The State of Nuclear Power—The Machine, the Accidents, the Fleet, the Costs, the Small Reactors, the Waste, and the Data Centers
Nuclear Power in 2026: The Operating Record, the Cost of Building New, the Small Reactors, and the Data Centers — Special Report
A nuclear power plant makes electricity the way a coal plant does, by boiling water and turning a turbine; only the furnace is different. This report files what nuclear power is and what it has done, from the record, at November 2026: the machine and the physical fact that governs its accidents, which is not the chain reaction but the heat that remains after the reaction stops; the three accidents that set the terms of every argument since, with their causes and their death counts kept in the categories the record keeps them in; the American fleet of 94 reactors, which runs better than it ever has and supplies a shrinking share of a growing country’s electricity; the newest American build, promised at $14 billion and delivered at about $31 billion, seven years late; the two price instruments that frame the economics and do not agree with each other; the small modular reactors, two of which operate anywhere on earth, both built by states, neither in North America; the data centers, for which the record’s plain answer is that no small reactor powers one and none is under construction to do it, while every contracted nuclear kilowatt a technology company actually holds comes from a large reactor that already exists or is being restarted; the spent fuel, which accumulates at about 2,000 tons a year into a national inventory with no operating repository beneath it; and the claims now in circulation—about prices, dates, waste-burning, and thorium—that the record, checked claim by claim, does not support. Where the record does not agree with itself, the report prints the disagreement and does not average it. The companion filing on the weapons prohibition is this morning’s report, The Nuclear Taboo; this one is about power.
The File in Brief
The machine first, because everything else in this file is downstream of it. A nuclear reactor is basically a steam engine—with atoms instead of coal. Fission is only a heat source: it boils water, the steam turns a turbine, and the turbine turns a generator, in the same Rankine cycle a coal plant runs. Because the water in a light-water reactor must stay liquid under pressure, its steam is cooler than a fossil plant’s, and the fleet converts about a third of its heat to electricity—32.7 percent on the Energy Information Administration’s 2024 heat-rate table, against about 44 percent for the gas fleet and about 32 percent for coal on the same instrument. The governing hazard is not the chain reaction, which stops when control rods go in; it is decay heat, the roughly 6 to 7 percent of full power that fission products keep generating after shutdown, still about 1 percent hours later. Every serious accident in the record is, at its physical core, a failure to remove that residual heat. The rest of this report—the fleet, the costs, the small reactors, the waste—is the record of what that machine has done in seventy years of commercial operation.
The accidents are three, they are different from each other, and their numbers are kept in separate ledgers in this report because the record keeps them separate. Three Mile Island in 1979 was a stuck valve, a misleading indicator, and a warning from an identical failure eighteen months earlier that the industry did not circulate; about half the core melted, the average dose to two million people was about one millirem, and no acute deaths occurred—the cancer question remains contested between two bodies of research that turn on a dose reconstruction nobody can redo. Chernobyl in 1986 was a reactor design whose physics ran the wrong way—power rising as its coolant boiled—with no containment building; thirty men died acutely, thyroid cancers in the exposed young number in the thousands, and eventual-death estimates run from documented counts under one hundred to a modeled 4,000 to an alleged 985,000, each printed here with its grade. Fukushima in 2011 was a plant that survived the earthquake and drowned in the tsunami: three meltdowns, no documented radiation-attributable health effects among residents on the United Nations’ assessment, and 2,351 certified disaster-related deaths in Fukushima Prefecture by May 2026, a Japanese administrative category covering the whole disaster’s disruption, overwhelmingly of the elderly—a real toll, in a different ledger from radiation, and not to be merged with it in either direction.
The fleet that exists works; the arithmetic around it is what moves. The United States operates 94 reactors at 54 plants in 28 states, average age about 44 years, and in 2025 they ran at a 91.0 percent capacity factor and generated 784,781 gigawatt-hours—slightly more electricity than the year before, and a smaller share of the nation’s total, 17.7 percent against 19.5 percent earlier in the decade, because total generation grew around them, not because nuclear output fell. Five American reactors have entered service since 1993. The last two, Vogtle Units 3 and 4 in Georgia, were promised at about $14 billion for service in 2016 and 2017 and were delivered at about $31 billion in owners’ costs, in service in July 2023 and April 2024; the Georgia Public Service Commission capped what ratepayers could be charged for Georgia Power’s share and left about $2.6 billion with shareholders. The fleet is meanwhile being extended rather than replaced: about 8,030 megawatts of capacity uprates approved since the 1970s, 86 of 94 reactors licensed to sixty years, 25 units at 13 sites approved to eighty, and three shuttered plants—Palisades, Crane, Duane Arnold—in funded restart programs, none yet back in service.
The economics split cleanly in two, and the two price instruments disagree in ways this report prints rather than splits. Power from reactors that already exist is cheap at the margin: Lazard’s 2026 study puts existing nuclear at $26 to $36 per megawatt-hour, the lowest firm generation on its page. Power from reactors not yet built is the most expensive entry in the same study: $175 to $255 per megawatt-hour for new nuclear, at capital costs of $12,300 to $17,570 per kilowatt that Lazard states are derived from Vogtle. The Energy Information Administration’s forward model, pricing a plant entering service in 2031 with federal credits applied, arrives at $87.81—the two instruments measure different things (one realized cost, one modeled cost; one before most support, one after credits), and the bridge between them is printed in section 4. The construction record underneath both: France’s Flamanville finished about twelve years late at €13.2 billion before financing; Britain’s Hinkley Point C stands at £35 billion in 2015 prices with first power now expected in 2030; Finland’s Olkiluoto 3 was delivered on a fixed-price contract whose overrun fell on the vendor; Korea’s four-unit Barakah plant in the Emirates entered service across 2021 to 2024 on a 2009 contract of about $20 billion, its audited final cost not found; and Chinese builds average 6.2 years against a global mean of 9.9, with no primary-sourced Chinese cost per kilowatt in existence that this research could find.
The small reactors are a licensing story, not yet a power story; and the data-center answer is plain. Two small reactors of modern design operate anywhere in the world: Russia’s floating Akademik Lomonosov and China’s HTR-PM, both state-built, neither publishing a cost of power. No new-design small reactor sends power to any North American grid. The one fully priced American test of the model, NuScale’s Carbon Free Power Project in Idaho, watched its target price climb from $55 to $89 per megawatt-hour—already net of roughly $30 of federal support—and its cost estimate from $5.3 billion to $9.3 billion, subscribed a quarter of the capacity it needed, and was terminated by its buyers in November 2023, killed by price and exit, not by the regulator; the design the Nuclear Regulatory Commission certified is not the design now being sold. The year 2026 produced the first American construction permits of the new wave—TerraPower’s Natrium in March, the Tennessee Valley Authority’s BWRX-300 at Clinch River in September—which are construction permits under the old two-step law, not operating licenses, and no licensed case in the record runs from application to operating commercial power in under a decade. And the commissioned question, answered as the research answers it: no. No small-modular-reactor-powered or microreactor-powered data center exists, and none is under construction, anywhere in the world, as of November 2026; delivered small-reactor capacity to data centers is zero megawatts. Every nuclear kilowatt actually contracted to a named American data-center buyer comes from an existing large reactor or a restart—Microsoft at Crane, Google at Duane Arnold, Amazon beside Susquehanna, Meta at Clinton—while every small-reactor arrangement in the file is a future instrument: a master agreement, a memorandum, a prepayment, a power purchase agreement for a plant that is not licensed. One narrow footnote, printed where the record prints it: Kairos Power’s Hermes 2 test reactor in Tennessee, ground broken in April 2026, has its planned output contracted toward the Google arrangement through TVA; it is a demonstration reactor of about 35 megawatts thermal, not a data-center power plant, and it does not change the answer.
The waste is small in mass, permanent in custody, and unburied in America. The United States holds 95,117 metric tons of spent fuel at 73 sites as of the end of 2024, growing about 2,000 tons a year, a little more than half of it in dry casks. One year of American coal ash outweighs one year of spent-fuel growth about 35,000 to one, and the entire seventy-year spent-fuel stock equals about a day of the coal-ash stream by mass—a comparison printed in this report with its caveat attached, that the streams are not hazard-equivalent per ton. The licensed repository in law remains Yucca Mountain, its statutory cap of 70,000 tons already exceeded by the inventory, its licensing proceeding suspended since 2011; two commercial interim-storage licenses survived a Supreme Court procedural ruling in June 2025, after which one project was cancelled and the other stands unbuilt, and the number of tons in consolidated interim storage is zero. Finland’s Onkalo received its operating license in October 2026, the world’s first for a spent-fuel repository. Regulation, meanwhile, has measurably changed and has not changed in the way either side’s slogan says: the ADVANCE Act of 2024 and the executive orders of May 2025 cut advanced-reactor licensing fees by more than half, added a deployment clause to the Nuclear Regulatory Commission’s mission beneath a safety clause that remains first, produced a new optional licensing framework in force since April 2026, and coincided with the agency losing about 500 staff net and its first-ever removal of a sitting commissioner. Public opinion, on the longest-running instrument, is at its second-highest reading ever: Gallup measured 61 percent in favor of nuclear energy in March 2025, and did not ask the question in 2026.
1. The Machine: A Steam Engine, with Atoms Instead of Coal
A commercial nuclear power plant does not generate electricity from fission. It generates heat from fission and electricity from steam. Inside the core, uranium-235 nuclei split, releasing energy as heat and the neutrons that sustain the chain reaction; control rods absorb neutrons to regulate the rate. The heat passes to water, the water becomes steam, and the steam drives a turbine coupled to a generator—the Rankine cycle, identical in kind to the back half of a coal or gas steam plant, with the reactor substituted for the furnace. That is the whole machine, and the Nuclear Regulatory Commission’s own teaching materials state it in those terms. The consequences of the substitution run through everything that follows: the reactor is the most elaborately regulated furnace in industrial history, and the power plant behind it is ordinary.
The two reactor types that make up the American fleet differ only in where the steam is made. In a pressurized water reactor, the primary loop is held at high pressure so its water does not boil in the core; it carries heat to steam generators, where a separate secondary loop boils, so the turbine hall sits outside the primary radiological boundary at the cost of large, expensive steam generators. In a boiling water reactor there is no steam generator: water boils inside the vessel itself, and the steam goes directly to the turbine—simpler, with fewer components, and with the turbine side inside the radiological envelope because it carries primary coolant. Both are cooled and moderated by ordinary water. Both run on steam whose temperature is capped by the need to keep that water liquid under pressure, and the cap is what sets the efficiency of the entire technology.
The efficiency record, on one instrument run against every fuel in the same year: the Energy Information Administration’s 2024 heat-rate table puts the American nuclear fleet at 10,443 British thermal units per kilowatt-hour, a thermal efficiency of 32.7 percent. The same table puts the gas fleet at about 44 percent and the coal fleet at about 31.7 percent. For every unit of electricity a light-water reactor delivers, roughly two units of fission heat are rejected to a river, a cooling tower, or the sea. Cause, manner, baseline: the figure is caused by the steam conditions a light-water reactor can safely produce, measured as fuel heat in per net kilowatt-hour out across the operating fleet, and its baseline is the fossil fleets measured identically. It is a materials and thermodynamics constraint, not an operating failure—and it is why advanced designs in this file chase higher temperatures, and why the waste-heat stream of a nuclear plant is proportionally the largest in the industry.
The second physical fact is decay heat, and it is the one that matters for safety. When a reactor shuts down, the chain reaction stops; the heat does not. The fission products accumulated in the fuel keep decaying, generating about 6 to 7 percent of full thermal power in the first moments after shutdown, about 1 percent within hours, about half a percent after a day. One percent of a gigawatt-scale reactor’s thermal rating is tens of megawatts of heat that must be carried away, continuously, by systems designed for exactly that duty in a plant that may have no other power. Loss of decay-heat removal—not loss of the ability to stop fission—is the initiating physical problem in both Three Mile Island and Fukushima, and it is the reason a shut-down reactor is not a cold one. Every safety system described in the rest of this report exists, in the end, to answer this one curve.
The same machine, packaged for ships, holds the longest operating record in the file. American naval reactors are pressurized water reactors with long-life cores, and the program that runs them, Naval Reactors, is its own safety authority: it reports, in its own program book with data as of December 2024, more than 177 million miles safely steamed on nuclear power, 97 reactors currently operating, and more than 7,600 reactor-years of accumulated operation. At the program’s seventy-fifth anniversary in July 2023, the Navy reported 273 reactor plants operated historically and 562 cores taken critical. The program’s standing statement is that it has never had a reactor accident, in its own definitions; two nuclear submarines, Thresher in 1963 and Scorpion in 1968, were lost with all hands for causes the record does not attribute to their reactors. The desk grades the naval figures documented—they are the operator’s own account, and no independent verification ledger exists—and notes, in section 11, what does and does not follow from them for civilian power: the fuel is weapons-grade material the service does not publicly specify, the crews are military, the reactors answer to no licensing authority and no ratepayer, and the equivalence sometimes drawn from the naval record to commercial small reactors appears in no primary source this research found.
One term in circulation is resolved here because the commission asked and the record answers plainly. There is no reactor type called a “seed reactor.” In the technical vocabulary, a seed is the enriched region of a seed-and-blanket core—a fissile center surrounded by a fertile blanket in which new fuel breeds. The configuration was instantiated at Shippingport in 1957, the first American commercial station, and tested to completion in the Light Water Breeder Reactor that ran in the same vessel from 1977 to 1982 on thorium and uranium-233. Whether that experiment’s final assay confirmed breeding—a ratio above one—could not be verified to a primary source in this research and is filed as not found. The term survives in modern marketing detached from that meaning; the record’s meaning is the one printed here.
2. The Accidents as Baseline: Three Mile Island, Chernobyl, Fukushima
Three accidents set the terms of every subsequent argument about this machine, and they are unlike each other in mechanism, in cause, and in consequence. The desk files them as the baseline of the record—what the technology’s worst days actually contain—before any question of viability is entertained, and keeps their numbers in the categories their sources keep them in.
Three Mile Island Unit 2, in Pennsylvania, on March 28, 1979, began in the non-nuclear half of the plant: a failure in the condensate system tripped the feedwater pumps and the turbine, and the auxiliary feedwater valves that should have answered had been left closed after maintenance. The steam generators boiled dry; heat removal from the core was lost. The principal cause, on the Kemeny Commission’s finding, was a pilot-operated relief valve that opened as designed and then stuck open, draining primary coolant for hours while its control-room indicator showed the signal sent to the valve, not the valve’s position—operators read a closed valve where an open one was emptying the system. Roughly half the core eventually melted. Two facts belong beside the mechanism. The same valve type had failed eleven times before; and eighteen months earlier, at Davis-Besse in Ohio, the identical sequence had begun and been caught, and the lesson was never disseminated through the industry—the organizational failure inside the technical one. The health record: an average dose of about one millirem to the roughly two million people in the area, a maximum under 100 millirem, and no acute deaths. The cancer question is genuinely contested: the Columbia University studies under Maureen Hatch found no link; a 1997 re-analysis by Steve Wing and colleagues reported increases along the plume pathway; the two turn on a dose reconstruction whose validity cannot now be re-established, and the bridge between them is unknown and printed as such. About 144,000 people left voluntarily at the peak; cleanup ran to about $1 billion across 1979 to 1993. No new American reactor order survived the decade after, a fact filed in section 3 with the ordering record that surrounds it.
Chernobyl Unit 4, in Ukraine, on April 26, 1986, was the opposite failure: not a loss of cooling in a sound design but a design whose physics amplified the accident. The RBMK’s moderator is graphite, which remains when coolant boils away; the water in an RBMK acts substantially as an absorber, so boiling raises reactivity—a positive void coefficient that the IAEA’s INSAG-7 reassessment quantified at roughly plus four to five beta on the night, voiding alone sufficient to add several times the reactivity of prompt criticality. Its control rods carried graphite tips that briefly added reactivity on insertion—a positive scram, in the record’s term. There was no containment building. During a turbine-rundown test at low power, the coefficients drove a prompt-critical excursion; steam explosions destroyed the reactor and displaced its roughly 2,000-ton shield; the exposed graphite burned for about ten days, lofting the release to altitude and across a continent: about 14 exabecquerels in total, including some 85 petabecquerels of cesium-137 and 1,760 petabecquerels of iodine-131. The death record, by category: two workers killed by the explosion and 28 more dead of acute radiation syndrome among 134 confirmed cases, thirty acute deaths in all; more than 6,000 thyroid cancers by 2005 among those exposed young, 15 of them fatal; UNSCEAR counted 19,233 thyroid cases across 1991 to 2015 and attributes roughly a quarter to the accident. The Chernobyl Forum’s projection of up to about 4,000 eventual cancer deaths is a model, printed as one. The figure of about 985,000 deaths circulated by Yablokov and co-authors is graded alleged in this file—asserted without an instrument the desk could verify. About 116,000 people were evacuated in 1986 and roughly 340,000 displaced in total across the following years; the World Health Organization’s assessment names mental health the largest public-health consequence. The record’s most recent entry: on February 14, 2025, a drone strike breached the New Safe Confinement arch over the ruin; the IAEA reports the structure has lost its primary confinement function, with no radiation increase measured, and attribution of the strike remains contested between the parties.
Fukushima Daiichi, in Japan, on March 11, 2011, was a plant that passed the test it was designed for and failed the one it was not. The magnitude 9.0 earthquake the reactors survived; the tsunami that followed stood about 14 to 15 meters against a design basis of about 5.7, drowned the emergency diesels and switchgear, and produced a station blackout in which decay heat had nowhere to go. Units 1, 2, and 3 melted; hydrogen from overheated cladding exploded in the buildings of Units 1, 3, and 4. Japan’s own investigation commission, NAIIC, named it “a profoundly manmade disaster,” the product of regulatory capture—a documented official finding, and the strongest institutional sentence in this file. Releases to atmosphere are estimated by the IAEA in ranges—iodine-131 at 100 to 500 petabecquerels, cesium-137 at 6 to 20—and the accident’s INES 7 rating is a threshold crossed, not an equivalence with Chernobyl established; the two events differ by an order of magnitude on these instruments and are not to be merged by their rating. On health, UNSCEAR’s 2020–2021 assessment finds no documented radiation-attributable effects among residents and future effects unlikely to be discernible. In the other ledger, Japan’s certification system has recognized 2,351 disaster-related deaths in Fukushima Prefecture as of May 1, 2026—3,810 nationwide—deaths from the disaster’s disruption, its evacuations and displacements, certified under a legal instrument that covers the earthquake and tsunami together and cannot be read as an epidemiological estimate of the nuclear evacuation alone; no bridge between the certification count and an excess-mortality estimate was found. The earthquake and tsunami themselves killed 15,901 people with 2,519 missing, none of them by the nuclear accident, and any use of that toll as a nuclear death count, in either direction, is a category error this report does not make. The aftermath runs on: a government cost estimate of ¥21.5 trillion set in December 2016, ¥12.1 trillion spent by fiscal 2021; treated water discharge under the ALPS program begun August 24, 2023, under an IAEA finding of consistency with international standards, some 188,000 cubic meters released in 25 rounds by the autumn of 2026; and about 880 tons of molten fuel debris inside the three reactors, of which gram-scale samples have been retrieved on trial from Unit 2, with full retrieval not planned before 2037 and the decommissioning roadmap’s completion, 2041 to 2051, graded a forecast. Japan took all 54 of its reactors offline by May 2012 and had restarted 15 by September 2026, paying on the order of ¥20 trillion in extra fossil fuel imports across 2011 to 2023 in the interval. Germany’s reversal is the accident’s documented political aftershock abroad: Berlin cited Fukushima explicitly in abandoning its own life-extension program in 2011, and its last three plants closed on April 15, 2023. What the closure cost in emissions is contested—a peer-reviewed synthetic-control study finds the penalty closed by 2023; an advocacy counterfactual attributes some 730 million tons of carbon dioxide and 19,200 to 24,000 deaths to the policy—and no official German net cost for the phase-out alone exists. Both instruments are printed; the bridge between them is unknown.
3. The Operating Record: The Fleet That Exists
The American fleet at the research cutoff, on the Energy Information Administration’s March 2026 figures: 94 reactors at 54 plants in 28 states, average age about 44 years. Its capacity prints three ways in three instruments—98,441 megawatts on the EIA’s survey form, 96,952 megawatts on the World Nuclear Association and IAEA PRIS figures, 96,880 in the EIA’s Monthly Energy Review—and the desk prints all three with their instruments rather than choosing, the reconciliation being a survey-definition difference no source resolves to the megawatt. In 2025 the fleet generated 784,781 gigawatt-hours and ran at a 91.0 percent capacity factor, the highest sustained performance in its history and the industry’s standard exhibit. Its share of national generation fell to 17.7 percent in 2025, from 18.1 percent in 2024 and 19.5 percent earlier in the decade, and the decomposition matters: nuclear output rose slightly while total American generation grew faster. The share decline is arithmetic, not decay. One internal disagreement is carried openly: the EIA’s 2024 capacity factor appears as 90.8 percent in its table and as 92 percent in its own summary publication, two readings from one agency that this report does not reconcile.
The world fleet prints two ways as well. The IAEA’s PRIS database counted, on October 27, 2026, 417 operating reactors of about 380 gigawatts and 78 under construction; the World Nuclear Association counted, on November 6, 2026, 443 operable units of 405,667 megawatts and 81 under construction. The gap of roughly two dozen units is in substance the suspended fleet—mostly Japanese reactors in long-term shutdown that PRIS lists separately at 23 units—but an exact reconciliation between the two instruments was not found, and the desk does not invent one. Of the construction actually underway, China holds 40 units, India 8, Russia 7, Egypt 4, Türkiye 4, Britain 2, and the United States none. The geography of new building is the single most legible fact in the file: the technology’s construction industry lives where states build it.
The American ordering record explains the fleet’s age. Five reactors entered service in the thirty-one years from 1993 to 2024: Comanche Peak 2 in 1993, Watts Bar 1 in 1996, Watts Bar 2 in 2016, and Vogtle 3 and 4 in 2023 and 2024. Vogtle is the modern build record in full, and it is filed here because every cost figure in section 4 stands on it. Georgia Power and its co-owners proposed two AP1000 reactors at about $14 billion, for service in 2016 and 2017. The units entered service on July 31, 2023, and April 29, 2024—about seven years late—at an owners’ cost of about $31 billion, about $35 billion counting the $3.7 billion Toshiba guarantee payment that followed its subsidiary Westinghouse’s bankruptcy. The derived capital cost is roughly $13,900 to $15,700 per kilowatt depending on the denominator chosen. The Georgia Public Service Commission’s stipulation of December 19, 2023, settled the division: of Georgia Power’s $10.19 billion share, ratepayer recovery was capped at $7.562 billion of capital plus $1.02 billion of financing, and shareholders absorbed about $2.63 billion. Westinghouse’s Chapter 11 filing of March 29, 2017—the event that also killed the twin V.C. Summer project in South Carolina, abandoned that July—is graded in this file as consequence and accelerant of the overruns, not their originating cause; the causes the literature assigns are design immaturity at construction start, first-of-a-kind status, an atrophied supply chain and workforce, regulatory change mid-build, project management, and financing costs compounding across a decade and a half. Every one of those causes is a statement about institutions, and none is a statement about physics.
What the existing fleet does instead of being replaced is extend and enlarge itself. The NRC has approved about 171 power uprates since the 1970s adding about 8,030 megawatts—the output of several large reactors, obtained from plants already built. License renewal has carried 86 of the 94 reactors to sixty years, and by November 2026, 25 units at 13 sites held approvals to eighty. Diablo Canyon in California holds renewal on paper to 2044 and 2045 while state law caps its operation at 2029 and 2030, both facts standing. Retirements, which ran to 13 units across 2013 to 2022 for unit-specific economic and mechanical reasons, stopped entirely from 2023 through 2026; the last American closure was Palisades in May 2022. And Palisades is now the lead case of the reversal: the first shuttered American plant to attempt a return to service, financed by a $1.52 billion Department of Energy loan closed in September 2024, its operating license basis restored by the NRC in July 2025, fuel loading begun August 30, 2026, and then paused—over a tilted fuel assembly, recovered under License Amendment 283, issued October 13, 2026, for the assembly designated DD41, in a plant where the NRC has also documented a chilled work environment in two departments. Its target date has slipped from late 2025 through 2026 toward a power-purchase deadline of March 1, 2027, and no firm date stands.
The other two restarts are contracted to data-center buyers, and their full record is filed in section 7; the operating facts belong here. Crane Clean Energy Center—the former Three Mile Island Unit 1, undamaged in 1979 and shut down for economics in 2019—is being restarted by Constellation under a twenty-year Microsoft power purchase agreement signed September 20, 2024, at a restart capital cost of about $1.6 billion, with a $1 billion federal loan closed in November 2025; its target has moved to the second half of 2027, and an environmental finding of no significant impact issued in September 2026 is not, as sometimes reported, a restart approval—one exemption and three license amendments remain pending; and while FERC approved the transfer of 760 megawatts of interconnection rights to the project on June 1, 2026, the PJM grid operator’s deliverability concern, which runs to 2031, stands unresolved in any document this research found. Duane Arnold in Iowa, closed in 2020, holds a twenty-five-year Google agreement signed October 27, 2025, and a federal loan of up to $1.9 billion closed September 8, 2026, against a company target of the first quarter of 2029 and an NRC licensing bundle still in review under which a license decision may not land before January 2028. The pattern across all three: the cheapest new nuclear capacity in America is old nuclear capacity, and its buyers are the data-center companies of section 7.
4. The Economics: Existing Power Is Cheap, New Power Is Not
The price of nuclear electricity depends entirely on whether the plant exists. Lazard’s Levelized Cost of Energy study, version 19.0 of July 2026—the latest edition, superseding the widely quoted version 18.0 of June 2025, whose figures are a different instrument and are not mixed into these ranges—puts the marginal cost of existing nuclear generation at $26 to $36 per megawatt-hour: the cheapest firm power in the study. The same edition puts new nuclear at $175 to $255 per megawatt-hour, unsubsidized, the most expensive entry on its page, at capital costs of $12,300 to $17,570 per kilowatt that Lazard states are derived from the realized Vogtle record. Between those two lines sits the whole economics of the technology in 2026: an asset that, once built, produces for sixty to eighty years at costs competitive with everything, and that, at the moment of building, costs more per unit of capacity than any alternative in the Western market. The comparison set in the same study: gas combined cycle at $51 to $129, utility solar at $40 to $98, onshore wind at $37 to $99—with Lazard’s own caveat, printed here because the paper’s doctrine requires it, that headline levelized costs exclude the transmission and firming costs that fall differently on different technologies. Costs rose across every technology in the 2026 edition; nuclear’s position relative to the field is not a 2026 event but a settled condition.
The second instrument reads lower, and the difference is the instruments, not an error in either. The Energy Information Administration’s Annual Energy Outlook 2026 prices an advanced nuclear plant entering service in 2031, in 2025 dollars, at $87.81 per megawatt-hour with federal tax credits applied and $110.87 before them—built on forward-engineering capital costs near $8,255 per kilowatt, not on realized build costs. Against it in the same model, gas combined cycle prices at $77.46 at the 40 percent capacity factor the model assigns under environmental rules, or $55.51 at an 87 percent capacity factor. The bridge between Lazard’s $175 to $255 and the EIA’s $88 to $111 is stated in the research and printed here: subsidy treatment (Lazard’s headline is unsubsidized; the EIA’s lower figure is after credits), capital basis (realized Vogtle outcomes against a forward engineering estimate), time horizon, and dollar year. A reader who quotes either figure without its instrument is quoting half a fact. Lazard also offers an illustrative figure for restarted plants, a midpoint near $148 per megawatt-hour, whose underlying range this research could not reliably extract from the study’s charts; it is filed as not found as a range, and the desk does not print the midpoint as more than a marker.
The international build record, country by country, is the evidence both instruments are summarizing. France’s Flamanville 3, an EPR begun in 2007, connected to the grid on December 21, 2024, about twelve years late, at €13.2 billion on the builder’s account excluding financing and €23.7 billion including it on the Cour des comptes’ account—two figures, two instruments, both printed. Britain’s Hinkley Point C has moved from £18 billion to £35 billion in 2015 prices, with Unit 1 now expected in 2030. Finland’s Olkiluoto 3, ordered at a fixed price of about €3 billion, cost about €11 billion in total and entered commercial service in May 2023, its overrun borne by the vendor consortium rather than the buyer—the one Western case where the contract, not the public, absorbed the failure. The Emirates’ Barakah plant, four Korean APR-1400s ordered in 2009 on a prime contract of about $20 billion, had all four units operating by September 2024; Unit 1 slipped about four years from its original date, and an audited final cost was not found—the $24 to $27 billion figures in circulation are graded claimed. China’s fleet builds at a mean of 6.2 years for units started in 2013 to 2022, against a global mean of 9.9 years; in the 2024 cohort, Zhangzhou 1 took 61 months from first concrete to operation against Flamanville’s 204. No primary-sourced Chinese overnight cost per kilowatt exists in the record this research could reach, and any precise Chinese dollar figure in circulation should be treated accordingly. The causes the literature assigns to the Western record—designs started before they were finished, first-of-a-kind units, supply chains and workforces allowed to atrophy between orders, regulation changing mid-build, weak project management, and interest compounding over ten-to-seventeen-year construction—are institutional causes, and their institutional character is why the same technology posts a different record in Korea and China. The physics is common. The building is not.
5. Small Modular Reactors and Microreactors: Permits, Not Power
The small modular reactor is the form in which the technology’s future is currently discussed, and its operating record, complete, is two machines. Russia’s Akademik Lomonosov, a floating plant of two 35-megawatt KLT-40S reactors of naval lineage, entered commercial service at Pevek on May 22, 2020. China’s HTR-PM at Shidao Bay, a pebble-bed high-temperature reactor of 210 megawatts gross and 200 net, entered commercial service on December 6, 2023—and generated 272 gigawatt-hours in 2024, a figure the sources characterize incompatibly (a load factor near 21 percent on one compilation’s rating basis, an operation factor near 72 percent on another’s), neither characterization constituting a published cost of power. No small reactor of a new commercial design sends power to any grid in North America. Neither operating unit publishes a levelized cost. The claim that an operating SMR demonstrates low cost is graded, in this file, not found as stated.
The American record begins with a certification that does not cover the product. NuScale’s 50-megawatt design was certified by NRC rule effective February 21, 2023—the first SMR certification in American history, at about $70 million in NRC fees and roughly 41 months of technical review inside a six-year proceeding. It is not the design NuScale now sells. The 77-megawatt US460 holds a Standard Design Approval, issued May 29, 2025, a document that states on its face that it does not constitute a commitment to issue any permit or license; no construction permit exists, and no firm American order stands. The subsidiary test already ran and produced the file’s most instructive numbers. The Carbon Free Power Project, a NuScale plant for the Utah Associated Municipal Power Systems, began with a target price of $55 per megawatt-hour, revised it to $58, then to $89 by January 2023—that final figure already net of roughly $30 per megawatt-hour of Inflation Reduction Act support, an unsubsidized equivalent near $119—while its cost estimate rose from $5.3 billion to $9.3 billion, about $20,139 per kilowatt all-in. Subscribers committed to 116 of the planned 462 megawatts, against an 80 percent subscription requirement, and on November 8, 2023, the project was terminated. The manner of its death is the finding: it was not rejected by the regulator, which never ruled against the design; it was declined by its buyers at its price. It remains the only fully priced American SMR offered to actual buyers on the record.
The two builds that stand furthest forward are Canadian and American, one under construction and both instructive about what a construction start costs. Ontario Power Generation’s Darlington project, four GE Hitachi BWRX-300 units, holds a construction license from the Canadian regulator issued April 4, 2025, for Unit 1—the only Western SMR construction license in force. The four-unit program is budgeted at 20.9 billion Canadian dollars all-in, about 17,400 Canadian dollars per kilowatt across the program; Unit 1 with the shared infrastructure stands at 7.7 billion, about 25,700 per kilowatt for the first unit—against the vendor’s original target of about US$2,250 per kilowatt. The basemat was placed in the spring of 2026, on a date the sources give as April 22 or May 1 and do not reconcile; the grid-connection target has slipped from the end of 2029 to the end of 2030. In the United States, the Tennessee Valley Authority’s BWRX-300 at Clinch River received the first American SMR construction permit on September 28, 2026, after a fourteen-month formal review resting on a 2019 early site permit and years of pre-application work; no construction start has been announced, TVA’s board has not committed to build, and the Department of Energy has awarded the project $400 million under its Generation III+ program. A permit, in the American two-step system, is permission to begin the part of the project that has historically consumed the money.
The advanced designs advance on the same clock. TerraPower’s Natrium reactor at Kemmerer, Wyoming, received its construction permit in March 2026—the Commission authorized it on March 4 and it issued March 9, the first construction permit ever granted for a commercial non-light-water reactor in the United States—and reactor construction began on April 23, 2026, on the company’s report; the IAEA’s database carries no under-construction entry for it that this research found. Its cost estimate of $4 billion derives to about $11,600 per kilowatt; an independent projection by the Southern Alliance for Clean Energy runs to about $29,000. Its target has moved from operation in 2028 to completion in 2030 and commercial service in 2031, and its first core requires 15 to 20 tons of high-assay low-enriched uranium, a fuel whose supply chain is the project’s binding constraint—one announced supplier’s enrichment plant does not yet exist. Its builder, Bechtel, exited the engineering and construction role in September 2026. X-energy’s Xe-100 project for Dow at Seadrift, Texas, filed its construction permit application on March 31, 2025, received its environmental finding in May 2026, and awaits a decision the parties variously expect at the end of 2026 or in the first half of 2027; Dow has not disclosed its cost. Its fuel supply took one concrete step in February 2026, when X-energy’s TRISO-X fuel fabrication plant received its Part 70 license—the fuel factory licensed before the reactor it would feed. Oklo, the most publicly valued of the microreactor companies, had its combined license application denied without prejudice in January 2022—an information failure, with no safety finding against the design—broke ground at Idaho National Laboratory in September 2025 under Department of Energy authority rather than an NRC license, holds a June 2026 preliminary safety document that is not authority to load fuel, and has fuel for one unit only, about five tons of HALEU recovered from the experimental EBR-II; it has never operated a reactor and is pre-revenue. The Department of Defense’s Project Pele transportable reactor, at about $300 million, has moved its assembly and testing to 2027 and operation to 2028. Westinghouse’s eVinci has no reactor application under NRC review; a scaled zero-power mock-up achieved criticality on August 24, 2026, which is a physics validation, not a power plant. Radiant’s Kaleidos has not yet gone critical, having missed its third-quarter 2026 target. China’s Linglong One, the ACP100, under construction since July 2021, reached hot functional tests in October 2026 without a fuel load, its original first-half-2026 target missed. No microreactor anywhere in the world has generated electricity commercially. Vendor targets for mature, nth-of-a-kind small reactors run $3,000 to $6,000 per kilowatt and are graded claimed; independent analysts’ first-of-a-kind levelized estimates run $80 to $150 per megawatt-hour and are graded claimed-analyst; Lazard publishes no SMR-specific cost of energy at all, and attributions of one in circulation are false.
6. Breeders, Fast Reactors, and the Waste
The oldest promise in the file is the reactor that makes or eats fuel, and its record is the inverse of its rhetoric: the machines that run are Russian and old; the machines that were celebrated are closed. Russia’s BN-600 fast reactor has operated since 1980 and holds a license extended to 2040, across a record that includes 27 sodium leaks and 14 fires between 1980 and 1997 with no public release. Its successor, the BN-800, has run on a full core of mixed-oxide fuel since September 2022—one pass, not a cycle—and discharged its first three pilot assemblies carrying minor actinides in April 2026, with the operator’s own caveat that the reduction in those materials “must be confirmed”; the planned throughput of the entire program, about four tons of minor actinides across some sixty years, is the scale against which claims of burning a national stockpile should be read. Against that record: France’s Superphénix, the largest breeder ever built, ran eleven years at a lifetime load factor under 7 percent, generated about 8.2 terawatt-hours, and was closed by political decision in 1997–98; its predecessor Phénix ran 1973 to 2009; France’s follow-on ASTRID was cancelled in 2019. Japan’s Monju accumulated about 250 days at power in some twenty-two years, survived a 1995 sodium leak and its cover-up in the record if not in public standing, and is being decommissioned through 2047. The American line is terminated at every point: Fermi 1’s partial meltdown in 1966, EBR-II’s thirty productive years ending in 1994 with its Integral Fast Reactor program, and the Clinch River Breeder cancelled in October 1983 after $1.5 to $1.7 billion spent. Breeding works. The record shows it working, at state expense, in one country’s program, and shows it abandoned on cost and politics everywhere else it was tried at scale.
Reprocessing—the chemical separation that any waste-burning scheme must start from—has a matching shape. Of about 448,000 tons of spent fuel discharged worldwide, about 28 percent has been reprocessed, against a world capacity near 2,000 tons a year; France’s La Hague plant has run at up to 1,700 tons a year since 1976, and a single pass through reprocessing and mixed-oxide fuel recovers only about a fifth more energy from the original uranium, on the Carnegie Endowment’s accounting. Japan’s Rokkasho plant has been under construction since 1993, postponed twenty-six or twenty-seven times, and has never operated. The United States has had no commercial reprocessing since the 1970s. The technical verdict of the file is threefold and is printed without softening: burning existing waste requires separation, a fast neutron spectrum, and multiple recycle passes; that combination operates nowhere on earth; and a geological repository is still required afterward for what remains, on the OECD Nuclear Energy Agency’s consensus—the claim that recycling eliminates disposal is graded, in this file, false. One further finding runs against the small-reactor claim rather than for it: a 2022 study in the Proceedings of the National Academy of Sciences by Krall and co-authors calculated that some SMR classes would produce two to thirty times more waste volume per unit of energy than the existing fleet, a result the industry contests and this report files as contested rather than settled. The commercial waste-burning ventures stand where their paper stands: Oklo’s announced $1.68 billion Tennessee fuel center is a claim with no license in existence; Moltex lost its sponsoring proponent in 2025; ARC-100 completed a Phase 2 vendor design review in July 2025 and holds no construction license. No commercial reactor anywhere is licensed, built, or operating to burn existing civilian spent fuel. That sentence is the section’s verdict, and section 11 returns to it.
The waste itself, in the American inventory: 95,117 metric tons of heavy metal in 329,237 assemblies at 73 sites, as of December 31, 2024, on the Idaho National Laboratory’s inventory for the Department of Energy; about 56 percent of it in some 4,200 dry casks, the rest in pools; growing at about 2,000 tons a year toward a projected lifetime total near 180,000 tons. Its legal destination remains Yucca Mountain, Nevada—the site in statute, its 70,000-ton cap already exceeded by the inventory on the ground, its licensing proceeding suspended since September 30, 2011, and unfunded since that year. The Department of Energy has pivoted to consent-based siting for federal interim storage, naming five states on August 4, 2026. The two commercial interim-storage licenses—Interim Storage Partners in Texas, licensed 2021, and Holtec in New Mexico, licensed 2023—were vacated by the Fifth Circuit and restored on June 18, 2025, by the Supreme Court in NRC v. Texas, a six-to-three decision on party-status grounds that expressly did not decide whether the NRC has the authority in question; Holtec then cancelled its project in October 2025, the Texas project stands unbuilt, and the tonnage in consolidated interim storage in the United States is zero. The one operating American deep repository, WIPP in New Mexico, has taken defense transuranic waste since 1999 and cannot by law accept commercial spent fuel. Finland’s Onkalo, which received its operating license in October 2026, is the world’s first repository for spent fuel to reach that step; the American program, older than Finland’s by decades, has not.
7. Nuclear Power for Data Centers: The Plain Answer, and What Is Actually Contracted
The question that commissioned this section admits a one-word answer, and the desk gives it the research file’s own sentence: no. No SMR-powered or microreactor-powered data center exists, and none is under construction, anywhere in the world, as of November 2026. Delivered small-reactor capacity to data centers is zero megawatts. The zero-power criticalities achieved in the summer of 2026 by three Department of Energy pilot projects produce no electricity and power nothing. The structure of everything that does exist is uniform and is the section’s second finding: every kilowatt actually contracted to a named American data-center buyer comes from an existing large light-water reactor or from a restart of one, and every small-reactor arrangement in the file is a future instrument—a master agreement, a memorandum of understanding, a letter of intent, a prepayment, or a power purchase agreement written against a plant that is not licensed. The distinction is not a technicality; it is the difference between the power market that exists and the one that has been announced.
The existing contracts, by buyer. Microsoft’s twenty-year agreement of September 2024 underwrites the Crane restart filed in section 3; its price is undisclosed, and the $98 to $115 per megawatt-hour figures in circulation are analyst estimates, graded alleged. Google’s twenty-five-year agreement of October 2025 underwrites Duane Arnold’s restart, target 2029. Google also holds, from October 14, 2024, a master agreement with Kairos Power for up to 500 megawatts of future reactors, first unit in 2030, full deployment by 2035, price undisclosed—the agreement whose test-reactor footnote the File in Brief carried: Kairos’s Hermes 1 is a 35-megawatt-thermal test reactor producing no electricity, permitted in December 2023 and under nuclear construction since May 2025, its completion window since extended to April 2029; Hermes 2 holds construction permits from November 2024, for a plant the permit documents describe at about 20 to 28 megawatts electric and an August 2025 TVA power agreement describes as “up to 50 megawatts”—a discrepancy between instruments the desk prints unresolved—with ground broken in April 2026 toward a 2030 target. Amazon led a $500 million investment round in X-energy in October 2024 and announces an ambition of more than five gigawatts by 2039, graded claimed; its physical project, the Cascade plant with Energy Northwest in Washington, stands in pre-application with no construction permit application filed, its filing date slipped from the end of 2026 to the first half of 2027; a Dominion Energy memorandum concerning North Anna is exploratory and dormant. Meta ran the field’s largest solicitation, a December 2024 request for one to four gigawatts, and signed in June 2025 a twenty-year agreement at Constellation’s operating Clinton plant—attributes from June 2027, a 30-megawatt uprate attached, price undisclosed; its January 9, 2026, announcement of “up to 6.6 gigawatts by 2035” disaggregates, on inspection, into 2,176 megawatts of Vistra capacity already operating, 433 megawatts of uprates not yet approved, an Oklo project of up to 1.2 gigawatts holding prepayments and no construction permit, and TerraPower funding for two Natrium units with rights to six more and no Meta-site permit: a headline composed entirely of existing generation, unapproved increments, and options, filed in section 11 among the claims not found as stated. Oracle’s case is the section’s cleanest: in September 2024 its chairman stated the company held building permits for three nuclear reactors to power a data center; in the twenty-six months since, no site, no vendor, no NRC filing, and no securities disclosure has been found, and the delivered capacity is zero. Oklo’s 12-gigawatt master agreement with Switch of December 2024 is non-binding on its face; its Equinix arrangement is a letter of intent with a $25 million prepayment that appears in filings and is not revenue; its announced pipeline of about 14 gigawatts is mostly non-binding. The International Energy Agency, in April 2026, tallied the world’s “conditional” SMR offtake pipeline at 45 gigawatts—the agency’s own word, conditional, is the accurate one—against delivered capacity of zero. The IEA’s Energy and AI report is the demand side’s standard instrument and is clipped below.
One data center in America physically operates beside its contracted plant, and its file shows what “beside” now means. Talen Energy sold its Cumulus campus next to the Susquehanna station to Amazon Web Services for $650 million in March 2024, then sought to expand a behind-the-meter supply arrangement from 300 to 480 megawatts; the Federal Energy Regulatory Commission rejected the expansion on November 1, 2024, by two votes to one, and denied rehearing in April 2025—a documented regulatory decision that co-location does not exempt a load of that size from the grid’s rules. The parties restructured on June 11, 2025, into front-of-the-meter grid delivery of 1,920 megawatts through 2042, an arrangement Talen values near $18 billion, transitioned in April 2026; and in December 2025 FERC ordered the PJM grid operator to define co-located-load services, the question Susquehanna raised having become general. The demand underneath all of it is measured by two instruments that must not be spliced: Lawrence Berkeley National Laboratory’s 2024 report puts American data centers at 176 terawatt-hours in 2023, 4.4 percent of national consumption, and projects 2028 in scenarios spanning 325 to 580 terawatt-hours—a scenario gap wider than the entire 2023 consumption it extrapolates from; the IEA’s global series runs 415 terawatt-hours in 2024 to about 950 in 2030 in its base case. And announced load is not built load: PJM’s independent monitor counted about 4,017 megawatts of forecast data-center load cancelled or delayed as of September 2026. The buyers in this section are real, their contracts with existing reactors are real, and the small-reactor supply they have announced exists, at cutoff, as paper—priced, in the one case where a price was tested against buyers, at a level the buyers declined.
8. Safety in Comparison: Deaths per Unit of Electricity, and the Waste in Tons
The comparative safety record is kept, by the standard instrument, in deaths per terawatt-hour, and on that instrument nuclear power stands at the safe end of the table by two to three orders of magnitude from the fossil fuels. Our World in Data’s series gives brown coal 32.72 deaths per terawatt-hour, coal 24.62, oil 18.43, biomass 4.63, gas 2.82, hydro 1.30, wind 0.04, nuclear 0.03, and solar 0.02. The desk prints the construction of the nuclear figure because the figure is built, not counted: it divides 433 Chernobyl deaths plus 2,314 Fukushima deaths—the Fukushima component being the certified evacuation-and-disaster ledger of section 2, not radiation deaths—by about 96,876 terawatt-hours of historical nuclear generation. Substitute the Chernobyl Forum’s modeled 4,000 for the confirmed count under 100 and the rate becomes about 0.07; substitute the alleged 93,000 and it becomes about 0.98. The ranking survives every assumption the desk tested; the second decimal does not. The peer instruments spread around the same ground—Markandya and Wilkinson at 0.074, Sovacool at 0.0097, a range of about eightfold across the literature—and the fossil figures are European-based and likely understate the global toll. Our World in Data’s safest-sources comparison is clipped below as the instrument’s own presentation. Nothing in this section assigns a cause of death a lesser standing; it assigns each number its instrument, which is the paper’s whole method.
Two parts of the industry’s own safety record belong inside the comparison, not outside it. The front end of the fuel cycle carries the industry’s documented historical wound: uranium mining on the Colorado Plateau in the 1950s exposed Navajo miners to radon under conditions the Public Health Service studied while the exposure continued; between 3,000 and 5,000 Navajo miners worked the mines from 1944 to 1986, more than 500 abandoned mines remain, and the Church Rock spill of July 16, 1979—the largest release of radioactive material by volume in American history, a milling accident, not a reactor event—is part of the permanent record. The Radiation Exposure Compensation Act of 1990 lapsed in June 2024 and was reauthorized and expanded in 2025. The modern record is a different record—imports, in-situ recovery, regulated doses—and the two are never to be merged into one characterization in either direction; no separate instrument counting current global uranium-mining deaths per terawatt-hour was found. The operating record, on the regulator’s instrument: the maximally exposed member of the public near an American plant receives typically a fraction of a millirem to a few millirem a year from routine effluents, against an average American total dose of about 620 millirem a year from all sources; American light-water reactor workers in 2024 accumulated a collective dose of 4,903 person-rem with an average measurable individual dose of 0.16 rem, about 3 percent of the five-rem occupational limit. A 2026 NRC proposal would consolidate the design objective for effluents to 25 millirem a year; it does not change the 100-millirem public or five-rem occupational limits, and it is a proposal.
The waste comparison is kept in tons, with its caveat printed in both directions as the research directs. American spent fuel grows at about 2,000 tons a year—about 2.7 grams per megawatt-hour. American coal ash runs at about 70 million tons a year in recent estimates, about 110 million tons in the EPA’s 2012 vintage: one year of coal ash is about 35,000 times one year of spent-fuel growth by mass, and the entire seventy-year American spent-fuel stock approximates a single day of the coal-ash stream. The streams are not hazard-equivalent per ton—spent fuel’s hazard is concentrated, long-lived, and heat-generating; coal ash’s is dispersed, voluminous, and chemically toxic in perpetuity—and any use of the mass ratio without that sentence, by either side, misstates the record. For the newer technologies the citable figures are thinner: the International Renewable Energy Agency’s projection of 78 million cumulative tons of solar panel waste by 2050, from its 2016 regular-loss scenario, is the only stable figure this research found; the agency’s 2026 figures do not reconcile with it and no bridge was found, and no battery-waste projection at the grade this file requires was found at all. Section 6’s verdict stands beside these numbers: the waste problem in America is not a volume problem or a technology problem; it is a custody problem—95,117 tons, inventoried to the assembly, awaiting a destination the political system has licensed abroad and not at home.
A live scientific dispute closes the section because it bears directly on the small reactors of section 5. High-assay low-enriched uranium—HALEU, the fuel the new designs require, enriched above 5 and below 20 percent—was assessed in Science on June 7, 2024, by Kemp and co-authors, who concluded that material above roughly 12 percent enrichment could be used directly in a practical weapon, and that several hundred kilograms to about a thousand kilograms at 19.75 percent enrichment approximates a Hiroshima-class yield. The American Nuclear Society answered in named correspondence that it is unconvinced by the analysis, citing sixty years of research-reactor HALEU use without a known diversion and noting that any redefinition must proceed through the IAEA; the National Nuclear Security Administration’s administrator commissioned a National Academies assessment in January 2025 whose outcome is not in the record. The desk grades the underlying question open, under formal assessment: the fuel that the small-reactor program runs on sits in a category its own government is still having measured. That is the documented state, and it is printed without a recommendation in either direction, as this paper files it.
9. Regulation and Siting in 2026: What Changed, and What Did Not
American reactor licensing runs on statutes whose ages explain the timeline problem of section 5. Part 50, the original two-step system—construction permit, then operating license—dates to the industry’s founding. Part 52, from 1989, offers one-step combined licenses, early site permits, and design certifications. Part 53, the technology-inclusive framework mandated by the Nuclear Energy Innovation and Modernization Act of 2019—not, as often misattributed, a product of the 2024 ADVANCE Act—became final on March 30, 2026, and effective April 29; both construction permits issued in 2026 nonetheless proceeded under Part 50, and no Part 53 application has been docketed that this research found. The timelines the frameworks have actually produced: Vogtle’s combined license took about three years and eleven months from application to issuance, and about fifteen years from application to power; NuScale’s design certification took about 41 months of technical review inside six years from application to effective rule; Kemmerer’s construction permit ran a technical review under eighteen months against a twenty-seven-month schedule; Clinch River’s ran fourteen months, resting on its 2019 site permit. No case in the record runs from application to operating commercial power in under a decade, and no NRC generic schedule promise exists. The acceleration now visible in individual dockets is real and is filed as real; it is measured in months at the front of a decade-long process.
The ADVANCE Act—Division B of Public Law 118-67, signed July 9, 2024—changed four things in evidence. It cut the hourly rate charged to advanced-reactor applicants, effective October 1, 2025; the fiscal 2026 fee rule sets $154 an hour for advanced-reactor work against $337 standard, a reduction of more than half. Its Section 207 sets an expedited combined-license target of a final decision in twenty-five months, qualified in the text by the words “to the maximum extent practicable,” the qualifier being part of the law and part of this report. Its Section 501 mission update produced, on January 24, 2025, a new NRC mission statement that begins “The NRC protects public health and safety” and adds the deployment-efficiency clause beneath it—addition, not substitution; the claim that safety was removed from the mission is graded in section 11 as not found as stated, and the mirror claim that nothing changed is graded the same, the two being the section’s paired slogans. And its incentive prizes under Section 202 have never been awarded, because the licensing events that would trigger them have not occurred. Implementation stood at 33 of 36 deliverables complete in August 2026. Credit for the new pace is itself contested on the record—the NRC’s leadership credits the executive orders, the industry association credits the Act—and attribution is graded claimed where outcomes are verified.
The executive orders of May 2025 supply the targets the statute does not. Order 14299 directs an Army reactor operating at a domestic base by September 30, 2028; its Janus program announced five vendor-site matches on August 26, 2026, and nothing operates. Order 14300 sets the policy goal of expanding American nuclear capacity from about 100 to 400 gigawatts by 2050 and directs decision deadlines of eighteen and twelve months that have not been converted into binding regulation that this research found; its rulemaking program stood, by August 2026, at roughly twenty-seven to thirty-three proceedings, seven final and fifteen proposed, with an agency reorganization effective June 15, 2026. Order 14301’s Reactor Pilot Program—eleven projects, no federal funding—set a goal of at least three test reactors critical by July 4, 2026, and met it on its own terms: Antares on June 4, Valar on June 18, Deployable Energy on June 30, Aalo on July 4, and Oklo’s Groves reactor on August 5, all zero-power criticalities under Department of Energy authorization, none of them an NRC license and none of them a power plant. Order 14302’s report on spent fuel, due in about 240 days around January 18, 2026, has had no public release found. Inside the agency, the period’s personnel record is documented and unusual: Chairman Hanson was removed on June 14, 2025, the first dismissal of a sitting commissioner in the agency’s history; Commissioner Caputo resigned that July; the Commission sat at a quorum of three through much of the second half of 2025 before Chairman Nieh’s designation on January 8, 2026; and the Government Accountability Office counts a net loss of about 500 staff from July 2024 to June 2026—some 740 separations against 270 hires—with the agency’s direct-hire authority unused. Commissioner Crowell’s testimony gives 510 lost and 59 added across a different sixteen months; the two counts cover different windows and are printed together, unreconciled, as the record leaves them.
The states have moved faster than the federal build record. Eight states retain statewide moratoria on new nuclear construction on the Department of Energy and industry count—California, Hawaii, Maine, Massachusetts, Minnesota, Oregon, Rhode Island, and Vermont—with Connecticut and New York maintaining limited regimes; the National Conference of State Legislatures’ restrictions table is clipped below. The repeal record is verified state by state: Wisconsin 2016, Kentucky 2017, Montana’s referendum bar removed 2021, West Virginia 2022, Illinois partial in 2023 and full on January 8, 2026—its repeal paired with a gubernatorial order soliciting at least two gigawatts—and New Jersey in April 2026, followed in July by a procurement law for at least 1,100 megawatts. The money followed in specific instruments: Indiana’s 20 percent SMR manufacturing credit, a Texas fund capped at $132.5 million on the legislatures’ conference count (the industry’s $350 million session-package figure stands unreconciled beside it in the ledger below), and Tennessee’s more than $50 million toward Clinch River with $10 million more for workforce. The industry counts 45 states enacting more than sixty nuclear measures in 2025; the legislatures’ own conference counts more than 250 nuclear bills introduced that year.
10. The Measurements: What the Public Says, by Instrument
The public is measured by the two instruments that have run it longest, and they are printed without blending, as everywhere in this file. Gallup’s favor-oppose series reached 61 percent in favor and 35 percent opposed in March 2025, the second-highest reading in the series’ history behind 62 percent in 2010 and far above its 2016 low of 44 in favor—a low Gallup attributed chiefly to cheap oil and gas, its March 2011 reading having been fielded, a detail the record insists on, before the Fukushima accident it is sometimes said to reflect. Gallup did not ask the series question in 2026; it asked two others, and both are printed, as the paper’s doctrine requires of any instrument quoted in part: 46 percent said more emphasis should be placed on nuclear energy, a record for that item, and 45 percent favored while 53 percent opposed a nuclear plant in their own area. Pew, on different wording, measured support for expanding nuclear power at 59 percent in the spring of 2025—69 percent of Republicans, 52 percent of Democrats—up from 43 percent in its 2020 reading. The industry-commissioned Bisconti reading of 72 percent is filed with its provenance and not in either series. The opinion record, like the capacity record, prints three ways; the trend in each instrument, read on its own terms, is upward.
11. The Claims That Could Not Be Verified
The desk closes the evidence with the claims now in circulation that the record, checked against primary sources, does not support as stated. Each is graded not found as stated—the file’s term for a claim whose documented version differs from its circulated one—and each is printed with what the record shows instead. Seventy-four not-found or unverified entries stand in the full research file; the load-bearing ones follow.
“Small reactor power at $30 to $60 per megawatt-hour.” Not found as stated. The figures trace to vendor targets for mature, nth-of-a-kind production and to aggregator pages, not to any priced plant. The record’s actual prices are filed in section 5: the Carbon Free Power Project’s $89 subsidized target at termination, Darlington’s budgeted 17,400 Canadian dollars per kilowatt, and independent first-of-a-kind estimates of $80 to $150 per megawatt-hour. No operating small reactor publishes a cost of power at all.
“Small reactors are deploying by 2028; commercial fleets by 2030.” Not found as stated, except of Darlington’s planning goal in the narrowest reading. Every dated claim in the sweep has moved: the Carbon Free Power Project’s 2029 operation died with the project in 2023; Oklo’s Aurora 2027 operation has no license to operate under; Natrium’s 2028 became 2030–31; Hermes 1’s December 2026 became a construction window to April 2029. The regulator’s dates in 2026 held or improved; the developers’ operation dates moved right, in every case in the file, without exception.
“We can burn America’s nuclear waste now.” Not found as stated, on the record of section 6: no commercial license for reprocessing exists in the United States, the experimental EBR-II material being recycled is government research fuel—about ten tons of HALEU promised by December 2028—and no reactor anywhere is licensed or built to consume the civilian stockpile. The claim confuses a demonstrated chemistry with an operating industry.
“The Navy’s record proves small reactors are safe and cheap for civilians.” The naval record itself is documented in section 1, in the Navy’s own terms. The equivalence conclusion appears in advocacy, not in any primary source: naval fuel is highly enriched material whose current specification is not public, naval crews are military and selected, naval reactors are licensed by no civilian authority, carry no rate base, and answer to a program that is its own regulator. What transfers from the naval record is reactor physics and operating discipline. What does not transfer, on any document found, is a cost or a licensing conclusion.
“America’s nuclear waste would fit on a football field.” Verified only in its original scoped form, not found in its common shortened forms. The source is the Department of Energy’s 5 Fast Facts About Spent Nuclear Fuel, which states that about 90,000 metric tons of commercial spent fuel, stacked as bare fuel, could fit on a single football field at a depth of less than ten yards. The retellings drop the commercial-only scope, substitute ten feet for ten yards, and omit the same page’s statements that the real inventory sits in pools and casks at more than 70 sites in 35 states, that the United States does not currently recycle spent fuel, and that it has no operating repository. The scoped original is true and is clipped below; the shortened forms are not the source’s claim.
“Thorium reactors are around the corner.” Not found as stated. The world’s only operating thorium-loaded molten-salt reactor is China’s TMSR-LF1: two megawatts thermal, a research machine producing no electricity, which achieved first criticality in October 2023 and the first confirmed thorium-to-uranium-233 conversion on November 1, 2025. Its own institute’s roadmap puts a 100-megawatt-class demonstration around 2035 and commercialization around 2040. No thorium design holds an American license, permit, or certification; the salt-cooled designs in the American pipeline burn uranium. On the primary program’s own schedule, the corner is fifteen years long and located in China.
“Fukushima killed 18,000 people”; and “Fukushima killed nobody.” Both not found as stated, answered in full at section 2: the 18,000-order figure is the earthquake and tsunami toll; the zero figure erases the certified disaster-related ledger and one recognized occupational cancer death. The narrow sentence is the defensible one, and it is the one this report prints.
“The ADVANCE Act and the 2025 orders removed safety from the NRC’s mission”; and “nothing changed.” Both not found as stated, answered at section 9: the mission statement begins with the safety clause the Act required it to keep, and the same period produced a fee reduction in force, a new licensing framework in force, an agency reorganization, a commissioner’s first-ever removal, and about 500 net staff lost. What is documented, and belongs in any honest account in either direction, is that Order 14300 directs reconsideration of the linear no-threshold model and the ALARA principle; proposed rules under that direction were still moving at cutoff. The documented state is directed and in rulemaking—not adopted, and not abandoned.
The shorter sweeps, each not found as stated. “400 gigawatts by 2050” is a policy goal in an executive order, not an order book; no American utility order for a new large reactor was found. Any bare count of states that “ban nuclear power”—nine, twelve, fourteen—is unverifiable without its classification rule; the documented count is eight statewide moratoria plus two limited regimes. “Criticality” does not mean a reactor is online or generating; the summer 2026 events were zero-power, and one pilot reported ten kilowatts of thermal output days later. No regulator has certified any design incapable of a severe accident; regulators issue design-specific findings of reasonable assurance, and the passive-safety claims of vendors describe design-basis behavior over specified windows, while the mirror claim—that passive designs change nothing—is equally unsupported by the Darlington and Clinch River safety cases, which credited them. The United States has no operating repository for commercial spent fuel, and Finland now has a licensed one; both halves of that sentence are the record. And no instrument found—French, American, or international—publishes a settled cost per kilogram that would let the claim “France proves recycling is cheap” be verified; the claim is graded on its missing economics, not on La Hague, which operates.
12. The Contradictions Ledger: Where the Record Does Not Agree with Itself
The desk closes the file the way the record leaves it: with its disagreements printed, not resolved. First, the world fleet: the IAEA’s database counts 417 operating reactors and 78 under construction; the World Nuclear Association counts 443 operable and 81 under construction, ten days later; the difference is in substance the suspended fleet, and an exact reconciliation is not found. Second, the American fleet’s capacity prints three ways—98,441, 96,952, and 96,880 megawatts—across the energy agency’s own survey, the Association and IAEA figures, and the agency’s monthly review, for the same 94 machines. Third, the agency against itself: the EIA’s 2024 capacity factor is 90.8 percent in its table and 92 percent in its summary publication. Fourth, the price of new nuclear: Lazard’s realized-cost instrument says $175 to $255 per megawatt-hour and the EIA’s forward model says $88 to $111, and the bridge—subsidy treatment, capital basis, horizon, dollar year—explains the construction of the gap without closing it; the two are answers to two different questions, and quotation of either as the cost of nuclear power is the error this ledger exists to prevent. Fifth, the Navy against itself: more than 7,600 reactor-years in the program’s 2025 book, more than 7,500 in the figure its other publications carry—different instruments, different years, both self-reported, no independent ledger. Sixth, Hermes 2’s size: about 20 to 28 megawatts electric in its permits, “up to 50” in its power agreement. Seventh, Darlington’s basemat: April 22, 2026, or May 1, 2026, by source. Eighth, HTR-PM’s first full year: 272 gigawatt-hours characterized at a 21 percent load factor on one rating basis and a 72 percent operation factor on another, the definitions unbridged. Ninth, Texas’s 2025 nuclear funding: a $132.5 million fund maximum on the legislatures’ conference count against a $350 million session package on the industry count. Tenth, the NRC’s staff losses: about 500 net on the GAO’s July 2024 to June 2026 window, 510 lost and 59 added on a commissioner’s testimony over sixteen months—different windows, both documented, neither bridged. Eleventh, the boundary of this report’s plainest answer: the research file’s data-center section answers that no small-reactor-powered data center exists or is under construction anywhere, and its claims section records Hermes 2 as one narrow construction case with output contracted toward a data-center buyer. The facts agree; the classification boundary is phrased differently in the two passes, and the report carries the plain answer with the footnote attached rather than either silently. Twelfth and last, the report’s own method, stated once: where this file’s instruments conflict, the desk has printed both readings and named the bridge unknown rather than building one, because a bridge the record does not contain is a claim, and claims in this paper are labeled. The final paragraph of the report is the first paragraph’s restatement, as the form requires: a nuclear reactor is a steam engine with atoms instead of coal; the fleet built on that principle runs at 91 percent of capacity and a falling share, cheap to operate and ruinous, in the Western record, to build; its three accidents killed, by category, thirty men acutely in a design that ran the wrong way, none acutely in the two that lost their cooling, and thousands in the evacuation ledger that belongs to disaster rather than radiation; its waste is inventoried to the assembly and unburied in the country that invented the machine; its small successors hold permits, budgets, and announcements in place of operating hours, power no data center anywhere on earth, and stand or fall, on the one occasion buyers were asked, on price; and its regulator enters 2027 cheaper to apply to, smaller than it was, and still, on the record of its own dockets, the front end of a decade-long road from application to power.