Small Reactors, Big Promises

By Nekud October 8, 2026 Updated October 22, 2026 15 min read

In October 2024, the four largest cloud providers — Microsoft, Amazon, Google and Meta — began signing the largest wave of nuclear power contracts in history, committing more than 9.8 gigawatts of capacity to data-centre demand. Two years on, one restarted reactor and 1.92 gigawatts of actual power sit on one side of the ledger, and a construction scoreboard that runs to a single broken-ground commercial project in Wyoming on the other. The promise of small modular reactors, which is what most of the new contracts depend on, has run ahead of any reactor ever built on schedule or on budget. This is the state of the hyperscaler nuclear bet in October 2026: how it started, what is actually being built, and how much of the data-centre demand curve it can realistically close.

Key findings

The hyperscaler nuclear bet is the largest coordinated corporate energy purchase in history. By July 2026 the four largest U.S. cloud providers had collectively committed to more than 9.8 gigawatts of nuclear capacity across 13 announced projects, but only about 1.92 GW — the Amazon–Talen deal at the Susquehanna plant in Pennsylvania — was actually delivering power.^{[1, 2]}
Microsoft is restarting Three Mile Island. A 20-year power purchase agreement with Constellation Energy, signed in September 2024, will return the unit-1 reactor — closed in 2019 for economic reasons — to service as the Christopher M. Crane Clean Energy Center, with the full 835 MW output going to Microsoft data centres from the second half of 2027.^{[1, 3]}
The first modern small-modular-reactor construction permit in 40 years has been issued. The U.S. Nuclear Regulatory Commission approved TerraPower’s Natrium sodium-cooled plant at Kemmerer, Wyoming, on 4 March 2026, and the company broke ground the following month; construction is expected to complete by early 2031.^{[4, 5]}
The track record of small modular reactors is poor. NuScale’s flagship Carbon Free Power Project in Idaho was cancelled in November 2023 after costs rose from $3.6 billion to $9.3 billion; Russia’s Akademik Lomonosov and China’s HTR-PM both came in years late and well over budget.^{[6, 7]}
The new reactors run on HALEU. Most Western SMR designs depend on high-assay low-enriched uranium — enriched to between 5 and 20 percent uranium-235, closer to weapons-grade than conventional fuel — and the U.S. currently has only one commercial HALEU enrichment supplier, with most of the world’s supply coming from Russia.^{[8, 9]}
Even at full delivery, hyperscaler nuclear covers less than a fifth of projected demand. The Carnegie Endowment’s June 2026 analysis found that the announced deals, if all built, would meet less than 20 percent of the data-centre electricity demand the same companies are creating through 2035.^[1]

The hyperscaler nuclear bet

In the space of about two years, the four largest U.S. cloud providers went from signing the occasional renewable power purchase agreement to making nuclear power their signature infrastructure bet. The trigger was a single, simple observation: the next generation of AI training and inference would be constrained not by chips but by electricity, and electricity was becoming harder to add quickly enough. The data centre of the late 2020s is no longer a hundred-megawatt warehouse; the hyperscale campuses now under construction in Texas, Virginia and Ohio are designed to draw one to five gigawatts each, the output of an entire nuclear power station.[1]

Microsoft went first and biggest in dollar terms. In September 2024 it announced a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island unit 1, an 835 MW pressurised-water reactor that had been shut down in 2019 for economic reasons.[3] The deal, reportedly worth about $16 billion, rebranded the plant as the Christopher M. Crane Clean Energy Center. A federal waiver from the Federal Energy Regulatory Commission followed in June 2026, and the first Microsoft-bound power is targeted for the second half of 2027.[1]

Amazon chose a different shape. In June 2024 it signed a 1.92 GW co-located power purchase with Talen Energy at the Susquehanna nuclear plant in Pennsylvania, an arrangement that is already running.[1, 2] On top of that, Amazon has placed a $700 million equity investment in X-energy in October 2024 and a long-term offtake for a 12-unit Cascade Advanced Energy Facility near Richland, Washington.[10] Amazon’s bets are unusually mixed — existing reactor, equity, and an SMR fleet — but the headline figure is the same: gigawatts.

Google‘s signature deal is with Kairos Power, an Oakland-based developer of fluoride-salt-cooled reactors. The October 2024 master agreement was described as the world’s first corporate procurement of SMR power, with a target of 500 MW of KP-FHR capacity by 2035 and the first reactor by 2030.[11] Google has separately signed a power purchase agreement with the Tennessee Valley Authority for the first 50 MW of Kairos’s Hermes 2 reactor, now under construction in Oak Ridge.[12, 13]

Meta has assembled the largest single bundle: 6.6 GW across agreements with Vistra, Oklo and TerraPower, layered on top of its earlier deal with Constellation for the Clinton nuclear plant in Illinois.[1, 2] None of the Meta–SMR contracts is yet supplying power, but together they make Meta the most exposed of the four to the success or failure of the small-reactor roll-out.

By July 2026 the four companies had signed 13 such deals, according to the SMR Intel deal tracker, totalling just over 9.8 GW of nuclear capacity. About 1.92 GW of that — the Amazon–Talen co-located power and a handful of small existing-plant extensions — was actually flowing.[2] The rest is forward-looking, tied to plants and reactors that are not yet licensed, not yet built, or not yet approved to operate.

What is actually being built

Behind the press releases, a small number of companies are doing the actual building. The construction scoreboard, as of October 2026, is small but real.

TerraPower broke ground in April 2026 on its first Natrium plant at Kemmerer, Wyoming, on the site of a retiring coal plant owned by PacifiCorp. The plant pairs a 345 MW sodium-cooled fast reactor with a molten-salt thermal store that can briefly boost output to 500 MW. The U.S. Nuclear Regulatory Commission issued its construction permit on 4 March 2026 — the first for a commercial non-light-water power reactor in the United States in more than 40 years — and the company has signed a contract with Meta for eight further Natrium units, with the construction-completion target for the first plant holding at early 2031.[4, 5, 12]

X-energy is in the queue behind TerraPower. Its Xe-100 four-reactor project at Dow‘s UCC Seadrift Operations site on the Texas Gulf Coast received a federal environmental approval with a “Finding of No Significant Impact” in May 2026, after a review that took less than a year; the company expects a full construction permit in the first quarter of 2027.[14, 15] The 80 MW helium-cooled modules are designed to feed both electricity and high-temperature process steam into Dow’s chemical plant, an arrangement X-energy’s chief executive has argued is where SMRs first become economic, because the high-temperature heat displaces on-site gas boilers. The company has also signed a deal to deliver 12 Xe-100 units to Amazon’s Cascade site in Washington, with a target of 320 MW in the first phase and expansion to 960 MW.[10]

Kairos Power broke ground on its Hermes 2 fluoride-salt-cooled reactor in Oak Ridge, Tennessee, in April 2026. The reactor is small — about 50 MW of generation — and is being built as the engineering test bed for the larger KP-FHR commercial design.[11, 12] Google has agreed to take the first 50 MW of output under a 2025 agreement with the Tennessee Valley Authority; Hermes 2 is expected to start up before the end of the decade.

In the United Kingdom, Rolls-Royce SMR signed a contract with Great British Energy – Nuclear in April 2026 to deliver the country’s first small modular reactors, with three units to be sited at Wylfa on the coast of Anglesey in North Wales.[16, 17] The 470 MW pressurised-water design is the furthest through any European regulatory process; the design received regulatory justification in March 2026 and is in Step 3 of the Generic Design Assessment.[18]

In Canada, Ontario Power Generation is preparing the Darlington site for a General Electric Hitachi BWRX-300, a 300 MW boiling-water reactor, with first commercial operation targeted for the end of 2030.[12]

In Russia and China, two small modular reactors are already operating: the Akademik Lomonosov, a barge-mounted KLT-40S, has been delivering power to the grid since 2019; and the HTR-PM, a 200 MW high-temperature gas-cooled reactor at Shidao Bay, has been operating since 2023.[12, 19] Both are sometimes treated as proof that SMRs work; both were also years late and substantially over budget, and they sit outside the Western regulatory framework that the new generation of SMRs will have to clear.

A bruising track record

The history of small modular reactors is short, and most of it is cautionary.

NuScale Power, the company that was first to win a U.S. design certification for a small modular reactor — in 2022, for its 50 MWe VOYGR design — saw its flagship project collapse in November 2023.[6] The Carbon Free Power Project at the Idaho National Laboratory, which would have deployed six 77 MWe modules for the Utah Associated Municipal Power Systems, was cancelled after estimated costs rose from $3.6 billion to $9.3 billion and the project could not attract enough subscribers to reach its minimum threshold.[7] The cancellation was followed by a class-action lawsuit by NuScale investors, alleging the company had overstated the project’s prospects.[20]

Westinghouse, the older nuclear hand, has been here before. Its AP1000 design, the predecessor of NuScale’s reactor, was the design used at the Vogtle nuclear plant in Georgia. Vogtle’s two new reactors came in at more than $20 billion over their original $14 billion budget and took more than six years longer than planned.[21] The overruns drove Westinghouse into Chapter 11 bankruptcy protection in 2017. The company has since emerged with a smaller reactor, the AP300, but has not yet broken ground on a commercial unit.

The international track record is similar. Russia’s Akademik Lomonosov, a barge-mounted KLT-40S reactor, took 12.7 years to complete instead of the planned 3.7 years. China’s HTR-PM, a 200 MW high-temperature gas-cooled reactor at Shidao Bay, was supposed to take four years and took nine.[19, 22]

The pattern, as the Institute for Energy Economics and Financial Analysis put it in a 2024 review, is that “projected costs significantly understated actual construction expenditures” across the small modular reactor sector.[22] The Natural Resources Defense Council, in an issue brief, noted that no SMR has been manufactured at scale anywhere, and that the “economy of multiples” that proponents argue will replace the lost economies of scale has no precedent in nuclear construction.[23] A 2024 paper from IEEFA examined the cost claims made by the four leading U.S. SMR developers and found that every one had moved its projected cost-per-kilowatt upwards over the previous three years, with NuScale’s most-recent figure more than double its 2015 estimate.[22]

The implication is that the current cost projections for the new generation of reactors — TerraPower’s Natrium, X-energy’s Xe-100, Kairos’s KP-FHR, the Rolls-Royce SMR — should be read as the floor of a range rather than a point estimate. None of the developers has built a commercial-scale unit, and the cost record of the closest analogues, the barge-mounted and pressurised-water reactors in Russia and China, is not reassuring.

A novel reactor, with a novel fuel

The designs on the table are not the light-water reactors that built the existing commercial nuclear fleet. They are smaller, factory-built, and use a range of coolants and fuel forms that no commercial power plant has run at scale.

TerraPower‘s Natrium runs on liquid sodium and uses metallic uranium fuel, an approach that lets the reactor run on a closed fuel cycle and burn the long-lived actinide waste from conventional reactors.[4] X-energy‘s Xe-100 runs on helium and uses TRISO fuel — tiny kernels of uranium wrapped in layers of silicon carbide and pyrolytic carbon that are designed to be meltdown-proof.[14] Kairos Power‘s KP-FHR uses fluoride salt as the coolant, an approach that the company says will let the reactor run at atmospheric pressure and operate flexibly.[11]

Most of the new designs depend on high-assay low-enriched uranium (HALEU), a fuel that is enriched to between 5 and 20 percent uranium-235, compared with the 3-5 percent enrichment of conventional reactor fuel.[8] HALEU sits in an awkward middle ground: too enriched to be made by the standard enrichment plants that supply the existing commercial fleet, but not enriched enough to be directly usable in a weapon. The United States has only one commercial HALEU enrichment supplier, in Tennessee, with the rest of the world’s supply coming from Russia, until a Department of Energy programme builds out domestic capacity.[9]

The shift to HALEU, and the proliferation concerns that come with it, is one of the more underexplored parts of the SMR story. A 2022 study at the Stanford Center for International Security and Cooperation found that the volume of nuclear waste from SMRs could be 2 to 30 times higher per unit of electricity generated than from conventional reactors, depending on the design, and that any SMR roll-out would also expand the number of sites handling weapons-usable material.[8, 23] A follow-on study from Argonne National Laboratory largely confirmed the waste-volume finding, while noting that the radioactivity per kilogram of spent fuel was lower, because the fuel is consumed more completely.

The existing commercial nuclear fleet, by contrast, runs on conventional low-enriched uranium, a fuel that has been produced at scale for sixty years, with a global enrichment capacity that is robust and diversified. The new fuel cycle that the SMRs depend on is, in commercial terms, still under construction.

Where this leaves us

The hyperscaler nuclear bet is the largest coordinated corporate energy purchase in history, and the engineering behind it is real. The first modern small modular reactor construction permit in the United States in more than 40 years has been issued. The first reactors are being built. The first nuclear-powered restart of a closed plant is on the books.[3, 4, 5]

It is also, in important ways, a bet on a category of reactors that has never been built on time or on budget, with fuel supply chains that do not yet exist at the scale the roll-out will require, and that even at full delivery will cover less than a fifth of the demand the same companies are creating.[1, 22, 8]

The next two years will be the test. The Kemmerer Natrium plant is expected to be complete by 2030, with construction ending in early 2031. The first BWRX-300 at Darlington is scheduled for the end of 2030. The Dow Seadrift Xe-100 is aiming for early-2030s operation.[4, 12, 15] Each of those milestones is contingent on a regulatory and supply-chain track record that has, so far, been more often broken than kept.

The same two years will also see whether the data-centre demand curve keeps its current slope. The International Energy Agency‘s base case projects global data-centre electricity consumption to roughly double, to about 945 TWh, by 2030, driven by AI workloads growing at 30 percent a year.[2] Even with every announced nuclear deal delivered, the four hyperscalers’ new power would still leave the majority of that demand to be met by gas, grid power, and renewables.

The deal, in other words, is real but partial. The reactors are being built, but they are not the whole answer; they are a substantial down payment on a question that is being asked faster than the answer can be delivered.

Hyperscaler	Partner	Technology	Committed MW	First power target	Status
Microsoft	Constellation Energy	PWR restart (Crane Clean Energy Center, ex–Three Mile Island 1)	835	H2 2027	FERC waiver approved Jun 2026
Amazon	Talen Energy	Existing PWR (Susquehanna)	1,920	Active since Jun 2025	Power flowing
Amazon	X-energy	Xe-100 HTGR (Cascade, Washington)	up to 960 (12 reactors)	2030+	Construction permit expected Q1 2027
Google	Kairos Power	KP-FHR fluoride-salt-cooled SMR	500	2030	First reactor under development
Google	Tennessee Valley Authority	Hermes 2 FHR SMR (Oak Ridge)	50	Late 2020s	Under construction since Apr 2026
Meta	Vistra	Existing PWR (Ohio, Pennsylvania)	2,600	Active	Power flowing
Meta	Oklo	Aurora compact fast reactor	up to 1,200	2030s	Early stage
Meta	TerraPower	Natrium sodium-cooled fast reactor (8 units)	up to 4,000	2030+	Construction permit issued Mar 2026
Meta	Constellation	Clinton PWR (Illinois)	~1,100	Active	Power flowing

Major hyperscaler nuclear deals announced since October 2024. Committed capacity is in megawatts. First power target is when the supplier aims to deliver electricity, not the date of the announcement. Company figures are self-reported. Sources: [1, 2, 3, 10, 14, 15].