> ## Content Index > Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt > Use this file to discover other available public pages before exploring further. # Small Modular Reactors: A Full Guide to Nuclear Power for AI Data Centers - URL: https://www.insidedeeptech.com/small-modular-reactors-nuclear-power-ai-data-centers/ - Published: 2026-08-17T07:02:53.000Z - Updated: 2026-08-17T07:02:53.000Z - Description: Everything you need to understand the nuclear option for AI power in 2026: what a small modular reactor actually is, which designs are licensed or under construction, the hyperscaler deals that matter, and why most of this electricity will not arrive this decade. - Author: Editorial Team - Tags: Energy, AI A small modular reactor is not a miniature version of a 1970s power plant, and it is not a 2026 product you can order for a campus. It is a class of nuclear fission machines, typically 300 megawatts-electric or less per module, designed to be factory-built, shipped, and stacked as demand grows. What changed is the buyer. The [International Energy Agency](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary?ref=insidedeeptech.com) now projects data-center electricity use roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global demand, with AI-focused sites tripling over that stretch. Hyperscalers that spent the last decade buying wind and solar PPAs have started signing nuclear offtake, equity, and restart deals because they need firm, carbon-free megawatts that show up every hour, not just when the weather cooperates. That does not make small modular reactors the near-term answer. The electricity that will actually reach AI campuses before 2028 is almost all existing nuclear: restarts, life extensions, and co-location next to plants that already run. New modules are a 2030s story. The useful question in 2026 is which designs have left the slide deck, who is paying, and which bottlenecks — licensing, fuel, and first-of-a-kind cost — still decide the schedule. ### Key takeaways - **A small modular reactor is a fission plant, usually ≤300 MWe per module, designed for factory fabrication.** The [IAEA](https://www.iaea.org/topics/small-modular-reactors?ref=insidedeeptech.com) definition is the one that matters. Microreactors (typically under 10–20 MWe) sit inside the same family. Some “advanced reactors” that data-center buyers talk about, including TerraPower’s 345 MWe Natrium, sit just above the cutoff. - **New-build modules will not power 2026 or 2027 AI clusters.** The near-term nuclear megawatts are restarts and existing plants. Microsoft’s 20-year deal with Constellation to restart Three Mile Island Unit 1 as the Crane Clean Energy Center is an 835 MW large light-water restart, not a modular reactor. - **A handful of designs have real paper, steel, or both.** NuScale’s original 50 MWe module is the first small modular design certified by the U.S. NRC. GE Hitachi’s 300 MWe BWRX-300 holds a Canadian construction licence at Darlington. TerraPower’s Natrium received the first U.S. construction permit for a commercial non-light-water reactor on 9 March 2026. - **Hyperscaler offtake is real and still mostly prospective.** Google’s Kairos Power agreement targets up to 500 MW of new modules from 2030 through 2035\. Amazon and X-energy have stated a goal of more than 5 GW of Xe-100 projects in the United States by 2039, starting with a 320 MW first phase in Washington. - **The industrial problems are fuel, cost, and first-of-a-kind construction, not neutron physics.** Many advanced designs need high-assay low-enriched uranium (HALEU). Factory fabrication is a claim until a production line exists. The cancelled NuScale–UAMPS project is the reminder that a certified design can still fail a power-purchase math test. - **This is complementary to fusion, not a substitute for it.** [Fusion](https://www.insidedeeptech.com/can-fusion-power-the-ai-data-center-boom/) is a longer-dated, still-unproven power source. Fission modules are the nearer nuclear option. Neither replaces gas, grids, or efficiency work this decade. ## Small modular reactors at a glance | Attribute | Detail | | ------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | | What it is | A nuclear fission reactor, typically up to 300 MWe per module, designed for serial factory build and multi-module plants | | What it is not | A battery, a fusion machine, a drop-in diesel replacement, or a 2026 SKU | | Core trick | Smaller thermal inventory, more passive safety, and (in theory) factory learning instead of one-off civil construction | | Key hardware | Reactor module, steam or power-conversion island, used-fuel storage, and — for several advanced designs — a HALEU or TRISO fuel line | | Transmission medium | Grid interconnection, or in a few cases behind-the-meter / co-located supply to a campus | | Main applications | Firm clean power for grids, industry, and, increasingly, AI data centers; some designs also sell process heat | | Deployed today | A few small or first-of-a-kind plants abroad (Russia’s floating KLT-40S units; China’s HTR-PM). No commercial U.S. or Canadian new-build module is generating yet | | Biggest unsolved problem | Turning a licensed design into a repeatable, financeable construction product before the AI power gap is filled by gas | ## What a small modular reactor actually is Start with the thing it is not. It will not appear next to a training cluster the way a row of diesel generators does. It is not [fusion](https://www.insidedeeptech.com/the-fusion-energy-startups-closest-to-commercial-power-in-2026/). And the ticker “SMR” is a stock, not a technology. The IAEA’s working definition is the cleanest: advanced reactors that produce up to 300 MWe per module, deployable as a single unit or a multi-module plant, and designed to be built in factories and shipped to site. World Nuclear Association uses the same 300 MWe ceiling and stresses serial production. Reactors below about 10–20 MWe are usually called microreactors. Conventional large light-water units run around 1,000 MWe or more. The “modular” claim is industrial, not nuclear. The physics is ordinary fission: a controlled chain reaction heats a coolant, the coolant drives a turbine or a power-conversion loop, and the plant produces electricity at a capacity factor that wind and solar cannot match without storage. What the vendors are selling is a construction method. Pour less custom concrete. Weld more of the nuclear island in a factory. Ship a module. Repeat. That method is still mostly a design goal. Almost every Western project in 2026 is a first-of-a-kind plant being licensed and built as a one-off. The factory comes later, if the first few units work and someone orders a fleet. ### The problem data centers are actually trying to solve An AI campus is a 24/7 load with a nasty transient. The IEA’s 2026 update is the best public baseline: data-center electricity grew 17% in 2025; AI-focused sites grew 50%; an advanced rack by 2027 can pull peak power comparable to 65 households and dump heat comparable to 30 domestic boilers. Training and inference also swing load by more than 50% of rated capacity within a second. That is a grid problem as much as a generation problem. The same report is blunt about the near-term substitute. Constrained by slow interconnections, U.S. developers are pushing onsite gas. Satellite tracking suggests around one-fifth of those gas-behind-the-meter projects have started land clearing or construction, and the IEA’s range for onsite gas serving data centers by 2030 is 15–27 GW, mostly in the United States. Nuclear is the carbon-free version of that instinct: put firm megawatts next to, or contracted for, the load. It is also why this topic sits next to, rather than on top of, the chip story. [HBM](https://www.insidedeeptech.com/high-bandwidth-memory-hbm-full-guide/) and packaging decide how many accelerators ship. Power and cooling decide how many of those accelerators can be plugged in. The [inference TCO](https://www.insidedeeptech.com/the-real-cost-of-running-ai-in-2026-inference-chips-compared/) conversation is incomplete if the megawatt-hour is assumed to appear on schedule. ## How the machines work A working small modular plant is four problems stacked on top of each other: the nuclear island, the power-conversion island, the fuel, and the licence. ### Light-water modules Most of the designs closest to construction are still light-water reactors. They use ordinary water as coolant and moderator, the same basic physics as the existing U.S. and Canadian fleets. That is a feature. Operators, regulators, and supply chains already know the materials. GE Hitachi’s BWRX-300 is a 300 MWe boiling-water reactor, the largest unit that still fits the IAEA ceiling. Ontario Power Generation holds a Canadian Nuclear Safety Commission construction licence (April 2025) for the first unit at Darlington, with commissioning of that first unit targeted around 2030\. BWXT is under contract for the reactor pressure vessel. Early site work for up to four units is underway. This is the Western project that most resembles a real construction job rather than a demonstration. NuScale’s VOYGR plant is an integral pressurized-water design: the steam generators sit inside the reactor vessel. The NRC certified the original 50 MWe US600 module in January 2023 — the first small modular design to clear that bar. In May 2025 the company received standard design approval for an uprated 77 MWe (250 MWt) module. The first six-module project is now aimed at Romania after the Utah Associated Municipal Power Systems plant at Idaho National Laboratory was cancelled in November 2023 on cost. Certification is not the same thing as a customer who will pay. Holtec’s SMR-300 and several other integral PWRs sit in earlier licensing. They matter as a class: if light-water modules work, they can use existing fuel-cycle infrastructure. They do not solve the HALEU problem because they do not have it. ### High-temperature gas and pebble-bed designs X-energy’s Xe-100 is an 80 MWe high-temperature gas-cooled reactor. Helium cools a pebble bed of TRISO fuel — uranium particles wrapped in ceramic layers that are themselves the containment for the fuel kernel. Four modules make the 320 MW first phase Amazon and Energy Northwest have described for the Cascade Advanced Energy Facility near the Columbia Generating Station in Washington, with an option to grow to 12 modules and 960 MW. Construction is targeted for the end of this decade; operations are a 2030s date. Amazon’s public goal with X-energy is more than 5 GW of U.S. projects by 2039. China’s HTR-PM, a 210 MWe high-temperature gas reactor, has been in commercial operation since 2023\. It is the existence proof that a pebble-bed plant can make power, not a template for a U.S. licence. ### Molten salt and sodium Kairos Power’s KP-FHR uses molten fluoride salt as coolant and a ceramic pebble fuel. The point of the salt is low pressure and high temperature: the nuclear island does not need a huge steel pressure vessel, and the heat is useful for a steam cycle. Google’s October 2024 agreement was the first corporate offtake for a fleet of these modules, sized at up to 500 MW of new 24/7 carbon-free power, with a first unit targeted for 2030 and further units through 2035\. A later three-party structure with the Tennessee Valley Authority covers 50 MW from the Hermes 2 plant in Oak Ridge for Google campuses in Tennessee and Alabama. Kairos’s earlier Hermes unit is a non-powered demonstration — useful engineering, not a data-center feed. TerraPower’s Natrium is a 345 MWe sodium-cooled fast reactor paired with a molten-salt thermal store that can boost output to 500 MW for several hours. That storage is the feature data-center and grid operators actually care about: the reactor can run flat while the plant follows load. The [U.S. Department of Energy](https://www.energy.gov/ne/articles/nrc-issues-construction-permit-terrapowers-natrium-advanced-reactor?ref=insidedeeptech.com) recorded the NRC construction permit for Kemmerer Unit 1, Wyoming, on 9 March 2026 — the first construction permit the NRC has ever issued for a commercial non-light-water power reactor. Completion is expected in 2030\. An operating licence is a separate application. At 345 MWe, Natrium is slightly above the IAEA small-modular ceiling; it is an advanced reactor that buyers lump into the same conversation, and it should be labeled as such. ### Microreactors Units of a few megawatts, often proposed for remote sites, military bases, or a single industrial customer, are a different product. They are easier to imagine next to a small campus and harder to finance. Several U.S. designs have DOE demonstration paths. Treat them as a research and early-deployment class, not as a 2026 hyperscale solution. ## Who is building what | Design | Developer | Type / size | Status as of August 2026 | Earliest credible power | | ------------------- | ------------------------------- | ------------------------------------------ | --------------------------------------------------------------------------------------------------------- | --------------------------- | | BWRX-300 | GE Hitachi / OPG | BWR, 300 MWe | Canadian construction licence; Darlington site work; RPV contracted | \~2030, first unit | | VOYGR / NuScale NPM | NuScale | Integral PWR, 50 MWe certified; 77 MWe SDA | NRC design certification (2023) and SDA (2025); U.S. lead project cancelled; Romania first commercial aim | Early 2030s if financed | | Xe-100 | X-energy | HTGR, 80 MWe | Amazon / Energy Northwest Cascade first phase (4 × 80 MWe); Dow Seadrift CPA under NRC review | 2030s | | KP-FHR | Kairos Power | Fluoride-salt, pebble fuel | Google offtake up to 500 MW; Hermes demo; Hermes 2 / TVA 50 MW | \~2030 first, 2030–35 fleet | | Natrium | TerraPower / GE Vernova Hitachi | SFR + salt store, 345 / 500 MWe | NRC construction permit 9 Mar 2026, Kemmerer, WY | \~2030 | | HTR-PM | China | HTGR, 210 MWe | Commercial operation since 2023 | Already generating | | Akademik Lomonosov | Rosatom | Floating PWR, 2 × 35 MWe | Operating since 2020 | Already generating | Share figures and “GW of offtake” totals published by trackers should be treated as announcements, not as capacity under construction. A power-purchase agreement for a plant that does not yet have a licence, a site, and a fuel contract is a demand signal. It is not a megawatt. ## Why AI buyers showed up Three facts, not a narrative. First, the load is large, concentrated, and fast. Five technology companies’ capital expenditure is, on the IEA’s 2026 figures, larger than global investment in oil and gas production. “AI factories” tracked from satellites more than tripled in capacity in 18 months. Interconnection queues and transformer lead times do not move on that clock. Second, the carbon math that made wind and solar PPAs sufficient no longer closes a 24/7 CFE ledger. A training cluster that must run through a windless week needs firm generation or a heroic quantity of storage. Nuclear’s capacity factor — existing U.S. plants routinely above 90% — is the attribute the buyers are paying for. Third, the alternative firm options have their own queues. Gas turbines saw a 70% surge in orders in 2025, per the IEA. Onsite gas still requires overbuilding generation 30–70% relative to demand if it is to cover a variable AI load reliably. That is why nuclear offtake and gas offtake are happening in parallel, not in a tidy substitution. The deals that are actually signed split into two piles. **Existing atoms, nearer dates.** Microsoft’s 20-year PPA with Constellation (announced 20 September 2024, filed as an 8-K) underwrites the restart of Three Mile Island Unit 1, rebranded the Crane Clean Energy Center, adding about 835 MW to PJM. Constellation’s own disclosures put restart capex around $1.6 billion and an original in-service estimate of 2028; later project communications have discussed 2027, still subject to NRC approval and a licence extension toward 2054\. Amazon’s earlier co-location with Talen at Susquehanna is the other mature pattern: put the campus next to a running plant. Meta’s life-extension offtake at Clinton in Illinois is the same idea. Google’s collaboration to restart NextEra’s Duane Arnold plant in Iowa (more than 600 MW, targeted early 2029) is another restart, not a module. **New modules, later dates.** [Google’s Kairos agreement](https://blog.google/company-news/outreach-and-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/?ref=insidedeeptech.com) and Amazon’s X-energy investment plus the Energy Northwest Cascade project are the two clearest small-modular offtakes. They are also the ones whose first electrons are scheduled after the current AI buildout wave. That is not a criticism of the contracts. It is the construction calendar. ## The bottlenecks that will slip the dates ### Licensing is faster than it was. It is still licensing. Natrium’s construction permit in 18 months, versus a 27-month published schedule, is genuine progress. It is also a permit to build, not a permit to operate. Every non-light-water design is teaching the NRC a new materials and safety case. Light-water modules move faster because the agency has seen the physics before. Neither path is a software release. ### HALEU is a fuel-cycle problem, not a press-release problem Low-enriched uranium for today’s light-water fleet is enriched to less than 5% uranium-235\. HALEU sits between 5% and 20%. Several advanced designs, including many sodium-cooled and some TRISO-fueled concepts, are specified around it. Commercial U.S. HALEU production has been a DOE demonstration item, not a commodity market. A reactor that cannot fuel cannot slip to “late 2020s” by working harder on concrete. TRISO manufacturing is a related, narrower chokepoint for Xe-100 and Kairos. X-energy’s TRISO-X line in Oak Ridge is part of the Amazon-backed financing story for a reason. ### First-of-a-kind cost is how projects die NuScale’s UAMPS cancellation is the canonical example. The design was certified. The target price was not. Serial factory production is the proposed cure, and it only starts after someone absorbs the first units. Hyperscaler offtake and DOE demonstration money exist to cover that gap. They do not delete it. ### Siting, water, and the grid still dominate the calendar A 300 MWe module is small by nuclear standards and large by data-center campus standards. It still needs a site, cooling, a transmission path or a private wire, emergency planning, and a host community. Behind-the-meter schemes that isolate a campus from the grid run into utility and reliability rules that are being rewritten in real time. The IEA’s warning applies here too: most data centers still prefer a grid connection. Nuclear that only works as an islanded private plant is a narrower product than the marketing implies. ### This will not replace the rest of the stack Even a successful 2030s module fleet does not retire the need for better chips, better cooling, or better siting. [Space-based data centers](https://www.insidedeeptech.com/data-centers-in-space-hype-or-the-next-infrastructure-frontier/) remain a research and publicity story. Fusion remains a later option. Efficiency — including the order-of-magnitude drop in energy per simple AI query the IEA documents — changes the slope. It does not remove the need for firm megawatts. ## A realistic timeline | Horizon | What is actually available | Confidence | | ----------------------- | -------------------------------------------------------------------------------------------------------------------- | ------------------------------------- | | Now (2026) | Existing nuclear PPAs and co-location; restarts in licensing; no new Western commercial module generating | Deployed (existing plants) | | Near term (2027–2029) | Possible Crane / other restarts if NRC and outage work hold; first Western modules still in construction | Medium: regulatory and execution risk | | Medium term (2030–2035) | First BWRX-300, Natrium, Kairos, and Xe-100 units if current schedules hold; FOAK cost discovery | Plausible, execution-limited | | Long term (late 2030s) | Factory-built fleets if the first units work and HALEU / TRISO supply exists; Amazon–X-energy 5 GW target lives here | Speculative on volume | One useful heuristic: watch construction permits, fuel-fabrication lines, and binding offtake that survives a cost reset. Do not watch announced gigawatts. ## The bottom line 2026 is the year nuclear re-entered the AI infrastructure conversation as a procurement category, not a climate slogan. The physics is old. Fission plants already run at the capacity factors data centers want. The new work is industrial: a smaller, more passive, factory-intended machine, a regulator willing to license non-light-water designs, and a set of buyers who can underwrite first-of-a-kind cost. A few projects have steel and paper. None of the Western new-builds are a 2026 power product. For anyone making decisions today, the split is clean. If the question is how to power a cluster this year or next, the nuclear answer is an existing plant, a restart, or a grid contract — and, in practice, a lot of gas. If the question is whether a vendor’s small modular design is real, ask for the licence status, the fuel assay, the construction permit, and a customer who is still there after the cost number moved. ## Frequently asked questions ### What is a small modular reactor? A nuclear fission reactor, generally 300 MWe or less per module, designed so that much of the nuclear island can be built in a factory and installed as one or more repeatable units. It is a construction and licensing concept built on ordinary fission physics, not a new way of splitting atoms. ### How is this different from a conventional nuclear plant? Output and construction method. A large light-water unit is typically around 1,000 MWe and is built as a unique civil-engineering project. A modular design aims for a smaller thermal inventory, more passive safety systems, and serial fabrication. Whether those aims survive first-of-a-kind construction is the open industrial question. ### Will small modular reactors power AI data centers in 2026? No. New Western modules are a 2030s generation story. The nuclear megawatts arriving for AI this decade are overwhelmingly existing reactors, life extensions, and restarts such as Constellation’s Crane Clean Energy Center. ### Is Microsoft’s Three Mile Island deal a small modular reactor? No. It is a 20-year power-purchase agreement to restart an existing \~835 MW pressurized-water reactor, Unit 1, under a new name. It is the most important near-term nuclear-for-AI contract in the United States, and it is a large light-water restart. ### Who is actually building these plants? GE Hitachi and Ontario Power Generation at Darlington (BWRX-300). TerraPower at Kemmerer, Wyoming (Natrium, construction permit issued March 2026). X-energy with Energy Northwest and Amazon in Washington, and with Dow in Texas. Kairos Power with Google and TVA in Tennessee. NuScale, certified in the U.S., is now aimed at a first commercial plant in Romania after its U.S. municipal project collapsed. China and Russia already operate small or first-of-a-kind units that are not U.S.-licensed designs. ### What is HALEU and why does it matter? High-assay low-enriched uranium, enriched above 5% and below 20% uranium-235\. Several advanced (especially non-light-water) designs specify it. Until a commercial U.S. supply exists, those designs have a fuel-cycle critical path that no offtake announcement can shorten. ### Are these reactors safer than existing plants? Vendors argue that a smaller inventory, passive decay-heat removal, and — in some designs — low-pressure coolants or TRISO fuel reduce the set of accidents a plant must survive. That is a design claim. Safety is determined by the licensed safety case, the operator, and the regulator, not by the marketing category “small” or “modular.” ### How does this compare with fusion for data centers? Fusion, if it works as a power plant, is a later and still-unproven option; we covered that separately in the [fusion-for-AI](https://www.insidedeeptech.com/can-fusion-power-the-ai-data-center-boom/) and [fusion startups](https://www.insidedeeptech.com/the-fusion-energy-startups-closest-to-commercial-power-in-2026/) guides. Fission modules use a working commercial physics package and a live licensing path. They also produce long-lived waste and require a fuel cycle. Buyers who need electrons before 2030 are not choosing between the two. They are buying gas, existing nuclear, and interconnection. ## Methodology This guide is based on primary company, regulator, and agency disclosures available as of 17 August 2026: the IAEA’s definition of small modular reactors (up to 300 MWe per module); the IEA’s *Key Questions on Energy and AI* executive summary (data-center demand of 485 TWh in 2025 to 950 TWh in 2030); the U.S. Department of Energy’s 9 March 2026 notice of the NRC construction permit for TerraPower’s Kemmerer Unit 1; Constellation’s 20 September 2024 Crane Clean Energy Center announcement and 8-K; Google’s 14 October 2024 Kairos Power agreement; and Amazon’s and X-energy’s October 2024 Cascade / Xe-100 disclosures. Offtake totals compiled by third-party trackers are treated as announcement tallies, not as capacity under construction. Vendor cost, schedule, and “factory fabrication” claims are company-stated and are not independently costed here.