Small Modular Reactors

A small modular reactor is a fission plant of roughly 300 MWe or less, designed to be built repeatedly from factory-made modules rather than assembled once on site. The engineering claim is specific and worth stating precisely, because the whole case rests on it: shrinking a reactor makes each unit more expensive per kilowatt, and the promise is that series production wins that penalty back and more.

State electrical and thermal megawatts separately, every time. This is not pedantry in this subject; it is the difference between a claim and its opposite. China's HTR-PM is two 250 MWt modules on one turbine, which is not a 500 MWe plant. Kairos's Hermes 2 was permitted at 70 MWt and 20 MWe. A specialist trade outlet reporting the Darlington programme in 2025 described four BWRX-300 units as totalling 4,800 MW; four units of 300 MWe are 1,200 MWe. This brief writes MWe or MWt after every figure and treats any unqualified megawatt number in this field as uninterpretable.

What this brief is about, and what it is not. This is a brief about a delivery model — factories, modules, licensing throughput, learning curves and outturn cost. It is deliberately not a brief about reactor physics, coolant chemistry or fuel cycles, which belong to advanced fission: molten salt, thorium, fast spectra, breeding and reprocessing are that brief's property, and where this one needs a fact from that domain it borrows rather than re-derives. Nor is it about what happens to the fuel afterwards, which belongs to nuclear waste solutions. Several designs discussed here are both small and advanced; the two properties are independent, and conflating them is how a coolant novelty gets credited to modularity. TerraPower's Natrium is the standing example: 345 MWe sodium-cooled, uprated to 500 MWe by molten-salt thermal storage, which makes it an advanced reactor of roughly conventional size rather than a small modular one.

The rungs of the deployment ladder are the organising device of this brief. Announced, design-reviewed, licensed to construct, under construction, operating. Almost every disputed claim in this subject turns out to be a slide between two adjacent rungs — a completed regulatory design review described as an approval, a memorandum of understanding described as an order, a right to purchase power described as a power-purchase agreement. The OECD Nuclear Energy Agency's third SMR Dashboard counts 127 designs worldwide, up from 98 in its previous edition, with 51 in some pre-licensing or licensing process and seven operating or under construction. That ratio, 127 to 7, is the honest summary of the sector's state and nothing else in this brief contradicts it.

Established The deployment record is short, and it is a record of cost escalation and cancellation rather than of series learning. Nothing in what follows is a physics failure. Every plant named here works or would work; the argument is entirely about what things cost, how long they take, and who signs a contract at the end.

Established The most-watched Western project died on price before it was built, and the price series is the single most instructive artefact in the subject. NuScale's Carbon Free Power Project with Utah Associated Municipal Power Systems was reconfigured in July 2021 from twelve modules of 50 MWe to six modules of 77 MWe, a plant of 462 MWe. Capacity fell while unit size rose — the opposite of the series-production story. The target power price was $55/MWh in 2021 against a levelised target of $58/MWh in 2020 dollars, and became $89/MWh in January 2023, an increase of about 53%. Total capital cost moved from $5.3 billion to $9.3 billion, a 75% increase, for a plant that had shrunk. The purchaser-side advisory that recorded this states the implied figure as $20,139 per kilowatt. Federal support reached $4.2 billion in cost-share and resources, on top of roughly $600 million of Department of Energy support for small-reactor commercialisation since 2014 and a $1.35 billion ten-year programme approved in 2020 subject to appropriations. Thirty-three of fifty members signed in 2017; twenty-six of twenty-seven remaining participants voted to continue in February 2023; the project needed 80% subscription by February 2024 and was terminated on 8 November 2023. Two sources give the pre-revision price as $55 and $58/MWh; the difference is almost certainly nominal against 2020 dollars, and this brief reports both rather than resolving it.

Frontier The cause of that failure is contested, and one of the readings is fatal to the modular claim itself. NuScale's own communications framed it as a customer problem: the project “presented unique challenges that NuScale does not expect will be replicated with other customers.” A pro-nuclear advocacy post-mortem partly agrees — the buyer “lacked experience with nuclear technology, couldn't undertake cost risks on behalf of its customers, and operated in a market with cheap natural gas and growing wind deployment” — but the same analysis identifies a design problem: construction of the large pool in which the modules are submerged “incurred a large, fixed cost regardless of the number of modules,” which “rendered the design more expensive and less adaptable than other SMR alternatives.” That is the scale penalty the modular architecture was supposed to abolish, appearing inside the modular architecture. Against both, the Union of Concerned Scientists offered the structural reading: small modular reactors “will generate more expensive electricity than large reactors, which themselves are not economical to build.” This brief does not adjudicate; it notes that the fixed-pool finding is the one that generalises least comfortably for the sector.

Established What is genuinely operating is two plants, and the published performance of both is poor or unavailable. China's HTR-PM at Shidaowan is two 250 MWt high-temperature gas modules driving one turbine — critical in September and November 2021, on the grid in December 2021, at full power in December 2022, in commercial operation from December 2023, against a first concrete pour in December 2012. Its electrical rating is quoted three ways: 210 MWe in tertiary sources, a combined 200 MW in the plant's original nominal figure, and 150 MW after a derating of one quarter that the independent World Nuclear Industry Status Report describes as made “for unknown reasons.” The same report states that “there is little to no information available as to the operational experiences with the two HTR-PM modules.” Any load factor quoted for this plant is only as good as the denominator it was computed against, and this brief declines to reduce it to a single number.

Established Russia's Akademik Lomonosov, a floating twin-KLT-40S barge at Pevek, entered commercial operation in May 2020 and passed its first billion kilowatt-hours on 16 January 2025. Peer-reviewed work reports it was commissioned after thirteen years instead of the anticipated three, with cost rising from 6 billion roubles to at least 37 billion. One terawatt-hour over roughly 56 months against 70 MWe implies an average capacity factor near 35% — but Pevek has a population of about five thousand on an isolated Chukotka grid and the plant also supplies district heat, so that figure is substantially demand-limited and is not a reliability number. These figures are carried forward from the brief's earlier sourcing and could not be re-verified in the September 2026 research round; no new number has been added to them, which is why none appears.

Established The cleanest test ever run was a head-to-head on one site, and the small reactor lost it. At Changjiang on Hainan the same owner, regulator and labour market built two reactors at once. The 1,100 MWe Hualong One Unit 3 poured first concrete in March 2021 and reached first criticality on 10 July 2026. The 125 MWe ACP100 “Linglong One” poured first concrete four months later, on 13 July 2021, against a planned total construction period of 58 months which would have delivered it in about November 2025. It completed cold functional tests on 16 October 2025 and now projects commercial operation at the end of 2026, about 65 months from first concrete. The full-size unit, nine times larger, started first.

Established Argentina's CAREM-25 is the long-running cautionary tale and it belongs in the position rather than the footnotes. A 32 MWe integral pressurised-water design, under construction since February 2014 against an original 2017–18 generation target, reported at roughly 85% completion by union estimates, with work halted amid layoffs after it received no 2024 budget allocation. It has been under construction longer than most of the companies now announcing 2030 delivery dates have existed, and it does not generate. As with Lomonosov, these figures are carried forward from the brief's existing sourcing and were not re-verifiable in the latest research round; no 2026 status is asserted here.

Established What is genuinely under construction in the West is one reactor, and it is Canadian. Ontario Power Generation's first BWRX-300 at Darlington received a construction licence from the Canadian Nuclear Safety Commission on 4 April 2025, valid until 31 March 2035, with four facility-specific licence conditions and three regulatory hold points. Ontario approved the four-unit programme on 8 May 2025 at CAD 20.9 billion for 1,200 MWe. The first regulatory hold point was removed on 30 March 2026, releasing the reactor building foundation and civil construction, and OPG applied for a twenty-year operating licence on 25 March 2026. This is the leading Western case by a wide margin and the engineering section treats its cost structure in detail.

Frontier Darlington's schedule has already moved, in public, before the reactor pressure vessel is in. Reporting in April 2025 gave Unit 1 commercial operation by the end of 2029 and Units 2 to 4 in service between 2034 and 2036. By 2026 both OPG and trade press give grid connection by the end of 2030. That is a one-year slip visible on the public record inside eighteen months. This brief reports both and resolves neither.

Established One small modular design has ever been certified by the United States regulator, and its vendor no longer sells it. The independent status report's summary of the Western position is blunt: there is “only one Design Certification and one Standard Design Approval (both NuScale, U.S.) but no constructions.” The certification is the US600, approved 19 January 2023. NuScale's current product is the US460, which holds a Standard Design Approval from 29 May 2025 — a lesser instrument that does not authorise construction.

Handwave Every claim that the data-centre wave has ordered gigawatts of small reactors. The largest headline of the wave resolves into about a third firm, and the whole of that firm third is electricity from large reactors built in the 1970s and 1980s; the institutional section gives the contract-by-contract breakdown.

Frontier The central open question is whether series learning exists at all in nuclear construction, and the honest answer is that nobody knows, because no small modular programme has completed a second unit of a series. That is not a rhetorical framing. It is the literal state of the evidence: the sector has produced zero observations of the quantity on which its entire economic case depends.

Established The strongest peer-reviewed evidence points the wrong way, and it comes from the best case rather than the worst. Grubler's 2010 study in Energy Policy examined the French pressurised-water programme — “arguably the most successful nuclear scale-up experience in an industrialized country,” enabled by “centralized decision making, a high degree of standardization, and regulatory stability.” Working from previously unavailable public records covering absolute, annual and reactor-specific costs, it found that “even this most successful nuclear scale-up was characterized by a substantial escalation of real-term construction costs,” while operating costs stayed roughly stable. Its conclusion names the phenomenon: “negative learning” in which “specific costs increase rather than decrease with accumulated experience,” and therefore that “uncertainties in anticipated learning effects of new technologies might be much larger than often assumed.”

Notice exactly what that result is, because it is usually mis-stated in both directions. It is not a claim that nuclear construction cannot get cheaper. It is the observation that standardisation, centralisation and regulatory stability — the three conditions the modular case says it will supply — were all present in France and did not produce a positive learning rate. The small-reactor argument is that a fourth condition, factory serial production, will succeed where those three failed. That may be right. It is a bet against the only large-scale natural experiment anyone has run.

Frontier This brief cannot give Grubler's escalation multiple and does not estimate one. The paper's abstract carries the finding without figures; the publisher's page is disallowed to automated access and the author's institutional repository sits behind an anti-bot wall. A secondary contemporary account establishes only that the interim analysis covered 1979 to 2000 and that French escalation was more moderate than the American experience. An unverified number is left out rather than rounded to a memorable one.

Frontier The counter-argument exists and its rebuttal has not been retired. Lovering, Yip and Nordhaus (2016) found that nuclear construction costs have not universally escalated. A peer-reviewed reply by Koomey, Hultman and Grubler raises five objections: overnight costs were compared without financing or construction duration, so a slower build looks identical to a faster one; the dataset was not disclosed, making the result irreproducible; roughly €10 billion of French construction-related engineering and labour was excluded; 1950s and 1960s demonstration reactors were pooled with modern plants; and costs were aggregated across countries with incomparable accounting. Their conclusion is that only South Korea shows genuine declines, on data without independent audit. This brief cites the original only through the reply, because the original could not be reached.

Frontier How many units would it take? The sensitivity is the finding. At the conventional scaling exponent of 0.6, moving from 1,000 MWe to 300 MWe raises cost per kilowatt by 62%. Clawing that back needs 8 units at a 15% learning rate, 24 units at 10%, and 671 units at 5%. Those are not three similar answers. A two-thirds reduction in an unmeasured parameter moves the required order book by a factor of eighty-four, from a programme a single utility could plausibly commit to into something no market on earth is going to buy. The learning rate is doing all of the work in that sentence, and it is the one quantity for which the industry has no observation at all.

Frontier A genuinely open question that does not merely sound open: whether the fixed, non-modular fraction of a small plant is small enough for the argument to work. The submerged-pool finding from the cancelled American project is one instance. Darlington's CAD 1.6 billion of infrastructure and services common to all four units — roads, sewers, bridges, ancillary buildings, fibre lines and cooling-water tunnels — is another, and it is 21% of the first phase's CAD 7.7 billion. Modularity reduces the cost of the parts that can be made in a factory. It does nothing to the site, the licensing, the grid connection or the civil works, and the empirical question of what fraction that leaves is unanswered.

Frontier Also genuinely open: whether a buyer with unlimited capital and an urgent load changes the outcome. This was the thing the data-centre wave was supposed to settle, and so far the contracts suggest not, because the firm commitments have mostly gone to existing large reactors rather than to new small ones. The test is not whether hyperscalers sign paper; they have signed a great deal of it. The test is whether any of them takes first-of-a-kind construction risk on a small reactor, and as of September 2026 none has.

Handwave What merely sounds open: whether the regulator is the obstacle. The regulatory throughput argument is real but much weaker than its advocates suggest, and the licensing section gives the numbers. The American regulator completed the Hermes 2 construction-permit review in sixteen months, roughly four months ahead of schedule; the British regulators completed the fastest generic design assessment on record in about twelve months; the Canadian commission granted a construction licence about two and a half years after the application. What took nineteen years at Darlington was not the reactor review.

Established The first bottleneck is arithmetic, and engineering does not remove it. Small reactors cost more per kilowatt than large ones because most reactor cost scales with surface area and volume rather than with output. A 2023 techno-economic study decomposes the penalty directly: $1,490/kW for a light-water small modular reactor at the high end and $776/kW at the low end. Everything the sector promises is an attempt to earn that back through repetition.

Established And the promise does not close even when granted for free. The same study assumes nth-of-a-kind experience throughout — conceding the entire series-learning argument — and still finds a light-water small modular reactor at $4,844/kW against $4,599/kW for a large reference pressurised-water plant. At nth-of-a-kind, with learning already banked, the small reactor is still the more expensive machine per kilowatt. Its levelised cost lands at $89.6/MWh against $62.3/MWh for gas without carbon capture.

Established The second bottleneck is the licensing and environmental clock, and it precedes the factory entirely. Darlington's chronology is the cleanest available measurement: OPG applied for a site preparation licence in September 2006; the environmental assessment went to a joint review panel in March 2008; the panel reported in August 2011; a ten-year site preparation licence was issued on 17 August 2012; the technology partner was announced in December 2021; the construction licence application went in on 31 October 2022; hearings ran from October 2024 to January 2025; and the construction licence came on 4 April 2025. That is nineteen years and seven months from the first site application to permission to build. None of it is module fabrication and none of it is addressed by a factory.

Established The third is that first-of-a-kind site work eats the factory's advantage. Kairos Power's Hermes 1, a 35 MWt non-power test reactor, had its NRC construction completion deadline extended from 31 December 2026 to 30 April 2029 — about 28 months — on an amendment requested on 24 March 2026 and approved on 21 April 2026. The stated cause was “development delays associated with the first-of-a-kind nature” of the design and construction, together with site preparation and legacy structure removal that required significantly more work than planned. A factory-built design lost 28 months to a specific patch of ground.

Established The fourth is fuel, and it binds now rather than in principle. The Nuclear Energy Agency's dashboard finds 30 designs requiring high-assay low-enriched uranium at 10 to 20% enrichment and nine more planning 5 to 10%, against 39 planning standard uranium oxide. Its summary judgement is the one that matters: 47 of the 74 designs analysed in detail — more than 60% — rely on fuels not available at commercial scale. The only American producer had delivered 20 kg in November 2023 and 900 kg by 25 June 2025, from a cascade of sixteen machines. A full commercial cascade is 120 machines producing about 6,000 kg a year, roughly 42 months from funding, with further cascades every six months thereafter. The Department of Energy's own inventory target was 21 tonnes available by 30 June 2026 against a Congressional schedule of 3 tonnes by September 2024, 8 tonnes by December 2025 and 10 tonnes by June 2026 — and it projects domestic demand reaching 50 tonnes a year by 2035. That is roughly eight full cascades, none of which exists; the arithmetic is this brief's and the inputs are sourced. A market analyst quoted in trade press put the position more simply: “there is no market,” only “concepts of a market.”

Established The fifth is the enrichment base underneath the fuel, which is larger and less discussed. American reactor operators bought 15 million separative work units in 2024, of which 19% was United States origin, 81% foreign origin and 20% Russian. Urenco's Eunice plant runs 4.3 million SWU a year, expanding by 700,000 by 2027 and 2.1 million more from 2032 to 2036; Orano's Oak Ridge project is licensed for 7.4 million SWU with $900 million of federal funding and a regulatory decision due 30 April 2027. A projection cited in the same reporting puts 2050 American needs at 31.4 to 96.5 million SWU against 8.8 million of near-term domestic capacity. The dedicated British high-assay facility targets 27 tonnes a year, enough for “up to 30 advanced reactors,” with construction from 2028 and production in the early 2030s.

Established The fuel constraint has already moved a real schedule. TerraPower announced a minimum two-year slip from a 2028 in-service date in December 2022 because Russia's invasion of Ukraine removed “the only commercial source of HALEU fuel.” That is not a modelling result; it is a project date changing.

Frontier A note on the unit of analysis, because both sides abuse it. The scale penalty applies to the plant, not the module. A 462 MWe six-module plant carries a far smaller penalty than a single 77 MWe module would, and quoting the module figure overstates the problem exactly as quoting the plant figure understates the novelty. The honest version states which is which.

Established Nothing here waits on a physics result, and that is the most important thing about the dependency structure. Light-water small modular reactors use chemistry and fuel forms that have operated for six decades. What they wait on is a factory that has never been built, an order book deep enough to justify it, a licensing and environmental process that does not re-litigate each unit, and an enrichment industry that does not yet exist at the required scale.

Established The fuel dependency is the one that is currently binding, and it runs partly outside this brief. More than 60% of the designs tracked by the Nuclear Energy Agency depend on fuel unavailable at commercial scale. For the advanced designs among them — sodium-cooled, salt-cooled, gas-cooled, TRISO-fuelled — the fuel qualification questions belong to advanced fission, which owns the reactor physics and the fuel cycle. What belongs here is the industrial fact: twenty-one tonnes of high-assay material on the books to mid-2026, allocated by rationing to eight companies out of fifteen or more applicants, against a projected fifty tonnes a year of demand by 2035.

Established It depends on a licensing process it does not control and largely does not shorten. The Canadian, American and British regulators have all demonstrated that a reactor design review can be done in one to two years. The nineteen-year Darlington clock was environmental assessment, joint review panel, site licensing, technology selection and hearings. A modular factory changes none of those.

Established It depends on a buyer willing to carry first-of-a-kind cost. The cancelled American project shows what happens when that buyer is a consortium of small municipal utilities that, in the words of a sympathetic post-mortem, “couldn't undertake cost risks on behalf of its customers.” Darlington's buyer is a provincially owned utility with a rate base. That difference may matter more than any technical difference between the two designs.

Established What depends on it is firm low-carbon power near load, which several topics elsewhere in this corpus assume without naming a source. The waste stream depends on it too, in the specific sense that a fleet of many small cores produces a different spent-fuel geometry and a different decommissioning problem from a few large ones — a question that belongs to nuclear waste solutions and that this brief notes rather than answers.

Established The decisive experiments in this subject are construction schedules and procurement documents, not physics tests, and two of them are already running.

Established Changjiang is a controlled comparison no study could have designed. Two reactors, one site, one owner, one regulator, one labour market, four months apart, differing by a factor of nine in size and by the modular claim itself. The 1,100 MWe unit reached criticality first. The 125 MWe unit has now run about 65 months against a 58-month plan and completed cold functional tests in October 2025. The decisive follow-on is whether the second ACP100 unit is materially faster than the first, which is the only direct test of series learning anyone will get this decade and which no source consulted reports as having begun.

Frontier Darlington is the Western version and its readout arrives at the end of 2030. Unit 1 against a forecast Unit 4 cost 33% lower is a falsifiable claim with a date attached, made by a public owner under a provincial regulator, on a budget whose scope is published even if its contingency is not. If Unit 2 comes in materially below Unit 1 on a like-for-like basis, the modular argument will have its first genuine datum. If it does not, the negative-learning literature will have gained an observation from a programme designed explicitly to refute it.

Established A negative result already returned, and it is precise. Hermes 1's 28-month deadline extension is a first-of-a-kind construction failure on a 35 MWt test reactor, attributed by its own builder to the novelty of the work and to legacy infrastructure found during excavation. Modules are made in factories; sites are not.

Established Two further negative results are corporate rather than technical, and they should be recorded as results. Ultra Safe Nuclear Corporation, whose 15 MWe / 45 MWt micro modular reactor completed Canadian Phase 1 design review in February 2019, entered bankruptcy on 29 October 2024 after roughly $100 million of prior investment and $25 million of founder loans, when the death of its anchor investor in May 2024 left an $18 million raise unable to hold the structure together. Newcleo suspended its United Kingdom lead-cooled programme on 31 July 2025 and cut British staff from about 150 to a minimal team, six weeks after the design was accepted into generic design assessment. In both cases the regulatory rung was reached and the commercial rung was not. A design review is not a business.

Frontier The positive datum worth watching is that concrete is now being poured on novel designs in the West. Kairos broke ground on Hermes 2 in April 2026 at Oak Ridge, and TerraPower began construction at Kemmerer on 24 April 2026. Neither is a small modular reactor in the strict sense — Hermes 2 is a demonstration plant permitted at 70 MWt and 20 MWe, Kemmerer is 345 MWe rising to 500 MWe on storage — but they are the first Western non-light-water construction in over half a century and their outturn schedules are data.

Frontier The experiment that has not been run is the one that would settle the argument fastest: a genuine multi-unit factory order. No factory anywhere is producing modules against a firm order book of the depth the learning-rate models require. Until one is, every learning-rate figure in this subject is an assumption with a citation attached.

Established The deployment ladder has five rungs and almost every disputed claim in this subject is a slide between two adjacent ones. Stating them separately is the single most useful piece of engineering discipline available here, because the difference between a design review and a construction licence is not a matter of degree.

Established A vendor design review confers nothing, and the Canadian regulator says so in its own words. From the CNSC's executive summary of the GE Hitachi BWRX-300 review: “a VDR does not involve the issuance of a licence under the Nuclear Safety and Control Act and is not part of the licensing process,” and “the conclusions of the VDR do not bind or otherwise influence decisions made by the Commission.” That review consumed more than 200 vendor documents across 19 focus areas, concluded there were no fundamental barriers to licensing, and identified seven technical areas requiring further development — among them severe accident analysis, demonstration of two independent means of reactor shutdown, and radionuclide release restrictions. The agreement was signed on 11 December 2019 and the combined Phase 1 and Phase 2 review completed in March 2023.

Established The full Canadian review record since 2008 is eleven designs and one licence. Completed: AECL's ACR-1000 through three phases from 2008 to 2010; Westinghouse's AP1000, 2010 to 2013; Candu Energy's EC6 through three phases, 2010 to 2013; ATMEA1, Phase 1 in June 2013; Terrestrial Energy's IMSR400, Phase 1 in November 2017 and Phase 2 in April 2023; USNC's MMR, Phase 1 in February 2019; ARC Clean Technology's ARC-100, Phase 1 in October 2019 and Phase 2 in July 2025; Holtec's SMR-160, Phase 1 in August 2020; Moltex's SSR-W300, Phase 1 in May 2021; GE Hitachi's BWRX-300, March 2023; and X-energy's Xe-100, December 2023. Westinghouse's 5 MWe eVinci applied for Phase 2 in June 2023 and the review is on hold. One of those eleven designs is being built.

Established The Darlington cost structure, stated as the budget actually defines it. The approved CAD 20.9 billion covers four units and 1,200 MWe. Unit 1 is CAD 6.1 billion, plus CAD 1.6 billion of infrastructure and services common to all four, for a first-phase figure of CAD 7.7 billion. Unit 4 is forecast at CAD 4.1 billion, about 33% below Unit 1; Units 2 and 3 are not finalised and are expected to fall between. The programme's claimed electricity cost is about 14.9 cents per kilowatt-hour, compared by its proponents against 13.5 to 18.4 cents/kWh for wind, solar and battery combinations. The CNSC accepted a CAD 167.18 million letter of credit as the financial guarantee for decommissioning and waste management.

Established Read that budget's scope before comparing it with anything. The estimate covers “licensing, engineering, procurement, construction, operations readiness, contingency, interest and escalation” across a ten-year horizon. It is therefore not an overnight cost and cannot be set against the $/kW figures in the techno-economic literature. A contingency reserve is included and its size is unpublished. Comparing a fully loaded ten-year figure with an overnight cost is the most common numerical error made about this project in either direction.

Frontier The 33% decline is a forecast made by the owner, and it is the entire thesis under test. OPG states it will “apply learnings from the first unit's construction to deliver cost and time savings on subsequent units,” on the model of its own refurbishment programme. That is the learning-rate claim, made by the interested party, before any unit is complete anywhere in the world.

Established Construction has genuinely begun and the milestones are physical. As of mid-2026 the basemat foundation is installed, the reactor building is rising, the tunnel boring machine's main drive is in, and concrete for the condenser cooling-water system is nearing completion. The next regulatory hold point governs installation of the reactor pressure vessel. Whatever one thinks of the economics, this is not a rendering.

The closest neighbour is advanced fission, and the boundary between the two is worth stating because the two subjects are routinely merged. That brief owns reactor physics, coolants and fuel cycles: molten salt, thorium, fast spectra, breeding ratios and reprocessing. This one owns the delivery model: factories, modules, licensing throughput, order books and outturn cost. Many designs are both small and advanced, but the properties are independent — a novel coolant does not make a plant modular, and a modular plant need not be novel. The two briefs share one constraint outright, high-assay low-enriched uranium, and each takes a different half of it: the fuel qualification and reactor-physics half there, the industrial supply and rationing half here.

Nuclear waste solutions is adjacent downstream, and the relationship is more specific than the usual gesture. A fleet of many small cores produces more decommissioning events, more spent-fuel handling operations per gigawatt-hour and a different cask logistics problem than a few large plants, and several advanced small designs produce spent fuel in forms for which no disposal route is qualified. Nothing in this brief's economics accounts for that.

Commercial fusion is adjacent in the most instructive way available: it runs the same argument, one rung further back. Serial factory production, learning curves, modular units, private capital and confident dates — with the additional handicap that no fusion machine has yet produced net electricity at all. Anyone assessing the modular learning claim here should notice that the same claim is being made there with less evidence, and should apply the same discount.

Beyond the nuclear cluster: advanced nuclear propulsion shares fuel-cycle vocabulary and almost nothing else, since a propulsion reactor optimises thrust-to-weight rather than cost per kilowatt-hour; advanced battery technologies and planetary-scale energy systems treat small reactors as one candidate firm generator among several, and the 14.9 cents/kWh Darlington comparison against 13.5 to 18.4 cents/kWh for wind, solar and storage is exactly the comparison those briefs have to make; and energy corridors is the alternative answer to the same question, since a distributed fleet of small plants and a long transmission line are two ways of getting firm power to a load.

Established The 2025–26 institutional story is a wave of announcements whose contracted core is mostly not small and mostly not new. The genuine first is real and should be credited: on 18 August 2025 the Tennessee Valley Authority signed the first United States utility power-purchase agreement for a Generation IV reactor — 50 MWe from Kairos's Hermes 2, targeted for 2030, at market-based prices fixed by time of day, with Google taking the clean-energy attributes for data centres in Tennessee and Alabama. Financial details, contract length and risk allocation were not disclosed. The accompanying figure of up to 500 MW of advanced nuclear by 2035 comes from a master plant development agreement of October 2024 and is a framework, not an order.

Established Beyond that, the qualifiers do the work. Meta's 9 January 2026 announcement of “up to 6.6 GW by 2035” resolves into twenty-year agreements for 2,176 MW from three existing large plants — Perry, Davis-Besse and uprates at Beaver Valley — with purchases beginning in late 2026 and 2,609 MW online by 2034; the remainder is development frameworks with TerraPower for “up to eight” Natrium units and with Oklo for a 1.2 GW campus in Pike County, Ohio, first phase “as early as” 2030. Microsoft's 835 MW is a restart of Three Mile Island Unit 1, an existing large pressurised water reactor. Amazon's Cascade is 80 MWe per Xe-100 module, a 320 MWe first block and 960 MWe at full build, and Amazon holds a right to purchase rather than a power-purchase agreement, with the owner still preparing for a construction permit application that has not been filed. TVA's 6 GW arrangement with NuScale's commercialisation partner is a confidential memorandum of understanding with no timeline and no financial terms.

Established Where federal money has actually gone, it has gone to 300 MWe-class units at existing nuclear sites. On 3 December 2025 the Department of Energy selected two projects at $400 million each: TVA's Clinch River BWRX-300, and Holtec's two 300 MWe SMR-300 units at Palisades. Neither is a novel small design at a greenfield site, which is what the modular thesis actually proposes.

Established The Canadian institutional arrangement is the most legible in the world and that is itself the finding. A provincially owned utility, a federal regulator conducting a five-day public hearing in the host municipality, a published licence with four facility-specific conditions and three regulatory hold points, a CAD 167.18 million letter of credit for decommissioning, a published budget scope, and a twenty-year operating licence application filed on 25 March 2026 that will itself go to public hearing. The owner's claims — about 3,700 jobs sustained annually over a 65-year life, and a GDP contribution given as CAD 35.1 billion in one place and CAD 38.5 billion in another — are the interested party's and are not audited here. But the process is inspectable in a way that no other leading project is, and that is the better explanation of why the leading Western case is Canadian than anything about the reactor.

Established The Canadian pre-licensing institution has thrown off a clean statistic. Eleven vendor design reviews completed since 2008, one on hold, one licence to construct. Two of the eleven vendors have since had a corporate failure. The vendor design review is an excellent institution for finding out whether a design is licensable and a useless one for finding out whether anything will be built.

Established Britain repeats the pattern. Four designs have completed full generic design assessment and the BWRX-300 completed the first two-step assessment on 12 December 2024, with regulators noting that no site had been identified and no deployment was planned. Rolls-Royce won the national competition and signed a contract on 13 April 2026 for three units totalling at least 1.4 GWe at Wylfa, explicitly ahead of a future final investment decision, with no construction start, no first-power date and no per-unit price disclosed; its own per-unit figure has moved from about £1.8 billion to £2–3 billion without a unit being built.

Frontier The institution that does not exist is a repeat buyer. Not a signatory of frameworks — there are many — but an entity that has ordered a second identical unit and is paying for the first while the second is fabricated. Every learning-rate model in this subject assumes one. None of the world's actual buyers is one: not a municipal consortium that could not carry the risk, not a hyperscaler holding an option, not a national programme that has bought three units ahead of a final investment decision. Ontario, with four units approved and one under construction, is the closest thing that exists to that institution anywhere on earth, which is why the end of this decade at Darlington is the date that matters.

Established And the regulator has moved faster than the market. The American regulator's new technology-inclusive licensing framework was issued on 25 March 2026 — the first new initial-licensing regulation since 1989 — nearly two years ahead of its statutory deadline, under legislation enacted in July 2024. Fuel fabrication moved too: the first American Category II fuel facility licence for high-assay TRISO was issued on 13 February 2026. The bottleneck that these reforms were built to clear was real. It was also not the one that killed the only Western project to get near construction, which died on price with a valid design certification in hand.

Frontier The distinctive ethical question is not radiation risk, which is governed by a mature and conservative regime, but who pays for first-of-a-kind cost and who captures the benefit if learning arrives. The cancelled American project is the clean case: $4.2 billion of federal support and cost-share, a consortium of small municipal utilities carrying the escalation, and subscription failure ending it. The price rose 53% and the capital cost 75% before a module was built; the participating towns bore the development cost of a learning curve that was never traversed. Ratepayers and taxpayers were the risk-bearers, and the vendor's public framing afterwards was that the customer was unusual.

Established Public money and public accountability are unevenly matched across this sector, and the pattern is worth naming. Darlington is publicly owned, its budget scope is published, its regulator holds public hearings, and the construction licence is a public document with named conditions. The Romanian project, by contrast, took a conditional final investment decision whose conditions are sealed under an eight-year confidentiality agreement, disclosed only to shareholders on restricted terms, with no published capital cost — while carrying letters of interest from an American export credit agency contemplating up to $3 billion. Public credit and private confidentiality is a defensible commercial arrangement and an uncomfortable accountability arrangement, and the brief records it as both.

Established Evidence quality in this subject is structurally poor and the reader should be told why. A large share of what is publicly known about schedules, costs and orders comes from vendors, owners, an industry association's trade press and government agencies with programmatic commitments. This brief marks interested parties in its reading list. The pattern to watch for is not fabrication — shipped milestones are generally accurate — but framing: capacity announcements presented as orders, options presented as contracts, design reviews presented as approvals, and headline figures whose firm fraction is buried. Where an independent critical source exists, this brief prefers it, which is why the derating of the only operating high-temperature plant and the count of Western design certifications are taken from the status report rather than from the industry.

Established A disclosure obligation this brief takes on itself. The September 2026 research round for this brief lost its search allocation to a workspace-wide quota after five queries, and the retrieval tool would only open pages already surfaced by a search. Several sources this brief carries — the peer-reviewed work behind the Akademik Lomonosov figures, the trade reporting behind CAREM-25, the head-to-head Changjiang report, two regulator pages and the British contract announcement — could not be re-opened in that round. They are retained with their existing attribution and no new claim rests on any of them. The relevant paragraphs say so where the figures appear. Stating this is preferable to letting a reader assume every number was checked on the same day.

Frontier Four questions this brief cannot answer and does not pretend to. What is Linglong One's outturn cost? No figure was found in any source consulted; the 58-month plan is public and the budget is not. What is HTR-PM's load factor, and against which of three nameplate ratings? The independent status report says the operational information is not available and the rating was cut by a quarter for unpublished reasons. What is the size of Darlington's contingency reserve, which sits inside a publicly approved CAD 20.9 billion? Unpublished. How much high-assay fuel did each of the eight allocated companies actually receive? Undisclosed in both allocation rounds. Each of those is a number a public decision was made against, and in each case the public cannot audit it. That is an accountability finding, not a research failure, and it is the reason it appears in this section rather than in a footnote.

Frontier The civilisational case for small modular reactors is firm, dispatchable, low-carbon power on a footprint small enough to site near load — and, in the version this map cares about, a reactor small and repeatable enough to be a component rather than a project. That is a genuine prize. It would make deep decarbonisation of industrial heat and of high-latitude grids substantially easier, and it is a precondition assumed by several topics elsewhere in this corpus without acknowledgement.

Established The general principle this case illustrates is about where constraints actually live, and it is the same principle that runs through the superconductor and fusion briefs. Nothing about small reactors is blocked by knowledge. The physics has been settled for six decades; the materials are qualified for light-water designs; the regulators have demonstrated that they can review a novel design in twelve to sixteen months. What blocks deployment is a manufacturing base that does not exist, an order book that cannot form until a learning rate is demonstrated, a learning rate that cannot be demonstrated until an order book forms, and a nineteen-year public-process clock that sits in front of all of it. A capability can be fully understood, fully licensed in principle, and still not built, and the reasons will be industrial, institutional and financial rather than scientific. This corpus's recurring question — what stands between a demonstrated capability and a built world — is answered here by order books and environmental assessments, not by physics.

Established The negative-learning result has a significance well beyond nuclear power, and it is the reason this brief matters to readers who do not care about reactors. A great deal of industrial policy, in energy and elsewhere, assumes that volume reliably drives cost down. The French programme supplied every condition that assumption requires — standardisation, centralisation, regulatory stability, dozens of units — and produced real-term cost escalation instead. Where the cost driver is not a function of throughput, volume does not deliver what volume usually delivers. Any policy that treats a learning rate as a property of a technology rather than as an empirical question about a specific industry is assuming a mechanism that this industry, on its best evidence, does not have.

Speculative The opportunity-cost question is the honest civilisational counterweight and it remains open. Capital and skilled construction labour spent on first-of-a-kind small reactors are not spent on transmission, storage, or large reactors of a proven design. The Darlington comparison — about 14.9 cents per kilowatt-hour against 13.5 to 18.4 cents for wind, solar and battery combinations — is the owner's own framing and puts the two options within a factor that is well inside the uncertainty on a first-of-a-kind estimate. If the learning arrives, the comparison moves decisively one way. If it does not, a great deal of decarbonisation capital will have been spent proving a negative.

• Already happened: Established Darlington's site application went in in September 2006 and its construction licence came in April 2025 — nineteen years and seven months. The NuScale–UAMPS project ran from 2015 to termination on 8 November 2023 with price up 53% and capital up 75%. CAREM-25 has been under construction since February 2014. The previous American construction permit for a non-light-water reactor before Hermes 2 in November 2024 was issued in 1968.
• 2–3 yr: Established Linglong One reaches commercial operation, currently targeted for the end of 2026 against a 58-month plan that expired in November 2025. Hermes 2 and Kemmerer continue building. Darlington Unit 1 approaches an end-2030 grid date that was an end-2029 commercial operation date in April 2025. First-of-a-kind outturn costs start becoming observable rather than forecast.
• 5 yr: Frontier The first genuine second-unit comparison — Darlington Unit 2 against Unit 1, or Linglong One's follow-on — produces the sector's first measured learning rate. This is the single date that matters, and everything in the economics section is conditional on it.
• 10 yr: Handwave A factory is either producing modules against a multi-unit order book, or the sector has consolidated around two or three designs delivered as conventional projects. Both outcomes are live; the second is the historical base rate. Dates in this row are flagged handwave because the sector's published dates have slipped in every case this brief examines — 58 months to 65 and counting at Changjiang, 28 months on a test reactor, one year at Darlington inside eighteen months, thirteen years against three at Pevek.
• 25 yr+: Speculative Small modular reactors as a routine grid component rather than a programme, conditional on a learning rate nobody has yet measured and on an enrichment industry roughly eight times its currently demonstrated scale for high-assay fuel alone. The alternative long-run outcome, equally live, is that the sector's durable contribution turns out to have been the restart and uprate of large reactors that the data-centre wave financed while the small ones were being designed. Nothing in the record to date makes that the less likely of the two.

• Depends on Nothing on this map, and that remains the finding: no result in any brief here is on this technology's critical path. Light-water small modular reactors need no pending physics. The constraints are industrial, institutional and financial, and the row below is where they are recorded.
• Requires (not on this map) A factory that has never been built at the required throughput; an order book deep enough to justify it, whose depth depends on a learning rate with zero observations; a licensing and environmental assessment process that does not consume nineteen years before first concrete, as it did at Darlington; high-assay low-enriched uranium capacity at roughly eight times what has been demonstrated, for the sixty percent of designs that need it; a first customer willing to underwrite first-of-a-kind cost, which the cancelled American project shows a municipal consortium cannot be; and the nuclear-grade forging and module fabrication supply chain underneath all of it. All six are industrial, institutional or market capabilities rather than scientific results, which is why none of them is a brief on this map — and why recording them here rather than nowhere is the point of the row.
• Enables Firm low-carbon power near load, which several Category VII topics assume without naming a source. No typed enabling edge is claimed: the assumption in those briefs is for firm power generally, not for this technology specifically, and typing it would overstate what they depend on.
• Adjacent Advanced fission, which owns the reactor physics and fuel cycles this brief deliberately excludes and which shares the high-assay fuel constraint with it; nuclear waste solutions, which owns what comes out; commercial fusion, which runs an almost identical series-production and factory-learning argument at an earlier stage and with even less deployment evidence; and lunar surface power, where the small-reactor case is strongest precisely because there is no grid to compare against.

“Small modular reactors are cheaper than large reactors.” Handwave They are cheaper per plant and more expensive per kilowatt, and the second determines the price of electricity. At nth-of-a-kind, granting the learning argument in full, a light-water small modular reactor still models at $4,844/kW against $4,599/kW for a large reference plant, with levelised cost at $89.6/MWh against $62.3/MWh for unabated gas. The scale penalty at the high end of the same study is $1,490/kW. The cheapness claim is true of the cheque and false of the electricity.

“Costs come down as you build more — that's how manufacturing works.” Established It is how much manufacturing works and it is not how nuclear construction has worked. The only detailed study of a large standardised national programme — France, with centralised decision-making, high standardisation and regulatory stability — found substantial real-term construction cost escalation and named the phenomenon negative learning, in which “specific costs increase rather than decrease with accumulated experience.” The counter-study finding no universal escalation has a peer-reviewed reply against it that has not been retired, and the reply's conclusion is that only South Korea shows genuine declines, on unaudited data. This brief does not claim learning is impossible. It claims the prior is against it and the sector has produced no observation either way.

“The regulator is the bottleneck.” Handwave Not on the evidence. The American regulator issued the Hermes 2 construction permit in sixteen months from acceptance, about four months early; British regulators completed the BWRX-300 assessment Step 2 in about twelve months, the fastest such engagement on record; the Canadian commission issued a construction licence about two and a half years after application. What consumed nineteen years at Darlington was environmental assessment, a joint review panel, site licensing and technology selection — a public-process clock, not a safety-review clock. Reforming reactor licensing addresses the shorter half of the problem.

“The design has been approved by the regulator.” Established Almost always this means a vendor design review or a generic design assessment, which the Canadian regulator describes in its own words as something that “does not involve the issuance of a licence” and “is not part of the licensing process,” whose conclusions “do not bind or otherwise influence decisions made by the Commission.” The British equivalent for the BWRX-300 concluded there were no fundamental shortfalls and stated that no site had been identified, no plans existed to deploy it, further assessment would be needed before safety-significant construction, and environmental permits could not be issued. Eleven designs have completed Canadian vendor design reviews since 2008 and one is being built.

“The NRC has certified several small modular designs.” Handwave It has certified one, NuScale's US600 in 2023, and the independent status report's summary of the Western position is that there is “only one Design Certification and one Standard Design Approval (both NuScale, U.S.) but no constructions.” The vendor has since moved to a different product, the US460, holding the lesser Standard Design Approval from 29 May 2025, which does not authorise construction.

“The hyperscalers have contracted gigawatts of new small reactors.” Handwave The firmly contracted portion of the 2025–26 wave is mostly power from existing large reactors. Of Meta's January 2026 headline of up to 6.6 GW by 2035, the firm component is twenty-year agreements for 2,176 MW from Perry, Davis-Besse and uprates at Beaver Valley, rising to 2,609 MW by 2034 — about a third of the headline, and all of it electricity from plants built decades ago. Microsoft's 835 MW deal is a restart of Three Mile Island Unit 1, a 1970s pressurised water reactor, and is not a small modular reactor at all. Amazon's Cascade gives Amazon a right to purchase from a first block of 320 MWe — four Xe-100 modules at 80 MWe — with no construction permit application yet filed with the regulator. The genuinely new small-reactor commitment that has reached a signed utility power-purchase agreement is 50 MWe.

“Factory production removes construction risk.” Frontier The clearest datum runs the other way: a 35 MWt test reactor slipped its construction deadline by about 28 months, attributed by its builder to the first-of-a-kind nature of the work and to legacy infrastructure found during excavation. And the cancelled American project contained a fixed cost that modularity could not touch — the submerged-reactor pool, which “incurred a large, fixed cost regardless of the number of modules.” Modules are made in factories; sites, pools, tunnels and grid connections are not. Darlington's own common infrastructure is CAD 1.6 billion against a CAD 7.7 billion first phase.

“A conditional final investment decision is a final investment decision.” Established NuScale's Romanian project took a conditional FID on 12 February 2026 for a 462 MWe six-module plant at a former coal site. The conditions sit in two lists and shareholder documentation states that “the fulfilment is the basis for the feasibility of the project and which, consequently, represent mandatory requirements.” They were disclosed to investors under an eight-year confidentiality agreement and are unpublished. No capital cost has been published. It is a commitment to proceed if undisclosed things happen.

“China's small reactor shows it can be done quickly.” Frontier China's two took 132 months against a 50-month plan and about 65 months against a 58-month plan without yet reaching operation, on the same sites and with the same supply chains that deliver its large reactors on schedule — including one on the very same site that poured concrete four months later and went critical first. If construction speed were the modular advantage, it would show up there first.

“Canada's Darlington cost works out to $X per kilowatt.” Established It does not, and this is the most common numerical error made about the leading Western project. The CAD 20.9 billion covers “licensing, engineering, procurement, construction, operations readiness, contingency, interest and escalation” over a ten-year horizon. It is a fully loaded programme figure, not an overnight cost, and dividing it by 1,200 MWe produces a number that cannot be compared with any $/kW figure in the literature. The contingency reserve is included and its size is not published.

“A design review means the company is viable.” Established Ultra Safe Nuclear Corporation completed a Canadian Phase 1 review in February 2019 and filed for bankruptcy on 29 October 2024. Newcleo entered British generic design assessment in June 2025 and suspended its United Kingdom programme on 31 July 2025. Regulatory progress and commercial survival are close to independent variables.

“Megawatts are megawatts.” Established They are not, and the conflation of thermal and electrical output is the characteristic error of this subject. HTR-PM is two 250 MWt modules and somewhere between 150 and 210 MWe depending on which rating one accepts. Hermes 2 was permitted at 70 MWt and 20 MWe and is now described as up to 50 MWe following an uprate — both figures correct, referring to different things. A specialist outlet reported four 300 MWe units at Darlington as 4,800 MW. Any megawatt figure in this field without an e or a t after it should be treated as uninterpretable.