October 2024, Mountain View. Google signs a Power Purchase Agreement with Kairos Power for 500 MW of nuclear energy by 2035. Two weeks later, Amazon announces a $500M investment in X-energy. Microsoft reopens Three Mile Island Unit 1 with Constellation. Meta launches an RFP for 4 GW of nuclear capacity.
In six months, hyperscalers signed more nuclear contracts than the industry has seen in twenty years. And none of them is for a traditional 1,500 MW reactor.
They are all betting on SMRs: Small Modular Reactors.
What an SMR actually is
A conventional reactor is a one-off engineering project. It is designed on-site, built in 8 to 15 years, costs $10-15B, and produces 1,000-1,600 MW. Vogtle 3 and 4 in Georgia, the most recent reactors completed in the US (2023-2024), came in at $35B total for 2.2 GW—roughly $16,000/kW installed. Finland's Olkiluoto 3 racked up 14 years of delays and €11B for 1.6 GW.
An SMR is something different. It is built in a factory, shipped to the site by truck or barge, and assembled like Lego blocks. Single-unit size ranges from 50 to 300 MW. Footprint is 1-2 acres versus Vogtle's ~600. Safety is "passive": in an accident scenario, the reactor cools by gravity—no pumps, no operator, no external power.
The honest technical analogy: an SMR is to a traditional reactor what a MacBook is to an IBM mainframe. Less absolute power per unit, but mass-producible, rapidly deployable, replicable.
Why now
Three forces converge in 2024-2026.
First, AI electrical demand. A training cluster with 100,000 H100s draws 150 MW continuously. A modern hyperscaler campus reaches 1-2 GW. McKinsey projects US data center demand to grow from 25 GW (2024) to 80 GW (2030). The public grid cannot keep up: in Virginia, Dominion Energy has already turned down new interconnection requests.
Second, regulation. The US NRC certified the NuScale VOYGR design in January 2023—the first SMR ever certified in the US. The DOE has unlocked $900M in grants to TerraPower and X-energy. The UK started GDA (Generic Design Assessment) for Rolls-Royce SMR. Canada selected GE Hitachi BWRX-300 for Darlington.
Third, gas cost. Long-term natural gas prices have stayed volatile after 2022. Add EU ETS carbon cost at €80-100/ton and combined-cycle gas LCOE blows past $70-90/MWh. The economic window has opened.
The players: a map
| Vendor | Design | Capacity | Technology | Status | Anchor customer |
|---|---|---|---|---|---|
| NuScale (US) | VOYGR | 77 MW | PWR | NRC certified 2023 | RoPower (Romania) |
| Kairos Power (US) | KP-FHR | 75 MW | Fluoride salt | Hermes under construction (TN) | Google (500 MW @ 2035) |
| TerraPower (US) | Natrium | 345 MW | Sodium + molten salt storage | Kemmerer WY, target 2030 | Berkshire Hathaway Energy |
| X-energy (US) | Xe-100 | 80 MW | TRISO + helium | NRC pre-application | Amazon (Energy Northwest) |
| GE Hitachi (US/CA) | BWRX-300 | 300 MW | Simplified BWR | Darlington 2028-2029 | Ontario Power Generation |
| Rolls-Royce (UK) | RR SMR | 470 MW | PWR | GDA UK in progress | Czech Republic, target 2030 |
| Westinghouse (US) | AP300 | 300 MW | PWR | Design frozen 2023 | TBD, target 2030+ |
| Newcleo (IT/UK) | LFR | 200 MW | Lead-cooled + MOX from waste | R&D, prototype 2031 | $310M round (2023) |
Two entries in the table deserve a closer look. Newcleo, founded by Stefano Buono (ex-CERN, ex-Advanced Accelerator Applications), is headquartered in London but has a strong base in Rome and Turin. The design is Gen IV lead-cooled, fueled by MOX recycled from existing spent fuel—a non-trivial political detail in a country that banned nuclear power by referendum in 1987 and again in 2011.
The hard part: real numbers
SMR cost targets are $5,000-7,000/kW capex and $60-90/MWh LCOE. Those numbers are hypotheses. No commercial SMR has yet been built in the West. FOAK risk—first-of-a-kind—is real and well documented.
The NuScale UAMPS case is the most instructive. In 2020 the Idaho project targeted $5,300/kW. By November 2023 it had risen to $9,300/kW (LCOE $89/MWh) and the Utah municipal consortium cancelled the project. NuScale repositioned to industrial customers and Romania, but the market got the message: NOAK (nth-of-a-kind) requires building the FOAK first.
The same problem applies to Gen IV. Kairos is building Hermes in Tennessee as a non-electric prototype (35 MW thermal) precisely to validate costs before commercial deployment. TerraPower at Kemmerer is waiting on HALEU (uranium enriched 5-20%) whose Western supply chain is today nearly nonexistent—Russia controlled ~40% of global capacity before 2022.
Why they match data centers
Technically, a cluster of 4-8 SMRs sums to 1-2 GW. Exactly the size of a hyperscaler campus. The combined footprint stays under 20 acres, manageable within an industrial property. Passive safety enables flexible siting—the NRC is debating rules to shrink emergency planning zones from 10 miles to less than half a mile.
Economically, the SMR PPA structure resembles a solar+battery purchase: the hyperscaler signs a 20-year take-or-pay, the vendor finances against that contract, the bank accepts it because the customer's credit (Google, Amazon) is AA+. This is exactly the model that unlocked 200 GW of utility-scale solar in the past decade.
Operationally, nuclear is baseload. An AI data center runs 24/7 at >90% load factor. Solar+storage costs less per peak MWh but cannot cover 8,760 hours per year without massive overbuilding. An SMR can.
What still doesn't work
Three problems remain open.
Component supply chain. Heavy forgings for pressure vessels come today from Japan Steel Works, Sfar (China), Doosan (Korea). Western capacity is limited. Building 50 SMRs in the US by 2035 requires capacity that does not exist today.
Certification timelines. NuScale took 8 years for its NRC design certificate. Gen IV will need similar cycles. The 2028-2030 timelines vendors quote assume the NRC introduces accelerated licensing (Part 53), currently in public consultation.
Public acceptance. In Italy, Newcleo will operate outside the country until at least 2032. In Germany, the nuclear phase-out is politically cemented. Even where acceptance is growing (US, UK, Poland, Czech Republic), local siting takes 3-5 years of permitting. Poland has signed Westinghouse AP1000 on the large side and GE Hitachi BWRX-300 on the modular side—a signal that even "easy" markets prefer hedging technology bets.
A note on Gen IV versus Gen III+. Designs like Kairos KP-FHR and TerraPower Natrium operate at 600-700°C, not 320°C as in a traditional PWR. This potentially opens cogeneration (hydrogen, process heat, district heating). For data centers, though, waste heat remains a byproduct: what counts are electric MW. The "integrated reactor-DC-molten salt campus" promises that some vendors make stay theoretical until we see the first deployment.
Realistic timing
No commercial SMR on the grid before 2028-2029. The first will likely be BWRX-300 at Darlington (Ontario), followed by Natrium at Kemmerer and RoPower in Romania. Real scale-up—dozens of units per year—is a 2032-2035 story.
Hyperscalers know this. That is why the PPAs signed in 2024 target 2030-2035 delivery, and meanwhile Microsoft, Google, and Amazon are extending contracts on existing reactors (Three Mile Island, Susquehanna, Diablo Canyon). SMRs are the Plan B for not depending on gas after 2030—not the solution for the training cluster being switched on next month.
Put differently: modular nuclear is coming. But "when" and "how much" will depend on the first FOAK that exits the construction site without 3x budget overruns. That site is today at Darlington. Worth watching.