Small Modular Reactor for a Data Centre: How It Works, What It Takes

Key takeaways

• An SMR is unlikely to power a European data centre before the 2030s. No new SMR supplies a data centre in the West today. The UK review of the Rolls-Royce design was still in its third step at the regulator’s last update (ONR, 3 February 2026).
• Every nuclear plant boils water to drive a turbine. A split of one uranium-235 atom releases about 200 MeV, of which 197 MeV is usable heat (NRC training manual). A 300 MWe BWRX-300 splits about 0.92 kg of uranium-235 a day (our calculation).
• Small helps safety because of decay heat, the heat fuel still makes after shutdown. The DOE handbook gives about 7% of full power at once, 2% after an hour and 1% after a day. For a BWRX-300 that is about 61 MW, 17 MW and 9 MW (our arithmetic).
• Most data-centre deals are not supply contracts. Meta’s deal with Vistra covers existing plants. Its Oklo deal is a prepayment and its TerraPower deal is funding plus rights to energy. The one PPA with a new advanced reactor that we found is Kairos and TVA, for up to 50 MW.
• A 1 GW IT campus at PUE 1.3 draws 1,300 MW. At a 92% capacity factor that needs 5 BWRX-300, 4 Rolls-Royce SMR, 5 Natrium, 19 NuScale or 18 Xe-100 units, plus one spare, and a grid connection sized for the full load. Darlington’s first unit costs C$25,667 per kW (our division of Ontario’s figures).

A small modular reactor (SMR) is unlikely to power a European data centre before the 2030s. The first Western units are still being built, and no European design has finished its regulatory review. In our reading of the dates below, a developer who needs power before 2030 cannot count on a new SMR anywhere in Europe. The decisions that can be made now sit elsewhere: secure a grid connection for the full load, contract energy from existing nuclear plants where the market allows it, and ask each SMR vendor for the signed contract, the last regulatory milestone and the fuel supply. The checklist at the end lists them.

The rest of this article explains why. It starts with the mechanism, because the mechanism decides what an SMR can and cannot do for a data centre. Then it compares seven designs, separates signed deals from announced ones, and sizes the reactors for a 100 MW and a 1 GW campus. Our wider reporting on the grid limit on data-centre growth and on where the power is sets the demand side.

From atom to turbine

Every nuclear plant is a steam plant with an unusual heat source. When a uranium-235 atom splits, it releases about 200 MeV, a unit of energy for one atom. The NRC’s Westinghouse training manual gives 207 MeV in total and 197 MeV as usable heat. The rest leaves as neutrinos, particles that pass through everything. Most of the usable energy is the motion of the two fragments, which turns into heat inside the fuel. In everyday units, 200 MeV is 3.2 × 10^-11 joules.

That is tiny for one atom and huge for a kilogram. Fully splitting 1 kg of uranium-235 gives about 22,800 megawatt-hours (MWh) of heat (our calculation). A BWRX-300 produces 870 MW of heat, written 870 MWt, with t for thermal. That is 20,880 MWh a day, so about 0.92 kg of uranium-235 splits each day. The fuel also burns some plutonium that forms inside it, so treat 0.92 kg as an order of magnitude.

The heat boils water, the steam spins a turbine, and the turbine drives a generator. Electrical output is written MWe. Divide MWe by MWt and you get thermal efficiency. We computed it from the IAEA SMR Catalogue 2024, which carries vendor-supplied data. The BWRX-300 gives 300 / 870 = 34.5%. NuScale gives 77 / 250 = 30.8%. Rolls-Royce gives 470 / 1,358 = 34.6%. Xe-100 gives 82.5 / 200 = 41%, and Natrium 345 / 840 = 41%.

Water-cooled designs land between 31% and 35%. The high-temperature gas and sodium designs reach about 41% because the coolant leaves the core hotter; the Xe-100’s helium leaves at 750 °C, per the same catalogue. The heat that does not become electricity must go somewhere. For a BWRX-300 that is about 570 MW (870 minus 300, our subtraction), so each plant needs a large cooling source of its own.

Moderator, coolant and fuel: what separates the designs

A fission starts when a neutron hits an atom, and each split releases more neutrons. Fresh neutrons move fast. Uranium-235 splits far more readily when the neutrons are slow, so most reactors contain a moderator, a material that slows neutrons down. A coolant is the fluid that carries heat out of the core. Some designs use one material for both jobs.

Light-water designs use ordinary water as moderator and coolant. The BWRX-300 boils water inside the reactor vessel at 7.2 MPa and uses natural circulation, so it has no recirculation pumps (GE Vernova General Description). NuScale puts the core and the steam generators inside one vessel, also with natural circulation. The Rolls-Royce SMR is a three-loop pressurised water reactor with pumps at power. These three burn ordinary low-enriched uranium (LEU), which is below 5% uranium-235. The BWRX-300 averages 3.81% and has a 12 to 24 month refuelling cycle.

High-temperature gas designs such as the X-energy Xe-100 use helium as the coolant and graphite as the moderator. The core holds about 220,000 graphite pebbles, each carrying about 18,000 TRISO particles, tiny fuel kernels wrapped in ceramic and carbon layers. Pebbles are added and removed while the reactor runs, about 173 fresh ones a day at full power (IAEA catalogue). The fuel is enriched to 15.5%, which counts as HALEU, high-assay low-enriched uranium: between 5% and under 20% (DOE).

Sodium-cooled fast designs such as TerraPower’s Natrium have no moderator, so they use fast neutrons and need more concentrated fuel: HALEU metal, up to 19.75% in the IAEA listing. Liquid sodium runs near atmospheric pressure. A tank of molten salt stores heat, which lets output rise from 345 to 500 MWe for up to 5.5 hours. Oklo’s Aurora is also a liquid-metal fast reactor without a moderator.

Molten-salt-cooled designs are represented here by Kairos Power. Its reactor uses fluoride salt as the coolant and graphite pebbles with TRISO fuel, according to the NRC’s Hermes project page. We could not open that page ourselves, so we rely on trade-press reports of it.

The split matters commercially. The three light-water designs avoid HALEU, so their limits are cost and licensing, not fuel. The gas, sodium and salt designs depend on a fuel supply that barely exists yet (see the limits section).

Why small helps with safety: decay heat

When control rods drop into the core, the chain reaction stops within seconds. The fuel keeps making heat, because the fission products inside it keep decaying. The DOE Fundamentals Handbook puts this at roughly 7% of the previous power straight after shutdown, about 2% within the first hour and about 1% within the first day. Applied to a BWRX-300, that is about 61 MW at once, 17 MW after an hour and 9 MW within a day (our arithmetic: 0.07, 0.02 and 0.01 times 870 MW). If that heat is not removed, the fuel overheats. Decay heat, not the chain reaction, is the safety problem.

Size helps for a reason of geometry, which is our reasoning and not a vendor claim. Double the linear size of a core and its volume, which makes the heat, grows 8 times. The surface that can shed heat grows only 4 times. A smaller core has more surface for each watt, so conduction, radiation and natural circulation can carry decay heat away without pumps or electric power.

Decay heat, not the chain reaction, is the safety problem.

Designs use this in different ways. NuScale submerges its steel containment in a large water pool. Its vendor data in the IAEA catalogue claim unlimited time for core cooling without AC or DC power, water addition or operator action. The NRC issued the US460 a standard design approval on 29 May 2025 (Federal Register). The cooling claim is the vendor’s analysis. The BWRX-300 uses a tall chimney inside the vessel to drive the water flow. The Xe-100 and Kairos build safety into the fuel, since the TRISO layers hold fission products at high temperature. The IAEA catalogue says the Xe-100 core cannot melt by design, again on vendor-submitted data. Natrium’s sodium is at low pressure, so there is no high-pressure pipe break to cool after.

“Passive” is therefore a design goal that each regulator tests, and it does not apply equally to every design. The Rolls-Royce SMR keeps both active and passive safety systems, and uses natural circulation only as a backup for decay-heat removal.

What “modular” means

The word carries three meanings. Size is the first: the IAEA describes SMRs as advanced reactors of up to 300 MWe per module. By that definition the 470 MWe Rolls-Royce design and the 345 MWe Natrium sit above the line, although the IAEA catalogue lists both. Factory fabrication is the second: modules are built in a factory and shipped to site. Repetition is the third: a plant is built as copies and units are added over time, as in NuScale’s six-module VOYGR-6 plant of 462 MWe.

Cost savings rest on the second and third meanings, and the West has not yet shown them. Darlington is the first four-unit build, and its lower cost for later units is a budget, not an outcome. “Modular” does not mean plug-and-play. Each plant still needs a site licence, a grid connection and a cooling source, as the cost section shows.

Seven designs compared, with licensing status on 4 October 2026

The table gives technical values from the IAEA catalogue, which holds vendor-submitted data, and licensing status from regulators and companies. Efficiency is our division of MWe by MWt.

Three designs have permission to build something. The BWRX-300 has its Canadian licence to construct. Natrium and Hermes 2 have US construction permits. NuScale has a design approval and no site, and the Rolls-Royce design is still under review. Timing for Darlington unit 1 is not settled: NucNet reports commercial operation by the end of 2029, while ANS says grid connection at the end of 2030. No commercial SMR operates in the West. Outside it, Russia’s floating Akademik Lomonosov began commercial operation in May 2020 (WNN), and China’s HTR-PM gas-cooled plant entered commercial operation in December 2023 (WNN). We did not assess either in detail.

Signed, funded or announced: the data-centre deals

A press release that mentions gigawatts can describe a supply contract, a prepayment, a funding round or a letter of intent. Only the first delivers energy. Our analysis of Equinix’s nuclear agreements applies the same test, so this section covers the wider picture. We do not total the figures, because “up to” capacities and different kinds of agreement do not add up.

The ladder at the end of this section sorts each agreement by type. Select one to read what it commits and its source.

A press release that mentions gigawatts can describe a supply contract, a prepayment, a funding round or a letter of intent. Only the first delivers energy.

The only PPA with a new advanced reactor that we found is Kairos and TVA, for up to 50 MW. That is one twentieth of the load of a 1 GW IT campus. The nuclear energy that physically supplies data centres today, or will within a few years, comes from existing plants and restarts: Talen, Constellation, NextEra and Vistra.

Each new-build announcement is waiting on a named physical step. Oklo’s 10-Q says the company is evaluating the timing and form of future NRC applications, so no application is pending. The NRC has not issued the Long Mott permit, and it has not received one for Cascade. Meta’s TerraPower units follow a Natrium plant whose own construction permit dates from 4 March 2026, with completion expected in 2030.

Sizing a campus: 100 MW and 1 GW of IT load

The assumptions come first. IT load runs at 100% of nameplate, which is conservative because real use is lower. Power usage effectiveness (PUE) is total facility power divided by IT power, and we use 1.3. The Uptime Institute’s 2026 survey puts the industry average at 1.52, and says recent large builds routinely reach 1.3 or better. Facility load is IT load times PUE, so 100 MW of IT is 130 MW and 1 GW is 1,300 MW.

We apply a 92% capacity factor to every design. Capacity factor is the share of the year a plant produces at full power. EIA data give a US nuclear mean of 92.3% over 2016 to 2025, with a range of 90.8% to 93.4% and 91.0% in 2025. Mature plants earn that figure, and first-of-a-kind units will probably fall short for years. N is the smallest number of units whose output at 92% covers the load. N+1 adds one spare unit.

The calculator at the end of this section applies this method to all five designs, with your own IT load, PUE and capacity factor.

Take the 1 GW case. Facility load of 1,300 MW over 8,760 hours is 11,388,000 MWh a year, or 11.4 TWh. One BWRX-300 makes 300 × 8,760 × 0.92 = 2,417,760 MWh a year. The campus needs 4.71 of them, so 5, and a sixth as the spare. Three Rolls-Royce units fall 3 MW short of the load once the capacity factor applies, so the answer is 4 and 5. Each 0.1 of PUE adds 100 MW of load on a 1 GW IT campus.

At 100 MW the match is poor. One BWRX-300 is 2.3 times the 130 MW load (300 / 130), and the campus would use only 47% of its energy (1,138,800 of 2,417,760 MWh). A single reactor trip would remove all supply. The fit is better with 77 to 80 MWe modules, or with a shared plant that serves several customers through the grid.

Why the grid connection stays

N+1 covers one unit missing, whether tripped or refuelling. It does not cover a trip during an outage. Light-water units refuel on cycles of 12 to 24 months for the BWRX-300 and 18 months for NuScale and the Rolls-Royce SMR (GE Vernova, IAEA catalogue). The outage lengths are not in the sources we read. The Xe-100 refuels while running. Plan so that at most one unit is in an outage at a time.

A grid link sized for the full 130 MW or 1,300 MW covers the gap, and the first units of any design are the likeliest to trip. Batteries or UPS units still bridge the seconds between a trip and the grid picking up. The generators that a Tier III or IV facility carries stay in place. Nuclear changes the source of the energy and leaves the resilience stack as it was. A flat data-centre load does suit a plant that runs best at steady full power.

Siting adds one more question. In the US, the NRC’s rule 10 CFR 50.160 lets an SMR size its emergency planning zone by consequence analysis, using a threshold of 10 mSv over 96 hours, and the zone can end at the site boundary (Federal Register, 16 Nov 2023). We did not verify the European and UK rules. Ask the national regulator before you assume a distance. Our siting cost tool covers the grid and land side.

Nuclear changes the source of the energy and leaves the resilience stack as it was.

What it costs: Darlington and the cancelled UAMPS plant

The best Western price is Darlington. The Ontario government’s 8 May 2025 release says the first unit costs C$6.1 billion, with C$1.6 billion of systems common to all four units, and that OPG’s C$20.9 billion budget covers all four plus site preparation and engineering to date (release text as reproduced here). Our arithmetic: C$7.7 billion over 300,000 kW is C$25,667 per kW for unit 1. C$20.9 billion over 1,200,000 kW is C$17,417 per kW for the programme. The second figure includes work already done, so it is not a pure construction cost.

NuScale’s Carbon Free Power Project for UAMPS ended in November 2023. The trade outlet SMR Intel reports a final estimate of US$9.2 billion for six modules (462 MWe), which is about US$19,900 per kW (our division). It reports that the cost of power rose from US$58 to US$89 per MWh. Participants would not subscribe. We could not verify these figures at the primary level, and we do not state the subsidy assumed.

Both numbers come from first plants. Cost falls only if many units get built, and the order book today holds a handful.

Strongest counter-evidence and what we do not know

The case against our caution is real. Darlington has its licence to construct and its first regulatory hold point was lifted on 30 March 2026 (CNSC project page). Natrium has a construction permit. If those schedules hold, the first Western SMRs will make power around 2030. If the Rolls-Royce review ends well, European units could follow in the 2030s. Our conclusion is a judgement about regulatory gates and dates, and no physical limit forces it.

Our conclusion is a judgement about regulatory gates and dates, and no physical limit forces it.

Four limits cut the other way. Schedules slip: Hermes 1’s deadline moved from 31 December 2026 to 30 April 2029. Costs are high at first. The 92% capacity factor is a fleet average from mature plants, so our counts are the optimistic case. And the HALEU fuel supply is thin. The Centrus 10-Q describes a US$900 million DOE task order to deliver 1 metric ton of 19.75% HALEU by July 2032, with two optional lots of 5 tons each at US$17.0 million per ton (Centrus 10-Q). We did not quantify fleet demand, so we draw no conclusion in tonnes.

Gaps remain. We did not read the contracts. The Rolls-Royce review outcome after February 2026 is unconfirmed. Technical data are from vendors. Emergency-planning rules outside the US are unchecked.

What a European operator should do and watch

1. Secure the grid connection first. A campus needs a grid link sized for its full load, whichever contract it signs. Our regulation tracker lists the European rules that apply.
2. Contract existing nuclear energy. Long-dated PPAs with operating plants, and life-extension or uprate deals, are the nuclear energy that can reach a campus before 2030. The Talen, Constellation and NextEra contracts above show the pattern, in the US.
3. Treat the Rolls-Royce review as a gate. Check the ONR page for a Design Acceptance Confirmation. We found none after the 3 February 2026 update.
4. Watch the European projects. Sweden: Vattenfall selected the supplier on 15 June 2026, which starts detailed planning and is not a construction contract (Vattenfall). Poland: the BWRX-300 at Wloclawek (WNN). Czech Republic: an early-works contract at Temelin in April 2026 (ANS). UK: three Rolls-Royce units at Wylfa (Neutron Bytes). We did not reconcile the first-power dates in these secondary sources, so we give none.
5. Ask every vendor six questions. Is there a signed PPA or only a memorandum? What was the regulator’s last milestone, and when? Who supplies the fuel? What is the price per MWh, and which subsidy does it assume? Who carries a cost overrun? How big is the grid link?
6. Check again in 2027. Watch the Crane restart, the NRC decision on Long Mott after its November 2026 safety evaluation, Hermes 1 progress, Darlington unit 1, an Oklo application, and the first binding PPA for a new SMR.

Sources

• IAEA, Advances in Small Modular Reactor Technology Developments (SMR Catalogue), 2024
• NRC, Westinghouse Technology Systems Manual, section 2.1
• DOE Fundamentals Handbook DOE-HDBK-1012/3-92, decay heat (text hosted by Engineering Library)
• GE Vernova Hitachi, BWRX-300 General Description, rev J
• DOE, What is HALEU
• CNSC, Commission authorises OPG to construct a BWRX-300 at Darlington, 4 April 2025
• CNSC, Darlington New Nuclear Project
• Federal Register, NuScale US460 standard design approval notice, 4 June 2025
• ONR, Rolls-Royce SMR generic design assessment
• Vattenfall, Videberg Kraft selects Rolls-Royce SMR, 15 June 2026
• NRC, Long Mott Energy project page
• TerraPower, NRC approves Natrium construction permit
• Oklo, Form 10-Q for the quarter to 30 June 2026
• Meta, nuclear energy projects, 9 January 2026
• TerraPower, deal with Meta
• NuScale, Form 10-Q for the quarter to 30 September 2025
• Centrus Energy, Form 10-Q for the quarter to 30 June 2026
• Uptime Institute, Global Data Center Survey 2026
• EIA, Electric Power Monthly, Table 6.07.B
• NRC, Emergency preparedness for SMRs and other new technologies, Federal Register, 16 November 2023
• Ontario government release of 8 May 2025 (text reproduced by Today’s Northumberland)