stand real accelerants: construction permits arriving in months rather than years, standardized designs that let factory output ramp, and a customer class that signs 20-year offtakes before the first concrete pour. The 9.8 gigawatts of commitments are best read as an industrial policy instrument — hyperscalers using procurement power to rebuild a supply chain the government subsidized but never guaranteed.

06 The verdict: an option, not a rescue

The verified facts: 13 deals, 9.8 GW committed, ~1.9 GW flowing, first new-design power targeted 2030-2031, demand doubling by 2030. The analysis: SMRs solve the data-center power problem only for companies whose growth plans survive to the 2030s. For the 2026-2029 crunch — the seven-gigawatt gap, the delayed campuses, the capacity auctions — the working answers are existing-plant PPAs, restarts, gas, batteries, and demand flexibility. SMRs are the answer to the next power problem: the moment when AI baseload stops being a wave and becomes a permanent feature of the load curve.