by | Oct 25, 2026 | Physics | 0 comments

Small Modular Reactors for Data Centers: Where Things Actually Stand

· @Thomas Lee Abshier, ND

Small modular reactors crossed an important line in the summer of 2026: they moved from paper to physics. Four privately built American reactors went critical between June 4 and July 4. None of them has yet made a kilowatt of electricity for a customer, and none is yet powering a data center.

I stand on the side that small reactors will become a major source of firm power for computing, and that the question is when, not whether. The physics works, the regulators are moving faster than at any time in fifty years, and the buyers have real money on the table. The honest timeline, however, is the early-to-mid 2030s for meaningful data-center supply, not 2027. The rest of this essay explains why both halves of that statement are true.

The problem: computing has collided with the grid

American data centers used about 176 TWh in 2023, roughly 4.4% of the nation’s electricity. Lawrence Berkeley National Laboratory projects 325 to 580 TWh by 2028, which is 6.7% to 12% of national demand (LBNL via Berkeley Lab Energy Analysis). A single large AI campus now draws as much as a small city, and it draws it around the clock.

Three features of that load matter for the choice of power source:

  • It is constant. Training and inference clusters run 24 hours a day, so the buyer needs firm power, not power that follows the sun or the wind.
  • It is concentrated. Hundreds of megawatts at one address strain local transmission, and new grid interconnections commonly take years.
  • It is carbon-constrained. The large buyers have public net-zero commitments they intend to keep.

Natural gas meets the first two needs and fails the third. Solar and wind meet the third and fail the first without very large storage. Nuclear fission is the one mature technology that meets all three at once. That is the whole reason hyperscalers have come knocking at the door of an industry that built almost nothing for forty years.

The physics: what “small,” “modular” and “critical” actually mean

A conventional reactor produces about 1,000 MWe. A small modular reactor produces roughly 50 to 300 MWe per module, and a microreactor produces under about 20 MWe, often 1 to 10. “Modular” is a manufacturing claim, not a physics claim: the bet is that reactors built on a factory line will be cheaper and more predictable than reactors built by hand on site.

The central quantity in any fission reactor is the effective multiplication factor, the ratio of neutrons in one generation to neutrons in the generation before it.

k_{\mathrm{eff}} = \frac{N_{n+1}}{N_n}, \qquad k_{\mathrm{eff}} < 1 \;(\text{subcritical}), \quad k_{\mathrm{eff}} = 1 \;(\text{critical}), \quad k_{\mathrm{eff}} > 1 \;(\text{supercritical})

“Critical” means k equals exactly 1: each fission produces, on average, exactly one further fission, and the neutron population holds steady. Criticality says nothing about how many neutrons are present. A reactor can be critical at a few watts or at a few hundred megawatts.

That is why the 2026 milestones were “zero-power” criticality tests. The operators assembled real fuel in a real core and showed that the chain reaction becomes self-sustaining exactly where their neutronics models predicted. It is the first test of a design and the most basic one. It does not test heat removal, materials under irradiation, fuel burnup, power conversion, or years of continuous operation, and those are where reactors are won or lost.

Control depends on a fact of nuclear physics that is easy to overlook. About 0.65% of the neutrons from uranium-235 fission arrive seconds to a minute late, emitted by decaying fission products. Operating just below “prompt” criticality lets those delayed neutrons set the pace, stretching the reactor’s response from microseconds to seconds, which is slow enough for control rods and temperature feedback to manage.

Two fuel terms recur in this industry:

  • TRISO (tri-structural isotropic) fuel encases poppy-seed-sized uranium kernels in layers of carbon and silicon carbide. Each particle is its own containment vessel and holds its fission products to temperatures well above 1,600 °C, which is the basis for many of the passive-safety claims in gas-cooled and salt-cooled designs.
  • HALEU (high-assay low-enriched uranium) is enriched to between 5% and 20% U-235, against 3–5% for today’s commercial fleet. Higher enrichment allows a smaller core and a longer interval between refuelings. Most advanced designs need it, and that need is one of the industry’s binding constraints, as discussed below.

The summer of 2026: four reactors go critical

Executive Order 14301, issued in May 2025, directed the Department of Energy to bring at least three advanced test reactors to criticality by July 4, 2026. DOE did it by authorizing reactors itself, outside the NRC commercial-licensing process, through its Reactor Pilot Program and a companion Nuclear Energy Launch Pad at Idaho National Laboratory (ANS Nuclear News). Companies paid their own way; DOE supplied the authorization. Four reactors made the deadline.

Date (2026) Company and reactor Site Design Stated next step
July 4 Aalo Atomics, Aalo-X critical test reactor Idaho National Laboratory Full-scale core of a 10 MWe sodium-cooled reactor Commercial 10 MWe unit powering an on-site data center in 2027
June 30 Deployable Energy, Unity Idaho National Laboratory 1 MWe water-moderated, gas-cooled “nuclear battery”; conventional sub-5% fuel Full-power test at INL’s DOME facility in 2027
June 18 Valar Atomics, Ward 250 Utah San Rafael Energy Lab Helium-cooled high-temperature gas reactor, HALEU TRISO fuel Further experiments, then power ascension
June 4 Antares Nuclear, Mark-0 Idaho National Laboratory Heat-pipe microreactor Electricity in 2027; military installations in 2028

Three observations keep this in proportion.

First, every one of these was a zero-power test. Valar’s own description is that Ward 250 demonstrated its reactor physics at essentially no measurable energy output (NucNet). The milestone proves the core is designed correctly. It does not prove the plant.

Second, the speed is real and new. Antares was the first privately developed non-light-water reactor to go critical in the United States in more than four decades. Deployable Energy went from project start to criticality in roughly 150 days. Valar’s was the first DOE-authorized reactor built outside a national laboratory. For an industry that measured schedules in decades, that is a cultural change, not just a technical one.

Third, most of these reactors are not primarily data-center products. Antares is aimed at military bases, Deployable at remote and defense sites, Valar at mobile and industrial use. Aalo is the exception; its Project Ascension plans a commercial 10 MWe reactor and a co-located data center at INL. Press coverage that bundles all four as “AI power” overstates the link.

The commercial track: steel in the ground, power around 2030

The test reactors above are a parallel road. The plants that will actually sell electricity go through NRC licensing, and on that road the leading projects are now under construction or close to it, with first power clustered from 2028 to the early 2030s.

Developer and design Size Status (Oct 2026) Target first power
TerraPower Natrium, Kemmerer, WY 345 MWe, sodium-cooled fast reactor NRC construction permit March 2026, the first for a commercial reactor in nearly a decade; nuclear construction began April 2026 About 2030
Kairos Power Hermes 2, Oak Ridge, TN 50 MWe, fluoride-salt-cooled, TRISO Broke ground April 2026; 50 MW contracted to TVA’s grid for Google data centers About 2030
Oklo Aurora-INL, Idaho 75 MWe, sodium-cooled fast reactor Excavation and DOE review under way; total cost not yet published 2028
GE Vernova Hitachi BWRX-300, Darlington, Ontario 300 MWe, boiling-water Licensed and under construction since 2025; four units planned First unit 2030, all four by 2035
X-energy Xe-100, Dow Seadrift, TX 4 × 80 MWe, helium-cooled, TRISO pebbles NRC review due to finish late 2026; permit expected Q1 2027 Early 2030s
NuScale VOYGR 77 MWe per module, light-water Only SMR design with full NRC approval; no plant under construction Not yet scheduled

Two things stand out. The light-water designs (BWRX-300, NuScale) are the most licensable because they use familiar physics and ordinary fuel. The non-light-water designs (Natrium, Hermes, Xe-100, Aurora) promise higher temperatures and better economics but carry more first-of-a-kind risk. Kairos illustrates the pattern: the NRC extended the completion deadline for its non-power Hermes 1 test reactor from the end of 2026 to April 2029.

Abroad, the only land-based SMRs actually running are Chinese and Russian. China’s HTR-PM, a 210 MWe pebble-bed gas reactor, has operated since December 2023, and Russia’s floating Akademik Lomonosov since 2020. China’s 125 MWe Linglong One missed its first-half-2026 operation target; local reports in early October said it had begun hot functional testing and aimed at fuel loading and grid connection before the end of 2026 (NucNet). Even the world’s most practiced reactor builder is running a year late on its first commercial SMR.

The buyers: gigawatts signed, tens of megawatts built

The technology companies have committed to small reactors on a scale that would have been unthinkable five years ago. The commitments differ greatly in how binding they are.

Buyer Developer Headline commitment What is firm today
Switch Oklo 12 GW by 2044 A non-binding framework agreement
Amazon X-energy More than 5 GWe by 2039 Equity investment; first project 320 MWe at Energy Northwest, expandable to 960 MWe, with a permit application targeted for end of 2026
Meta Oklo 1.2 GW campus in Pike County, Ohio A prepayment agreement; Phase 1 of about 150 MW around 2030
Google Kairos Power 500 MW fleet Hermes 2 under construction, 50 MW to TVA’s grid for Google data centers

Add these up and the announced pipeline exceeds 18 GW. The capacity actually under construction for a named data-center customer is about 50 MW, at Oak Ridge. That ratio is not a scandal; it is what an early market looks like. But the reader should treat headline gigawatts as options on future reactors, not as reactors.

The buyers know this. Their deals are structured as prepayments, equity stakes and frameworks precisely because they are financing the learning curve. A hyperscaler that spends a few hundred million dollars to help a vendor reach its tenth unit is buying a seat at the table if the economics work, and it can afford to lose the stake if they do not.

The hard constraints

The physics is not the bottleneck. Four things are.

1. Cost

Small reactors lose the economy of scale that drove reactors to 1,000 MWe in the first place, and they must win it back through serial factory production. Until they reach that point, first units are expensive. The cautionary case is NuScale’s Carbon Free Power Project in Idaho, cancelled in November 2023 after its projected cost rose to about $9 billion and its power price to $89/MWh and beyond (SMR Intel).

The best current Western benchmark is Darlington. Ontario Power Generation budgets C$20.9 billion for four 300 MWe units, all-in including interest and escalation, and projects about 14.9 Canadian cents per kWh over 60 years (World Nuclear News). That is roughly US$12,500 per kW, against a DOE “nth-of-a-kind” target of about $3,600 per kW (iRecruit cost guide). The whole SMR thesis rests on closing that gap, and no one has closed it yet.

2. Fuel

Most advanced designs need HALEU, and outside Russia and China there is almost no commercial-scale supply. A U.S. GAO review this year found those two countries are the only ones able to produce it at scale (Telangana Today/IANS). Centrus’s plant in Piketon, Ohio, the only American producer, had delivered about 900 kg by mid-2026; on June 30 it signed a $900 million DOE contract to build commercial-scale enrichment (Simply Wall St). For scale, TerraPower has said its first core alone needs about 15 tonnes. A Carnegie Endowment paper this month calls fuel a major risk factor for the whole sector and describes a chicken-and-egg standoff: no fuel plants without reactor orders, no reactor orders without fuel (Carnegie). The light-water SMRs avoid this problem, which is a large part of their appeal.

3. Licensing for commercial service

The DOE pathway that made the summer’s criticalities possible authorizes test reactors on federal programs. Selling power to a private data center still requires NRC licensing, and the NRC, though faster under the ADVANCE Act, has issued only a handful of construction permits for advanced designs. Every commercial schedule in this essay assumes the regulator keeps its new pace.

4. Construction and supply chain

Nuclear-grade components, qualified welders and experienced builders are scarce in a country that finished only two new reactors (Vogtle 3 and 4) in thirty years, years late and billions over budget. First-of-a-kind delays are the norm, as the Kairos and Linglong One schedules already show. Spent fuel also remains a legal and political question: the United States still has no permanent repository, so every new reactor adds to on-site storage.

A realistic timeline, and why the case still holds

Reading the schedules against the constraints, I expect three phases:

  1. 2027–2028: first electrons. Microreactors of 1 to 10 MWe (Aalo, Antares, Deployable, Valar) attempt power operation, and Oklo’s 75 MWe Aurora-INL aims to start. Aalo’s co-located data center would be the first real nuclear-powered computing site, but at 10 MWe it is a demonstration, not a supply.
  2. 2029–2032: first commercial plants. Natrium, Hermes 2, the first Darlington BWRX-300 and the first Xe-100 units come on line, if they hold schedule. Meaningful, contracted power begins flowing to data-center buyers, mostly through the grid rather than behind the fence.
  3. 2032 and after: the factory test. Only when developers are building their fifth through tenth identical units will we learn whether modular construction actually lowers cost. This is the decade in which the gigawatt-scale agreements either convert into fleets or quietly lapse.

Why, given all the obstacles, do I still stand on the side of the technology?

  • The physics is sound and has now been demonstrated in new designs. Four novel cores went critical where their models said they would. That is not a small thing after forty years of paper reactors.
  • The need is structural. A constant, concentrated, carbon-constrained load is precisely the load that fission is suited to serve. That fit does not change with the news cycle.
  • The money is patient. Hyperscalers are among the few buyers in the world able to underwrite a learning curve, and their prepayments and equity are doing exactly that.
  • The regulatory climate has shifted. Construction permits for Natrium, Hermes 2 and, soon, Xe-100 show the NRC issuing approvals for non-light-water reactors at a pace not seen since the 1970s.
  • Competition is real. China’s operating HTR-PM and nearly finished Linglong One mean that if the United States does not learn to build small reactors, it will eventually buy them.

The areas of genuine uncertainty are cost, fuel and schedule, and they are large. None of them is a physical impossibility. They are engineering and industrial problems of the kind this country has solved before when it chose to.

Plain Language summary

AI data centers need huge amounts of steady, clean electricity, and small nuclear reactors are one of the few sources that can supply it. This summer four new American reactors “went critical,” meaning their chain reactions became self-sustaining, which proves their basic design works. They did this at almost zero power, though, so none of them is making electricity yet. Larger commercial small reactors are now being built and should start delivering power around 2030. The main obstacles are cost, a shortage of the special fuel many designs need, licensing and construction delays. The technology is real and the need is real, but large-scale nuclear power for data centers is a 2030s story, not a next-year story.

Sources

Submit a Comment