Announced is not
the same as built.

Is Anyone Actually Building One?

The story is everywhere: artificial intelligence needs enormous amounts of electricity, and small nuclear reactors are going to supply it. Tens of gigawatts have been announced. So it is worth asking a narrow, checkable question: is there a specific place, with an owner and a regulatory filing, where a small modular reactor is being built to power a data center?

The answer, as of August 2026, is almost, but not quite — and the gap between that and the impression most coverage leaves is the reason this page exists.

Three projects come closest, and each falls short in a different way that is worth understanding, because the ways they fall short are the shape of the whole industry.

1. 1
   Hermes 2, Oak Ridge, Tennessee — the closest thing to real

   Kairos Power. The NRC issued construction permits for two 35 MWt units on 21 November 2024, and ground was physically broken on 17 April 2026. A power purchase agreement with the Tennessee Valley Authority was announced in August 2025, with Google taking the clean energy attributes. What it is not: the electricity goes to the TVA grid, not down a wire to a server hall, and no operating license has been applied for. A construction permit is permission to build, not permission to run.
2. 2
   Project Matador, Carson County, Texas — real filing, wrong reactor

   Fermi America. 5,769 acres leased from the Texas Tech University System, and a combined license application the NRC accepted on 5 September 2025. It is explicitly a data center campus. But the reactors are four Westinghouse AP1000s — large reactors of about 1.1 GW each, not small modular ones — no tenant has been named, and site construction was paused in February 2026 pending a state air permit.
3. 3
   Pike County, Ohio — named customer, nothing filed

   Oklo and Meta, announced 9 January 2026: 1.2 GW at the former Portsmouth gaseous diffusion plant, on land Oklo holds under a Department of Energy site use permit. A named site and a named customer, which is more than most. But no application has been docketed with the NRC. Note also that in May 2023 the same Ohio project was two 15 MWe units online by 2028; it is now 1.2 GW by 2034. It was upsized fortyfold and delayed six years at the same time.

What the Data Centers Actually Need

For roughly two decades US electricity demand went nowhere. Efficiency gains in lighting, appliances and industry canceled out population and economic growth almost exactly. Retail sales in 2020 were below 2014. Across the whole decade from 2014 to 2024, growth averaged about half a percent a year.

Then it changed. The most authoritative US estimate comes from Lawrence Berkeley National Laboratory, and there are two versions of it — which matters, because the older one is the one everybody quotes.

The 2026 update is the current figure and is much less widely circulated. It revised the method as well as the numbers: the earlier report counted only graphics processors, while the update also counts the custom chips that Google and Meta design themselves, and it models a category of rack-scale server that did not exist at all in 2018 and was over four percent of shipments by 2025.

The number that actually concentrates minds is not consumption but the forecast revision. Grid Strategies tracks what utilities say they expect five years out. In 2022 the industry forecast 24 GW of peak demand growth. In 2023, 38 GW. In 2024, 64 GW. In 2025, 166 GW — more than a factor of six in three years, with data centers accounting for roughly 55 percent of it. NERC's assessment published in January 2026 raised its ten-year peak forecast by 24 percent, itself 69 percent higher than the year before.

Why They Cannot Just Plug In

If you could connect a new power plant to the grid quickly, none of this would be happening. Nobody builds a nuclear reactor next to a warehouse full of servers because it is cheap. They consider it because the alternative is waiting.

Read the last two rows again. A developer who plans around a specific queued power plant is planning around something with roughly a one in eight chance of existing. That is the counterparty risk that makes building your own generation look reasonable.

There is a second queue nobody had a process for. Connecting a large load has historically been a state matter with no federal standard, and the requests are extraordinary: ERCOT reported roughly 474 GW of large-load interconnection requests in August 2026, about 90 percent of them data centers, against a record actual peak of 85.5 GW. The Governor of Texas ordered an audit of those requests on 3 August 2026. FERC opened a rulemaking on large loads and received over 3,500 pages of comments.

What a Small Modular Reactor Actually Is

The IAEA definition is a reactor producing up to 300 MWe per module. The US Energy Information Administration uses the same threshold and treats microreactors as a subset at 20 MW or less — though note that EIA states that in electrical terms while the Department of Energy states its microreactor range of 1 to 20 MW in thermal terms, and a 20 MW thermal machine is only about 6 to 8 MW electrical. The same number means different things in different documents.

"Modular" bundles three separate ideas that are constantly conflated: major components built in a factory rather than on an open site, shipped and assembled where they are used, and several reactors sharing one site, control room and turbine hall so capacity can be added incrementally.

The Regulatory Ladder

Most confusion in this subject is a licensing confusion. There are several different instruments, they mean different things, and press releases blur them constantly.

• Design certification approves a reactor design by rulemaking. It is the strongest form of design approval and says nothing about any particular site.
• Standard design approval is a staff finding that a design meets requirements. Referenceable, but weaker — the NRC says it "does not ensure license approval".
• Construction permit authorises building. It does not authorise operating, which requires a separate later application.
• Combined license covers construction and conditional operation in one.
• Part 53, effective in 2026, is a new optional risk-informed framework that nobody has yet used to license a plant.

Where the Megawatts Actually Come From

Roughly 8.5 GW of nuclear capacity is now under long-term contract to four technology companies. Every megawatt of it comes from a reactor that already exists or one being brought back from retirement. This is the real market, and it is contractual, dated and large.

This matters for the SMR story in a specific way. Holtec's proposed Pioneer SMR-300 units, docketed in February 2026 and behind Clinch River and Long Mott in the queue, are at the Palisades site. Everything downstream is anchored to a plant that has not yet produced a kilowatt-hour since 2022.

Can You Plug Straight Into a Reactor?

The appealing version of this idea is that a data center sits next to a reactor, takes power directly, and never touches the grid at all — no queue, no transmission charges, no waiting. That is called behind-the-meter co-location, and its legal status was fought out over eighteen months.

1. 1
   November 2024 — FERC says no

   Amazon had bought a 300 MW data center campus physically co-located at Talen's Susquehanna plant, and PJM filed an amended interconnection agreement to expand it. FERC rejected it two to one, finding the non-standard terms had not been justified. Rehearing was denied in April 2025.
2. 2
   June 2025 — the market answers

   Rather than keep litigating, Amazon and Talen restructured the whole arrangement to front-of-the-meter: Susquehanna sells into PJM, Talen acts as Amazon's retail supplier, and the local utility handles delivery. The largest co-located nuclear data center in America de-co-located itself and accepted grid charges.
3. 3
   December 2025 — FERC writes a rulebook

   FERC found PJM's tariff unjust and unreasonable for lack of clarity, and ordered it to create four defined transmission service options for co-located load, plus reforms to behind-the-meter netting rules.
4. 4
   April 2026 — partially settled

   FERC accepted some of PJM's compliance filings and rejected others, notably PJM's attempt to narrow the definition of co-located load. Questions about storage pairing and cost allocation remain open.

The Fuel Problem

Light water SMRs run on the same low-enriched uranium the existing fleet uses, under 5 percent U-235, and the supply chain for that exists. Many advanced designs — the fast reactors, the pebble bed and molten salt machines — need high-assay low-enriched uranium, enriched between 5 and 20 percent. That supply chain barely exists in the West.

For years the only commercial source was Russia. The United States banned Russian uranium imports, and domestic production is being stood up from a very small base, with Department of Energy allocation programs distributing limited quantities to selected companies. This is widely described as the binding constraint on advanced reactor deployment.

What It Costs

The case for SMRs rests on a claim about cost: that factory production and repetition will make the tenth unit far cheaper than the first. The technical term is nth-of-a-kind cost reduction. It is worth being precise about the status of that claim.

The one real number. Ontario Power Generation's BWRX-300 at Darlington, the furthest advanced grid-scale SMR project in the Western world, is costed at CAD 20.9 billion for all four planned 300 MWe units, following a final investment decision on 8 May 2025. OPG breaks the first unit out separately: CAD 6.1 billion for the reactor plus CAD 1.6 billion of shared site works serving all four, so CAD 7.7 billion all in. The fourth unit is forecast at CAD 4.1 billion.

That spread is the learning-curve argument stated in a budget rather than in a brochure: a third off between the first unit and the fourth. It is also, so far, a forecast of learning rather than a measurement of it, since none of the four has been built. Note too that Darlington is not the most advanced SMR construction project in the world. The IAEA counts four SMRs in advanced construction, in Argentina, China and Russia; China's Linglong One at Changjiang has been under construction since 2021.

The cautionary tale. NuScale's Carbon Free Power Project at Idaho National Laboratory was to be the first US SMR: six modules, 462 MWe, for a consortium of Utah municipal utilities. Its target price went from $58 per megawatt-hour to $89 in January 2023, a 53 percent increase. It was canceled on 8 November 2023 because it could not reach the subscription level it needed. Note the causal chain: the design had been certified by the NRC ten months earlier. It did not fail technically or in licensing. Customers would not commit at the revised price.

For scale on the large-reactor side: Vogtle units 3 and 4 in Georgia are the only new US reactors this century. The project ran from 2009 to 2024 and cost over $30 billion.

What Is Actually Being Built

It is possible to read a great deal about nuclear power and data centers without encountering the fact that almost nothing being built to serve this load is nuclear.

Solar and battery storage together are about four fifths of planned 2026 additions. Natural gas is about seven percent. In the interconnection queues, Lawrence Berkeley counts 773 GW of solar and 749 GW of storage against 10.4 GW of nuclear — a ratio of roughly one to a hundred and fifty.

Where gas is being built for this load, it is being built fast and on site. Global Energy Monitor counted 252 GW of US gas capacity in development in January 2026, nearly tripled year on year, with more than a third of it intended to power data centers directly. Apply the same discipline there as everywhere else on this page: in development is not under construction, and the actual 2026 gas additions are 6.3 GW against that 252 GW pipeline.

The honest summary: every gigawatt-scale AI campus operating in the United States today runs on some combination of the existing grid mix and on-site gas turbines. Nuclear is a 2030s proposition at the earliest.

The Case That This Is Overstated

The strongest skeptical position is not that AI demand is imaginary. It is that interconnection requests are a terrible proxy for load, and that a great deal of forecasting has been built on them.

A developer shopping one project to five utilities generates five interconnection requests for one data center. None of them is fraudulent; each utility sees real demand; the system counts it five times. Wood Mackenzie reported in August 2026 that US grid operators and utilities had received requests totalling 1,066 GW for data center projects — about 83 percent of the entire US utility-scale generating fleet, which stood near 1,281 GW at the end of 2025. Wood Mackenzie's own expectation is that operators will commit to roughly 28 percent of it. A number of that size, on that expectation, is evidence of intent and of duplication in unknown proportions rather than a forecast of load.

The historical precedent argument deserves a hearing too. Around 1999 a widely cited claim held that the internet would consume half of all electricity within a decade. It did not; efficiency gains absorbed most of the growth. The analysts who debunked that forecast are, in several cases, the same people raising questions now.

Note also that the near-term forecasts are being revised down even while the long-term ones go up. EIA cut its 2026 generation growth estimate; PJM cut its summer 2028 peak forecast. And the range of published 2030 estimates is so wide that, as Rhodium put it, the gap between the lowest and highest analyst expectations is itself larger than the total current electricity consumption of some entire sectors.

Safety, Waste and the Awkward Paper

Passive safety is the strongest technical argument for these designs. Where a conventional plant relies on pumps, power and operator action to remove decay heat after shutdown, passive designs use gravity, natural circulation and large water inventories to do it unaided. A small core makes that easier, because there is less heat to remove and a better surface-to-volume ratio. This is why some SMR designs argue for a much smaller emergency planning zone — potentially the site boundary rather than a ten-mile radius, which is what makes siting one next to an industrial customer conceivable at all.

The disposal problem is unchanged either way. The United States still has no operating repository for commercial spent fuel, and every reactor built adds to the inventory sitting in dry casks at plant sites. Whatever else small reactors solve, they do not solve that.

What to Watch

Most news in this subject is noise: another framework agreement, another gigawatt figure with no site attached. A small number of things would be real signal.

• Palisades actually restarting. It would be the first US nuclear plant ever brought back from retirement, and it is the anchor for an SMR project at the same site.
• Three Mile Island Unit 1 restarting on the 2027 target, which would make the Microsoft contract real rather than prospective.
• TVA's Clinch River construction permit, a decision expected in Fall 2026 — the first BWRX-300 permit in the United States.
• X-energy's Long Mott permit, safety evaluation targeted November 2026, hearing date not yet set. The second most advanced SMR construction permit application in the country, and it is for a chemical plant.
• An operating license application from Kairos or TerraPower. Both are building under construction permits and neither has applied for permission to run.
• Any construction permit application naming a data center as the offtaker. As of August 2026 there is not one.
• Amazon's Cascade project filing anything at all with the NRC.

Where This Comes From

Every figure here was checked against a primary source in August 2026: NRC docket and licensing pages, FERC orders, EIA generator inventories, Lawrence Berkeley National Laboratory reports, NERC reliability assessments, IAEA and DOE definitions, and grid operator filings. Company press releases were used only where nothing else exists, and where a release conflicted with a docket, the docket won and the page says so.

This page will age faster than anything else on this site. Treat any claim here as carrying its date. If a decision turns on it, go to the docket.

• US Nuclear Regulatory Commission — licensing pages, dockets and Federal Register notices for Kairos Hermes, TerraPower Kemmerer, TVA Clinch River, X-energy Long Mott, Holtec Pioneer, Oklo Aurora and NuScale
• Federal Energy Regulatory Commission — the Susquehanna interconnection orders and the PJM co-located load dockets
• Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report (December 2024) and the 2025 Update (June 2026); Queued Up 2026
• Energy Information Administration — Electric Power Annual, Preliminary Monthly Electric Generator Inventory
• NERC Long-Term Reliability Assessment 2025; PJM and ERCOT load forecasts and system assessments; Grid Strategies national load growth reports
• IAEA and DOE Office of Nuclear Energy definitions; Idaho National Laboratory advanced reactor cost meta-analysis
• Krall, Macfarlane and Ewing, Proceedings of the National Academy of Sciences, 2022, on nuclear waste from small modular reactors