Understanding Nuclear Power

§The central insight

Nuclear power is the peaceful face of the same physics as a weapon: the density that makes its fuel almost free is the density that makes its failure modes severe and its fuel cycle dual-use (the same facilities that serve civil power also serve weapons) — and almost everything puzzling about the industry follows from that one fact.

Start with the sentence most people can produce. Big reactors split atoms, boil water, spin turbines; clean but dangerous, and expensive. Every clause is true. Together they explain nothing. They do not explain why a power reactor cannot detonate, why an industry with near-zero fuel cost produces some of the costliest electricity ever built, why the nations most alarmed by proliferation (the spread of nuclear weapons to states that do not have them) depend on the identical fuel cycle for their electricity, why the West stopped building for thirty years, or why the buyers now reopening shuttered plants are software companies.

This primer builds the working model instead. It walks the enterprise in the order a reader has to learn it — the physics, the fuel cycle (the industrial chain running from ore to fuel to whatever comes back out), the machine, the waste; then eighty years of history told as cause and effect; then the systemic role that history produced; then the frontier, assessed for what is genuinely underway against what is perennially promised. The organising fact runs through all of it. A nucleus of uranium-235 splitting releases about 200 million electron volts, roughly fifty million times the energy of burning one carbon atom, and that single ratio propagates outward into everything: the fuel is cheap, the machine is capital-intensive, the accidents are rare and vivid, the regulator is powerful, the enrichment plant (the machinery that concentrates uranium-235) is a weapons facility with a different setting, and the public is afraid.^[4]

At the end of 2024 the world operated 417 power reactors with 377 gigawatts of net capacity (the power the fleet can deliver at any instant, as distinct from the energy it produces over a year) across 31 countries, and had accumulated 20,163 reactor-years of operating experience (one reactor running for one year, summed across every machine ever built) from 653 reactors built since 1954.^[3] In 2025 those machines produced 2,812 terawatt-hours, an all-time record, and 8.9% of world electricity — the lowest share since the early 1980s.^[8] Both facts are true at once, and holding them together is the beginning of understanding nuclear power. The fleet is running harder than it ever has. The world is simply growing its electricity faster.

Nuclear supplies roughly a fifth of the world's low-carbon electricity, not the quarter often cited. In 2025 low-carbon sources — renewables plus nuclear — reached 42.6% of global generation, of which nuclear was 8.9 points. That is 20.9%, and the number has been falling for a decade because wind and solar have been growing and nuclear has not.^[8] The distinction is doing analytical work: it separates a technology that anchors decarbonisation from one that contributes to it while ceding the margin to something cheaper.

Kazakhstan mined 39% of the world's uranium in 2024.^[7] Four organisations — one of them a Russian state corporation holding 44% — perform essentially all commercial uranium enrichment.^[6] Five plants in five countries do the conversion step that feeds them: turning mined uranium oxide into the gas that enrichment machinery requires.^[45] France runs 57 reactors that supplied 67.3% of its electricity in 2024, the highest national share on earth.^[3] Over the past decade, 94% of reactors that began construction anywhere in the world were of Chinese or Russian design.^[2] A technology that looks, from a distance, like a global industry is on inspection a set of narrow channels with very few operators in them.

Four confusions this primer keeps straight

A reactor is not a bomb, and the reason is arithmetic rather than engineering caution. Fusion is not fission, and every one of the 417 machines counted above is a fission machine. A reactor design is not a built plant; nuclear's cost and schedule risk lives almost entirely in construction, not in physics. And radiation dose is not the same thing as public perception of radiation dose, which is why the honest accident record reads so differently from the remembered one. Each of these is drawn out where it belongs rather than announced here.

A note on what this document is. It is an explanatory primer built for a permanent corpus. It carries no buy, sell or hold view, no valuations, no price targets, and no grading of companies. Firms and states appear as actors in a system: what they do, what they control, and what breaks downstream if they stop.

1The physics: a chain reaction held exactly critical

A uranium-235 nucleus is a marginal object. It holds together, but only just; it is large enough that the electrostatic repulsion among its 92 protons nearly balances the strong force binding it. Add a single slow neutron and the nucleus becomes uranium-236 in an excited state, deforms, and splits — typically into two unequal fragments, two or three fresh neutrons, and about 200 million electron volts of energy, some 3.2 × 10⁻¹¹ joules. Per unit mass that works out to roughly 82 terajoules for a kilogram of U-235 fully fissioned, about fifty million times the energy released when a carbon atom burns.^[4]

The neutrons are what make it a technology rather than a curiosity. If, on average, exactly one neutron from each fission goes on to cause another fission, the reaction sustains itself at constant power. Reactor physics gives that condition a name: the effective multiplication factor, k, equals one, and the system is critical. Below one it dies away. Above one it grows. A power reactor spends its operating life held at k = 1 with a precision that would be alarming if the margin were not so forgiving.

Delayed neutrons: the reason a reactor can be steered

Most fission neutrons appear within about 10⁻¹⁴ seconds. If those were the only neutrons, the population would double in microseconds and no mechanical control rod, no operator, no computer could respond in time. But a small fraction of neutrons emerge later — seconds to minutes later — from the radioactive decay of certain fission fragments. For uranium-235 that delayed fraction is 0.66%.^[4] Sixty-six parts in ten thousand.

That sliver changes everything. A reactor is operated deliberately sub-critical on prompt neutrons alone and made critical only with the delayed contribution included. The effective response time of the system is then set not by microseconds but by the half-lives of the delayed-neutron precursors, and power changes take tens of seconds. The World Nuclear Association puts it plainly: delayed neutrons are "the crucial factor enabling a chain reacting system to be controllable", and without them any departure from balance would produce "a virtually instantaneous and uncontrollable rise or fall in the neutron population."^[4] Control rods, boron in the coolant, the temperature feedbacks in the fuel — all of it works because delayed neutrons buy the machine time.

A nuclear weapon works by driving a mass of highly enriched material prompt supercritical: k above one on prompt neutrons alone, so the chain multiplies in microseconds before the assembly blows itself apart. That requires uranium enriched to roughly 90% U-235 or better, and it requires assembling the geometry faster than the material can disperse.^[6] Reactor fuel is 3–5% U-235, held in ceramic pellets inside metal tubes inside a water-filled pressure vessel.^[52] At that enrichment the neutron economy cannot go prompt critical at all; there is too much U-238 absorbing and scattering. Push a light-water reactor hard enough and the fuel melts, the coolant boils away, the containment is tested — Chernobyl demonstrated an explosion of steam and hydrogen violent enough to lift a thousand-tonne cover plate.^[16] None of that is a nuclear explosion. The physics of the fuel forbids it.

Moderation, and why the coolant choice determines the whole plant

Neutrons come out of fission fast, and fast neutrons are inefficient at splitting U-235. They must be slowed by roughly a factor of ten thousand in energy, and the way to slow a neutron is to bounce it off something of comparable mass. Hydrogen is ideal, which is why ordinary water is the world's dominant moderator. It is also, conveniently, an excellent coolant. That coincidence is the reason 315 of the world's reactors are pressurised water reactors and 60 are boiling water reactors.^[52]

Ordinary water has one drawback: its hydrogen absorbs neutrons as well as slowing them, which costs enough of the neutron economy that natural uranium at 0.7% U-235 will not sustain a chain reaction in it. Hence enrichment. Heavy water, in which the hydrogen carries an extra neutron already, barely absorbs at all; a heavy-water reactor will run on natural uranium, which is why Canada's CANDU design and India's PHWRsPressurised heavy-water reactors — the family that uses heavy water as both moderator and coolant, of which Canada's CANDU is the original design. need no enrichment plant.^[52] Graphite is a third option, used at Calder Hall, at Chernobyl and at Chicago in 1942. Each choice cascades into pressure, temperature, refuelling method, fuel form and — as Chernobyl showed — into the safety physics of the whole machine.

Decay heat: why "off" is not "safe"

Drop the control rods and fission stops within a second. The reactor does not stop making heat. About 6% of a reactor's thermal output at full power comes not from fission but from the radioactive decay of accumulated fission productsThe fragments left behind when a nucleus splits — a zoo of lighter elements, many of them intensely radioactive and most of them short-lived. They are what makes spent fuel dangerous in its first decades. and transuranic elementsElements heavier than uranium — plutonium, americium, curium — built up inside the fuel when a uranium nucleus absorbs a neutron without splitting. A small part of the mass, and almost all of the very long tail: they are the reason spent fuel gets discussed in tens of thousands of years rather than hundreds., and that decay does not care about control rods.^[4] The peer-reviewed review published for the NEA's Working Party on Nuclear Criticality Safety puts the figure at about 7% of total thermal power immediately after shutdown, and notes what changes with time: fission products dominate for the first decade, and after roughly a century americium-241 alone generates almost half the remaining heat.^[91] One hour after shutdown a large reactor is still producing roughly 1.5% of full power as heat — for a 3,000-megawatt-thermal core, some 45 megawatts. A year later, about 10 kilowatts per tonne of used fuel; after ten years, about 1 kilowatt per tonne.^[4]

Forty-five megawatts is a great deal of heat in a sealed steel vessel. If it cannot be removed the water boils off, the fuel claddingThe sealed metal tube — a zirconium alloy — that holds the fuel pellets and keeps what they contain out of the coolant. One of the successive barriers a reactor's safety case is built from, and the one that reacts with steam to make hydrogen when it gets hot enough. oxidises in steam and generates hydrogen, and eventually the fuel melts. This single mechanism — decay heat with the cooling gone — is the accident. It is what happened at Three Mile Island when a stuck valve drained the coolant and the operators, misreading their instruments, throttled back the emergency injection.^[18] It is what happened at Fukushima Daiichi when a tsunami took out the power needed to run the pumps.^[17] Every safety system in a modern plant is, at bottom, a way of guaranteeing that decay heat has somewhere to go.

2The front end: from ore to fuel, and the chokepoint in the middle

Uranium is common, mining it is ordinary industrial work, and fabricating fuel is a specialist but contestable business. The step between them — enrichment — is the industry's strategic bottleneck and its proliferation hinge, and the same machines do both jobs.

Uranium is about as abundant in the earth's crust as tin, and more abundant than silver. Identified resources recoverable below $130 per kilogram stood at 5.93 million tonnes as of January 2023; at current consumption of roughly 67,000 tonnes a year that is about ninety years of supply, and known resources grew by at least a quarter in the preceding decade simply because people looked.^[40] Scarcity of uranium has never been the constraint on nuclear power and is not one now.

Mine output is concentrated all the same, for reasons of geology and cost rather than rarity. World production was 60,213 tonnes of uranium in 2024, of which Kazakhstan alone supplied 23,270 tonnes — 39% — mostly by in-situ leaching, in which a solution is pumped through a sandstone orebody and the uranium recovered from what comes back up. Canada supplied 24% from the extraordinarily high-grade Athabasca deposits; Namibia 12%, Australia 8%.^[7] On a corporate basis, Kazatomprom produced 21% of world output and Cameco 17%.^[7]

Mine product is uranium oxide concentrate, U₃O₈, the yellow-brown powder still called yellowcake. To enrich it, it must first be turned into a gas, and only one uranium compound is conveniently gaseous at modest temperature: uranium hexafluoride, UF₆. Conversion is the chemistry that gets there, and it is done at five plants in five countries — Rosatom at Seversk, CNNC at Lanzhou and Hengyang, Orano at Malvési and Pierrelatte, Cameco at Port Hope and ConverDyn at Metropolis, Illinois. Licensed capacity totals about 62,000 tonnes of uranium a year against actual UF₆ production nearer 42,000.^[45] Conversion is unglamorous, low-margin, and almost invisible in public discussion of nuclear power. It is also a genuine chokepoint: three countries hold roughly half the licensed capacity, and a converter that closes does not reopen quickly.

Enrichment, separative work, and the centrifuge

Natural uranium is 99.3% U-238 and 0.7% U-235. Light-water reactors need 3–5%. Since the two isotopes are chemically identical, the only handle is the 1% mass difference, and the only practical way to exploit it at scale is to spin the gas. A modern gas centrifuge is a rotor three to five metres tall turning at 50,000 to 70,000 revolutions per minute in a vacuum casing; the heavier U-238 molecules drift to the wall, the lighter U-235 concentrates toward the axis, and a thermal counter-current lets the enriched fraction be drawn off.^[6] One machine achieves very little. Thousands connected in cascades, the product of one stage feeding the next, achieve a great deal.

That trade-off is not a technical footnote. When uranium is cheap and SWU expensive, enrichers run "dirty" tails and buy more feed; when the reverse holds they strip the tails harder and effectively create uranium out of enrichment capacity. It is one of the reasons the uranium and SWU markets move against each other, and one of the reasons an enrichment shortage is not solved by mining more.

World enrichment capacity was about 61.5 million SWU a year on a 2022 basis, and it sits in four hands. Rosatom holds 27.1 million SWU — 44% of the world's capacity. Urenco, the Anglo-Dutch-German consortium with plants in three European countries and New Mexico, holds 17.9 million, or 29%. China's CNNC holds 8.9 million, 15%. France's Orano holds 7.5 million, 12%. Everyone else on earth accounts for one-sixth of one percent.^[6]

Dual use is not a metaphor here. The term names a specific and uncomfortable property: not two technologies that happen to resemble one another, but one technology whose civil and military products are separated by a setting rather than by a design. A cascade that enriches to 5% is the same cascade that enriches to 90%; the difference is how many stages the material passes through and how the pipework is arranged. Most of the separative work is spent getting from 0.7% to 5%, so a state that has mastered commercial enrichment has already done the hard part of the weapons problem. This is why the IAEA's safeguardsThe IAEA's treaty-based system of material accounting, inspection and surveillance, whose purpose is to verify that declared nuclear material stays in civil use and is not quietly diverted. It verifies; it does not police, and it can only look where the state has agreed to let it look. regime concentrates its attention on enrichment and reprocessing plants, and why the agency reported that Iran, "the only NPTThe 1968 Treaty on the Non-Proliferation of Nuclear Weapons. Its bargain, set out in full in §15: states without weapons agree not to acquire them and accept inspection of their nuclear material, states with them pursue disarmament, and every party keeps the right to civil nuclear power. non-nuclear-weapon State that is producing and stockpiling high enriched uranium", held an estimated 3.9 significant quantities of uranium enriched to 60% at the end of 2024 — a significant quantity being the amount for which "the possibility of manufacturing a nuclear explosive device cannot be excluded."^[31] Sixty per cent is not weapons-grade. It is also not far from it, and everyone involved knows exactly why.

Enriched UF₆ is then converted to uranium dioxide powder, pressed and sintered into ceramic pellets roughly the size of a pencil eraser, stacked into zirconium-alloy tubes and bundled into assemblies. Four firms dominate light-water fuel: Westinghouse (about 3,614 tonnes a year of rod and assembly capacity), Framatome (3,450), Rosatom's TVEL (2,760) and Global Nuclear Fuel (1,630).^[46] Fabrication looks competitive and largely is, with one qualification that turned out to matter enormously after 2022: assemblies are reactor-specific, qualification takes years, and utilities therefore have "limited choice in suppliers of fabricated fuel assemblies, especially for PWRs."^[46] Eastern Europe's VVERThe Soviet and Russian pressurised-water reactor family, and the design exported across Eastern Europe and beyond. Its fuel assemblies are specific to it, which is why a VVER operator cannot simply buy fuel from somebody else. fleet was built to run on Russian fuel and nothing else. Since 2022, all of Ukraine's VVER fuel for its nine operating reactors has come from Westinghouse's Swedish plant.^[46] That substitution took roughly a decade of prior qualification work, begun long before anyone needed it. It is the exception that shows the rule.

3The machine: how fission becomes electricity, and why light water won

A nuclear power station is a steam plant with an unusual boiler. The turbine, the condenser, the generator, the switchyard — all of that is nineteenth- and twentieth-century thermal engineering, and it is why nuclear plants convert heat to electricity at 33–37% efficiency, much like a coal plant and rather worse than a modern gas turbine.^[52] Everything distinctive sits upstream of the steam.

In a pressurised water reactor, water in the primary circuit is held at about 325 °C under roughly 150 atmospheres, pressure high enough to stop it boiling. It carries heat from the fuel to a steam generator, where it boils a physically separate secondary loop; the secondary steam drives the turbine and never touches the core.^[52] Two barriers between the fuel and the machinery, and a turbine hall that is not radioactive. In a boiling water reactor the water boils in the core itself at about 285 °C and 75 atmospheres, and that steam goes straight to the turbine — simpler, cheaper, one circuit, but the turbine is now part of the radiological envelope.^[52][5] Both are light-water reactors, both burn 3.5–5% enriched uranium oxide, and together they account for 351 of the world's 417 operating units.^[3]

The other families exist, and each teaches something. Pressurised heavy-water reactors — CANDU and its Indian derivatives, 45 units — use heavy water as moderator and coolant, run on natural uranium, and refuel on load through horizontal pressure tubes without shutting down.^[52] That last property is elegant and, from a safeguards perspective, awkward: on-load refuelling makes it easier to remove fuel at the low burnupHow much energy has been extracted from a given mass of fuel before it is discharged, conventionally quoted in gigawatt-days per tonne. Low burnup means the fuel came out early — which is exactly the condition that leaves the plutonium inside it in a form a weapons programme can use. that yields weapons-useful plutonium. Britain's gas-cooled reactors run carbon dioxide at 650 °C over graphite-moderated cores and reach about 41% thermal efficiency, the best of any commercial type — a design that won on thermodynamics and lost on capital cost and construction time.^[52] The Soviet RBMK, a light-water-cooled graphite-moderated pressure-tube machine, was cheap, big and refuellable on load, and it carried the flaw that destroyed Chernobyl. Ten remain in operation, all in Russia, all extensively modified; the type "has never been built outside the Soviet Union."^[52] Two commercial fast reactors run today, both Russian, with no moderator at all: they use fast neutrons and can extract more than sixty times as much energy from a given quantity of mined uranium.^[52]

Defence in depth, and what an operating plant actually does

Reactor safety is organised around a doctrine rather than a device: successive independent barriers, each assumed to fail. The ceramic fuel pellet retains most fission products. The zirconium cladding seals the rod. The reactor pressure vessel and primary circuit contain the coolant. The containment building — reinforced concrete, often a metre thick, with a steel liner — encloses the lot. Around those sit the engineered safety systems: emergency core cooling, residual heat removal, containment spray, redundant and diverse power supplies. Fukushima is instructive precisely because the barriers held in sequence and then failed in sequence, and the reason they failed was that redundancy without diversity is not redundancy: every diesel generator was in the same place, and the water came for all of them at once.

An operating plant is a remarkably steady thing. Fuel resides in the core for four to six years, reaching burnups near 50 gigawatt-days per tonne in the United States, up from around 25 in the 1980s.^[52] Roughly a third of the assemblies are replaced at each refuelling outage every 12, 18 or 24 months. In between, the plant runs. The industry's measure for that steadiness is the capacity factor: the electricity a plant actually produced over a period, divided by what it would have produced had it run at full rated output every hour of it. Load factor is the IAEA's name for the same idea. It is the number that decides how far a plant's enormous fixed cost gets spread, which is why it matters far more to a reactor than to any machine whose costs are mostly fuel. The IAEA's global median load factor was 86.4% in 2024, with the best quartile of reactors averaging 93.7%; boiling water reactors led the decade at a 90.4% median.^[3] American reactors averaged 91.0% capacity factor across calendar 2025.^[9] These are the highest utilisation figures of any generating technology in commercial use, and they are the reason nuclear's contribution to electricity runs well ahead of its share of installed capacity.

Reactor "generations" are a marketing vocabulary that has become useful shorthand. Generation I was the prototypes of the 1950s and early 1960s: Calder Hall, Shippingport, Dresden-1. Generation II is the great commercial fleet of the 1970s and 1980s — the machines that still supply most of the world's nuclear electricity today. Generation III and III+ are the designs of the 1990s onward, with more passive safetySafety that works by physics rather than by equipment: cooling driven by gravity, natural circulation and stored water, so the plant can protect itself without pumps, without electrical power, and without an operator deciding anything. and longer design lives: the AP1000 (1,250 MWe gross, passive safety, US certification 2005), the EPR (1,750 MWe, four loops), Korea's APR1400 (1,455 MWe, US certification May 2019), Russia's VVER-1200 and China's Hualong One.^[53] Generation IV, the six concepts under international coordination, is not yet commercial anywhere at utility scale — with one exception discussed in Part IV, and it broke ground in Wyoming in April 2026.^[58]

4The back end: heat, time, and a political problem dressed as a technical one

The volume of high-level nuclear waste is trivially small, its hazard falls by more than 99% within fifty years, and no country has yet opened a permanent repository for it. All three statements are true, and the gap between them is the subject.

Spent fuel comes out of a reactor intensely radioactive and thermally hot. It goes first into a water-filled pool at the plant, where the water provides both cooling and shielding, and stays there for years while the shortest-lived fission products decay. After that it can be moved to dry casks: steel canisters inside concrete overpacks, air-cooled, sitting on a pad at the site. The United States holds roughly 95,000 tonnes of used fuel, 43% in pools and 56% in dry storage, essentially all of it at current or former power-plant sites.^[39] Worldwide, about 430,000 tonnes have been discharged since civil nuclear power began; some 30% has been reprocessed and about 263,000 tonnes remains in storage.^[10]

The decay curve is the fact most often missing from public discussion. Within about forty to fifty years of removal from the reactor, the heat and radioactivity of spent fuel fall by more than 99%. After roughly a thousand years most of the radioactivity is gone; the residue is comparable in hazard to natural uranium ore, which reprocessed high-level waste reaches after about 9,000 years and unreprocessed spent fuel after about 300,000.^[10] The often-quoted quarter-million-year horizon is real, and it describes the tail, not the danger. The danger is front-loaded and it fades fast.

So is the volume. High-level waste is about 3% of the volume of radioactive waste produced and 95% of its radioactivity; low-level waste is 90% of the volume and 1% of the radioactivity.^[10] The world's entire inventory of solid high-level waste destined for disposal amounts to roughly 29,000 cubic metres — the WNA's illustration is a three-metre-tall building covering a football pitch.^[10] Seventy years of civilisation-scale electricity generation has produced a hazardous residue you could walk around in an afternoon. That is the energy-density argument arriving in its physical form.

Once-through or closed: the recycling argument

A light-water reactor extracts only a small fraction of the energy in the heavy metal it loads. Oklo's own SEC filing states the figure from the developer's side of the argument: conventional plants "only use approximately 5% of the energy content stored in nuclear fuel before needing to refuel."^[60] What comes out is therefore not waste in the ordinary sense. It is about 95% uranium, roughly 1% plutonium bred from U-238 during irradiation, and a few per cent fission products — and the uranium and plutonium are fuel.

Reprocessing separates them. The PUREX process dissolves the fuel in nitric acid and extracts uranium and plutonium by solvent extraction; the plutonium is blended into mixed-oxide (MOX) fuel and burned again. World commercial reprocessing capacity is about 3,860 tonnes a year — La Hague in France at 1,700 t/yr for light-water fuel, Sellafield's Magnox line at 1,500, Rokkasho in Japan at 800 (long delayed), Russia's Mayak at 400, India at 260. MOX production capacity is around 480 tonnes a year, predominantly French. Reprocessing cuts high-level waste volume to about a fifth of the original spent fuel and recovers 25–30% more energy from the original uranium.^[47]

Against which: it is expensive, it produces separated plutonium — the most proliferation-sensitive material in the civil fuel cycle — and it does not eliminate the need for a repository. The United States abandoned commercial reprocessing in the 1970s on exactly the second ground and has never returned. France, Russia, Japan, China and India persisted. Neither camp has been vindicated. The recycling question is an economic and political question wearing technical clothes: the technology works, the plutonium is real, and the choice turns on what a country believes about the price of uranium in fifty years and about who else will copy what it builds.

Geologic disposal: Onkalo opens, Yucca did not

Every serious national programme has reached the same technical conclusion: deep geologic disposal, several hundred metres down in stable rock, with engineered barriers around the waste. Finland got there first. Posiva's Onkalo facility at Olkiluoto uses the multi-barrier KBS-3 system: copper capsules enclosing boron-steel canisters of twelve fuel assemblies each, surrounded by bentonite clay, emplaced 400–450 metres down in the Olkiluoto bedrock, with an initial capacity of 6,500 tonnes of used fuel expandable to 12,000.^[66][13] Posiva applied for its operating licence on 30 December 2021; the regulator STUK began its formal review in May 2022 and had been given until 31 December 2025 to report.^[11] In December 2025 that deadline was extended to the end of June 2026, STUK citing "deficiencies in the documentation, updates due to plant modifications made by Posiva," and unresolved "uncertainties in the safety case review", particularly on the performance of the clay barrier.^[12] STUK missed that deadline too, saying on 30 June 2026 that it expected no further material from Posiva and would deliver its safety assessment to the ministry as soon as it was complete, without naming a date.^[66] Onkalo will almost certainly be the world's first operating repository. It is not yet operating, and the delays are documentary rather than geological.

The American story is the counterexample, and it is not primarily technical. The Nuclear Waste Policy Act of 1982 obliged the federal government to begin removing used fuel from reactor sites by 1998. Congress designated Yucca Mountain in Nevada as the sole candidate site in 1987, approved it in 2002, and DOE submitted a construction licence application in 2008; the project was defunded from 2009. The NRC nonetheless completed its technical safety review in 2014–16 and found the requirements met.^[39] Meanwhile utilities have paid $24.6 billion into the Nuclear Waste Fund since 1983, the fund balance stood at $51.4 billion at the end of FY2025, the government has paid out nearly $9 billion in damages for failing to take title to the fuel, and estimated remaining liabilities run to $30.8 billion.^[39] The Government Accountability Office, reviewing the same impasse, counted about 86,000 tonnes of commercial spent fuel stored at 75 sites, growing by roughly 2,000 tonnes a year, and recommended that Congress amend the Nuclear Waste Policy Act to authorise "a new consent-based process for siting, developing, and constructing" both interim and permanent facilities — a recommendation that has not been enacted.^[84]

The United States has spent, and expects to spend, roughly $40 billion in damages for not building a repository, while holding $51 billion collected to build one. That is not the signature of a technical obstacle. It is the signature of a political one, and it has a specific consequence downstream: interim dry storage at reactor sites, which was designed as a stopgap, has become the de facto American disposal policy by default. Whether that is dangerous is a separate question from whether it is a policy, and the honest answer to the first is that dry casks have performed without radiological incident for forty years while the second question goes unanswered.

51938–1945: the physics, and the programme that owned it first

In December 1938 Otto Hahn and Fritz Strassmann in Berlin bombarded uranium with neutrons and found barium among the products — an element little more than half uranium's weight, where the chemistry said there should be something heavier. Lise Meitner and Otto Frisch, working through the Christmas holidays in Sweden, supplied the interpretation: the nucleus had split, and the mass deficit had become energy on the scale Einstein's relation predicted. Within months, Joliot in Paris and Szilard and Fermi in New York had confirmed experimentally that fission released further neutrons, which meant a chain reaction was possible.^[44]

Almost nothing about the following six years was about electricity. The physics arrived in Europe in the winter before the war, and it was immediately and correctly understood as a weapons problem. Everything afterwards carries the mark of it: nuclear technology was born inside a military programme, under secrecy, on a wartime budget, with no requirement that it be economic and every requirement that it be fast.

On 2 December 1942 at 3:36 p.m., in a converted squash court under the abandoned stands of Stagg Field at the University of Chicago, forty-nine scientists led by Enrico Fermi withdrew the last control rod from a pile of graphite blocks and uranium and watched the neutron count climb and hold. Chicago Pile-1 achieved the world's first controlled, self-sustaining nuclear chain reaction.^[56] It ran at half a watt. There was no shielding, no containment and no coolant; the safety system was a rod on a rope, an axe, and a man with a bucket of cadmium solution on the balcony. It worked because of delayed neutrons, though the team's confidence that it would rested on a calculation rather than on a demonstration.

Chicago Pile-1's purpose was not power. The Metallurgical Laboratory's assignment was to prove that a reactor could breed plutonium-239 from uranium-238 at industrial scale, because plutonium was the second route to a bomb and the easier one to separate chemically. The pile at Chicago led to the production reactors at Hanford; the enrichment problem led to the gaseous-diffusion plants at Oak Ridge. Both paths converged in July and August 1945: the Alamogordo test on 16 July, a uranium-235 gun weapon over Hiroshima on 6 August, a plutonium implosion weapon over Nagasaki on 9 August.^[44]

Three inheritances from those years still shape the civil industry. Secrecy became the default posture toward nuclear information, and the reflex outlasted its justification by decades. The state, not the market, became the natural sponsor of nuclear projects, because only a state had ever built one. And the fuel cycle acquired a permanent second meaning: every enrichment plant and every reprocessing line built for electricity is a facility whose military application was demonstrated before its civil one.

61953–1960: Atoms for Peace, and how a submarine engine became the world's power reactor

On 8 December 1953, addressing the UN General Assembly in New York, President Eisenhower proposed that governments "begin now and continue to make joint contributions from their stockpiles of normal uranium and fissionable materials to an International Atomic Energy Agency" set up under the aegis of the United Nations. The agency would hold the material and, more importantly, "devise methods whereby this fissionable material would be allocated to serve the peaceful pursuits of mankind" — including "abundant electrical energy in the power-starved areas of the world." Contributing powers would thereby be "dedicating some of their strength to serve the needs rather than the fears of mankind."^[55]

"It is not enough to take this weapon out of the hands of the soldiers. It must be put into the hands of those who will know how to strip its military casing and adapt it to the arts of peace."

Dwight D. Eisenhower, address to the UN General Assembly, 8 December 1953

Read as diplomacy the speech is a Cold War manoeuvre, and it was. Read as institutional design it is the founding document of the entire international nuclear order. The IAEA that resulted became simultaneously the promoter of civil nuclear power and its inspector — a dual mandate that critics have called a conflict for seventy years and that is, more accurately, the bargain itself. Access to the technology in exchange for accepting verification. Everything in §15 flows from this speech.

The machines came quickly. The Soviet Union connected a 5 MWe reactor at Obninsk in June 1954, the first to supply electricity to a grid. Britain's Calder Hall, a 50 MWe gas-cooled Magnox unit, started in 1956 and is generally called the first commercial-scale station — though it was built to produce plutonium for weapons and sold electricity as a by-product. Shippingport, a 60 MWe pressurised water reactor in Pennsylvania, started in 1957. By 1960 Yankee Rowe and Dresden-1, both 250 MWe, were operating.^[44] Inside six years the industry went from prototype to something recognisably commercial.

Rickover's inheritance: why the world runs light-water reactors

The decision that determined the shape of the global fleet was not made on the merits of civil power generation at all.

Admiral Hyman Rickover's naval reactors programme needed a power plant that would fit inside a submarine hull, run for years without air, tolerate violent manoeuvring, and be as compact as physics allowed. Compactness pointed at water: it is a superb moderator and a superb coolant in the same substance, so the core can be small. Small cores need enriched fuel, and the United States, alone in the world, had industrial-scale enrichment plants left over from the Manhattan Project. USS Nautilus, the first nuclear submarine, put to sea in 1955; by 1962 the US Navy had 26 nuclear submarines operational and 30 more building.^[48]

Shippingport was, in substance, a naval reactor scaled up and put ashore — the design and the design team both came out of the same programme. Westinghouse and General Electric then carried that engineering into the commercial market, and the American export machine carried it to Europe and Japan. France, after building its own gas-graphite line in the 1960s, abandoned it and licensed Westinghouse PWR technology for its national build-out.^[25] The result is visible in Figure 9: 351 of 417 operating reactors are light-water machines.

The "roads not taken" question deserves a straight answer rather than a romantic one. Molten-salt reactors were demonstrated at Oak Ridge: the Molten Salt Reactor Experiment ran from January 1965 to December 1969, logged more than 13,000 hours at full power, and on 8 October 1968 became the first reactor anywhere to run on uranium-233 — the thorium cycle's fissile product — before the programme was wound down.^[86] Thorium fuel cycles were studied seriously and dropped. High-temperature gas-cooled reactors were built and abandoned in Germany and the United States. Sodium-cooled fast breeders absorbed enormous sums in France, Japan, Britain and the United States and produced two operating commercial units, both Russian. None of these were beaten in a fair engineering contest. Light water won because it had a decade's head start, a trained workforce, an existing supply chain, an enrichment industry paid for by the weapons programme, and a navy that kept ordering. Once the first hundred units were light-water, the qualification costs, the regulatory precedent and the trained operators locked the choice in. Path dependence, not superiority.

Whether that lock-in was costly is genuinely open. Light water is a mature, well-understood technology with an enormous operating base, and maturity is worth a lot. It also requires high pressure, requires enrichment, cannot make process heatHeat used directly for industrial work — refining, chemicals, hydrogen, desalination — rather than converted into electricity. The temperature a reactor's coolant can reach decides which of those uses it can serve at all. above about 330 °C, and burns roughly 5% of the heavy metal it loads. A fleet built on high-temperature gas or fast spectrum would have different problems, not fewer. What can be said is that the world did not choose its reactor technology; it inherited it from a submarine.

71960s–1970s: the great build-out, "too cheap to meter," and France's decision

Figure 12 makes the arc plain. Construction starts ran in single digits through the early 1960s, then 25 units in 1967, 37 in 1968, 37 in 1970, 38 in 1974, 38 in 1975, 43 in 1976. Installed capacity went from 2.2 GW in 1962 to 118 GW in 1979.^[3] In roughly fifteen years the world built a new heavy industry from very little.

Two forces drove it. The first was a genuine expectation that nuclear electricity would be extremely cheap — the phrase "too cheap to meter" belongs to a 1954 speech by Lewis Strauss, chairman of the US Atomic Energy Commission, and it has been quoted against the industry ever since. The expectation was not absurd. Fuel cost almost nothing, plants ran continuously, and early units came in on budget because they were small, simple and built by utilities with regulated returns and captive ratepayers. The second force was the oil shock. Crude prices quadrupled in 1973–74 and again in 1979, and every industrial economy that imported oil went looking for something that did not come by tanker from the Persian Gulf.

France: what a serial programme looks like

France is the clearest natural experiment in the history of the industry, and it is invoked loosely far more often than it is understood.

After the 1974 oil shock the French government decided to expand nuclear capacity rapidly using licensed Westinghouse pressurised-water technology, and it did so as a national programme rather than a series of projects: one utility, one regulator, one design family, standardised in a small number of series, built continuously.^[25] The result is 57 operable reactors totalling 63,000 MWe, all operated by EDF, supplying 67.3% of French electricity in 2024 — the highest national share on earth — and making France Europe's largest electricity exporter.^[3][25] The state completed the renationalisation of EDF in October 2023, raising its stake from 84% to 100%.^[25]

The uncomfortable coda is that France has since lost the capability that produced this. Flamanville 3, the country's first new reactor in twenty-five years, was estimated in 2005 at €3.3 billion with commercial operation expected in May 2012. It connected to the grid in December 2024 and reached full power in December 2025, at a cost of €13.2 billion — four times the estimate, seventeen years from first concrete rather than five.^[25] France's planned EPR2 programme of six reactors at Penly, Gravelines and Bugey carried a December 2025 cost estimate of €72.8 billion, with first concrete expected in 2027.^[25] The same country, the same utility, the same regulator. What changed was that the serial programme stopped, the supply chain aged out, and the institutional memory of how to build a reactor went with the people who had done it. That is not a French failing. It is the general mechanism, and §13 makes the argument properly.

81979 and 1986: the two accidents that changed different things

Three Mile Island did almost no radiological harm and transformed how reactors are regulated and operated. Chernobyl did enormous harm and changed what the public believes. Neither lesson is the one usually drawn.

Three Mile Island, 28 March 1979

The mechanism was mundane and the consequences were not. TMI Unit 2, a PWR near Harrisburg, Pennsylvania, was at 97% power when a secondary cooling malfunction tripped the turbine and scrammedA scram is the emergency shutdown of a reactor: control rods dropped into the core in seconds, killing the chain reaction. It stops fission at once. It does nothing about the decay heat that follows, which is why "off" is not "safe". the reactor. A pilot-operated relief valve opened as designed to shed pressure and then failed to close, and the control room had no instrument that showed the valve's actual position — only that it had been commanded shut. Coolant escaped for over two hours. High-pressure injection started automatically, and the operators, seeing the pressuriserThe vessel that sets and holds the pressure of a pressurised water reactor's primary circuit, part steam and part water. Operators read its water level to judge how much coolant the system is holding — which is what made a misleading reading so costly here. level rise and believing the system was going solidFilling the primary circuit entirely with water, leaving no steam space to cushion pressure changes. Operators are trained hard to avoid it, which is precisely why the belief that it was happening produced the wrong action., throttled it back. The core uncovered. At least 45% of it melted, some 62 tonnes, with 19 tonnes relocating into the lower plenum. The block valve was finally closed at 6:22 a.m.^[18]

Radiologically, almost nothing happened outside the plant. The average dose to people within ten miles was about 0.08 millisievertsA thousandth of a sievert, the unit of radiation dose weighted for the biological harm a given kind of radiation does. It is the unit that lets a chest X-ray, a transatlantic flight, a CT scan and a reactor accident be put on one scale — which §14 does., roughly a chest X-ray; the maximum individual dose was about 1 mSv, around a third of annual US background.^[18] The President's Commission — the Kemeny Commission, reporting in October 1979 — projected the total number of radiation-induced cancers among the affected population as an expected value of 0.7, and stated flatly: "There will either be no case of cancer or the number of cases will be so small that it will never be possible to detect them." It concluded that "the major health effect of the accident was found to be mental stress."^[20] A Pennsylvania registry followed more than 30,000 residents within five miles for eighteen years and found "no evidence of any abnormal number of cancers around TMI."^[18]

What the Commission actually indicted was not the equipment. Its overall conclusion was that preventing accidents as serious as TMI would require "fundamental changes in the organization, procedures, and practices — and above all — in the attitudes of the Nuclear Regulatory Commission and, to the extent that the institutions we investigated are typical, of the nuclear industry." And, in the sentence that ought to be carved above the door of every control room: "if the only problems were equipment problems, this Presidential Commission would never have been created. The equipment was sufficiently good that, except for human failures, the major accident at Three Mile Island would have been a minor incident."^[20]

The industry took that seriously in ways that are measurable. Operator training moved to a symptom-based approach — maintain core cooling first, diagnose the malfunction second — rather than requiring operators to identify the fault before acting. The Institute of Nuclear Power Operations was founded in 1979 on the Commission's recommendation as an industry self-policing body, with the National Academy for Nuclear Training following in 1985.^[18] Significant plant events per US reactor fell from 2.38 in 1985 to 0.10 in 1997, and median capability factor rose from about 65% in 1980 to above 90%.^[18] The reason American reactors now run at 91% capacity factor is, in a direct causal line, the accident at Three Mile Island.

The cost side was the opposite of benign. Regulatory requirements ratcheted, backfits were ordered on plants already under construction, review times lengthened, and the interest bill on half-finished projects compounded. Orders had already stopped — construction starts fell from 43 in 1976 to 27 in 1979 — but TMI ensured that nothing restarted, and it turned dozens of partially built plants into cancellations.^[3]

Chernobyl, 26 April 1986

Chernobyl is the accident that actually killed people, and its mechanism was a design pathology, not an operating error, though the operators supplied the occasion.

The RBMK-1000 had a positive void coefficient: when cooling water in the pressure tubes turned to steam, reactivity increased, because in a graphite-moderated core the water's main effect is neutron absorption rather than moderation. Worse, the control rods had graphite followers on their lower ends, so that inserting a fully withdrawn rod initially displaced water with graphite in the bottom of the core and briefly added reactivity there — the "positive scram" effect. That effect had been discovered at the Ignalina plant in 1983. Nothing was done about it, no compensating measures were taken, and the finding was not disseminated to operating stations.^[19]

During a test of the turbine's coast-down capability on 26 April, with automatic protection disabled and the reactor in a xenon-poisonedXenon, one of the fission products, absorbs neutrons voraciously. It builds up when a reactor's power is pulled down and then decays away over hours, so a machine in that state is fighting its own poison and is hard to hold steady at low power., low-power, unstable configuration outside anything the procedures contemplated, the operators pressed the scram button. INSAG-7, the IAEA's revised 1992 verdict, is precise about what happened next and admirably candid about how its own earlier report had been wrong: "Addition of further positive reactivity by insertion of the control and safety rods that had been fully withdrawn during the test was probably a decisive contributory factor. This latter effect was a result of faulty design of the rods."^[19] The power excursion fragmented the fuel, flashed the coolant, blew off the 1,000-tonne cover plate, jammed the rods and ruptured the fuel channels; a steam explosion and probably a hydrogen explosion followed.^[16] There was no containment building of the kind Western reactors had. About 14 exabecquerelsA becquerel is one radioactive disintegration per second; an exabecquerel is a billion billion of them. The unit counts decays, not dose — how much harm follows depends on which nuclides they are, where they land, and who is standing there. of radioactivity went into the atmosphere.^[16]

INSAG-7's central judgement is worth stating in its own words, because it moved the blame decisively away from the night shift: the accident "can be said to have flowed from deficient safety culture, not only at the Chernobyl plant, but throughout the Soviet design, operating and regulatory organizations for nuclear power that existed at the time," and the weight INSAG-1 had given in 1986 to the Soviet account, "which laid blame almost entirely on actions of the operating staff, is thereby lessened."^[19]

The human toll, as assessed by UNSCEARThe United Nations Scientific Committee on the Effects of Atomic Radiation — the UN body that reviews and reports the evidence on radiation exposure and its health effects. It is the closest thing this field has to an agreed scorekeeper, and this primer leans on it heavily.: of about 600 workers on site that morning, 134 received doses of 0.8 to 16 grayThe gray (Gy) measures the energy a tissue actually absorbs, per kilogram, before any weighting for the kind of radiation delivering it. For the gamma and beta exposures at Chernobyl it runs close to the sievert, so one gray here is near enough one thousand millisieverts on the ladder in §14. and suffered acute radiation syndrome; 28 died within three months, and a further 19 died between 1987 and 2004 of causes "not necessarily associated with radiation exposure." Roughly 350,000 people were evacuated or relocated. Average doses were about 120 mSv for 530,000 recovery workers, 30 mSv for 115,000 evacuees, and 9 mSv over two decades for those who remained in contaminated areas — against a global average natural background of 2.4 mSv per year.^[15] By 2005 more than 6,000 thyroid cancers had been diagnosed among those exposed as children, and "it is most likely that a large fraction of these thyroid cancers is attributable to radioiodine intake."^[15]

That last item is the one that should anger a reader, because it was avoidable by administrative action alone. Iodine-131 has an eight-day half-life and reaches children through milk. Distributing stable iodine and banning contaminated milk for a few weeks would have prevented most of those cancers. The Soviet authorities delayed. UNSCEAR's comparison with Fukushima makes the contrast explicit: Japanese food and water restrictions were "timely and effective", "whereas, in some areas of the Former Soviet Union, restrictions were delayed resulting in very high doses to the thyroids of those affected", and thyroid doses to Fukushima evacuees were about a hundred times lower than Chernobyl's.^[14]

Beyond thyroid cancer, UNSCEAR's conclusion is more restrained than the public memory: "Apart from the dramatic increase in thyroid cancer incidence among those exposed at a young age, and some indication of an increased leukaemia and cataract incidence among the workers, there is no clearly demonstrated increase in the incidence of solid cancers or leukaemia due to radiation in the exposed populations." And: "there were widespread psychological reactions to the accident, which were due to fear of the radiation, not to the actual radiation doses."^[15]

91987–2010: the wilderness, and the renaissance that fizzled

For twenty years after Chernobyl the Western nuclear industry did one thing well: it operated the plants it already had. Figure 12 shows construction starts running at two to six units a year worldwide through the 1990s, with a single year — 1995 — recording none at all.^[3] The fleet crept from 416 reactors in 1990 to 440 in 2005, and the additions were overwhelmingly in Asia.

The proximate causes are usually listed as public opposition and regulation. Both mattered. Neither is sufficient, and the operating data points somewhere else. Between 1980 and 2000 the existing fleet's median capability factor climbed from about 65% to over 90%.^[18] That improvement is worth roughly a third of the fleet in extra output, obtained at almost no capital cost — the cheapest new generation available anywhere, and it went on for two decades. Meanwhile natural gas got cheap, combined-cycleA gas plant that runs a gas turbine and then uses its exhaust heat to raise steam for a second, steam turbine — two cycles from one lot of fuel, and the most efficient way to burn gas for electricity. It is also the quickest and cheapest large generator to build, which is what made it nuclear's competitor rather than its complement. plants could be built in three years by a private developer with modest capital at risk, and electricity markets across the OECD were deregulated in ways that transferred construction risk from ratepayers to shareholders. A technology whose entire cost is up-front, whose schedule is measured in decades and whose regulatory environment can change mid-build is exactly the technology that deregulation punishes hardest.

The 2000s "renaissance" was real as an intention and thin as a construction programme. Rising gas prices, early carbon policy and a generation of new Gen III+ designs produced a wave of announcements; in the United States, the Energy Policy Act of 2005 offered federal loan guarantees under Title XVII, of which Vogtle 3 and 4 later became the flagship borrower. What actually got built in the West was Olkiluoto 3 in Finland (construction start 2005, grid connection 2022 — 200 months by the IAEA's count), Flamanville 3 in France, and eventually Vogtle 3 and 4 in Georgia.^[3] Shale gas arrived in 2008 and removed the price signal. The financial crisis removed the capital. By 2010 the renaissance was, in the OECD at least, a set of first-of-a-kindThe first unit built anywhere to a given design. It carries all the discovery — drawings that turn out not to fit, welders learning the procedure, a regulator seeing the thing for the first time — which makes its cost the least informative number in the industry about what the design will eventually cost. projects running late.

Nuclear construction is a serial-learning industry, and the mechanism recurs in Part IV. The phrase is doing real work, so it is worth spelling out. Cost in this business falls through repetition rather than through invention: the same design, built again by the welders, inspectors, planners and crane operators who built the last one, gets cheaper and quicker each time round. The knowledge lives in people and in supply chains rather than in drawings, which is the whole trouble — it walks out of the door when the orders stop, and a drawing cannot be handed to somebody who has never built the thing. The IAEA's own data shows what happens when the series breaks: worldwide median construction time fell to 59 months in 2001–05, when the units being completed were mostly Asian and mostly repeat builds, and has since risen to 102 months in 2021–24.^[3] The industry is not getting worse at physics. It is building fewer, more novel, less repeated machines with supply chains and workforces that have to be reassembled each time.

102011: Fukushima, and the difference between harm and consequence

At 14:46 local time on 11 March 2011 a magnitude 9.0 earthquake struck off Honshu. The reactors at Fukushima Daiichi scrammed correctly; the grid connection failed and the emergency diesel generators started, as designed. Forty-one minutes later a tsunami of about fifteen metres overtopped the site's seawall, flooded the generator and switchgear rooms, and removed all AC power and the ultimate heat sinkThe body of water or air a plant finally dumps its waste heat into — sea, river, cooling tower. Every cooling system in a reactor is in the end just a path to it, so losing it means the heat has nowhere to go however much else is still working. at once.^[17]

What followed is §1 in physical form. With cooling gone, decay heat at roughly 1.5% of nominal thermal power kept boiling away the inventory. Fuel in units 1, 2 and 3 uncovered and melted, unit 1's fuel reaching around 2,800 °C. Zirconium cladding oxidised in steam, liberating hydrogen, which accumulated in the reactor buildings and exploded — unit 1 at about 25 hours after scram, unit 3 at about 68 hours — destroying the upper structures.^[17] Releases of the radiologically significant nuclides, iodine and caesium, were about 10% of Chernobyl's, and only about 20% of the atmospheric release drifted over Japanese land; the rest went over the Pacific.^[14]

The health findings, from the United Nations Scientific Committee on the Effects of Atomic Radiation's 2020/2021 report, are unambiguous and are not what most people believe. "No adverse health effects among Fukushima residents have been documented that could be directly attributed to radiation exposure from the accident, nor are any expected to be detectable in the future." The Committee found "no credible evidence of excess birth defects, stillbirths, premature births or low birthweights". Increases in cardiovascular and metabolic conditions among evacuees were observed and are "probably associated with concomitant social and lifestyle changes and are not attributable to radiation exposure."^[14]

On the thyroid question, which was the great fear: "Although a substantial number of thyroid cancers have been detected among exposed children, the Committee believes that, on the balance of available evidence, the (relative to expected) large increase in thyroid cancers is the result of ultrasensitive screening procedures that have revealed the prevalence of thyroid abnormalities not previously recognized in the population, and is not a result of radiation exposure."^[14] The Committee went further, warning that over-diagnosis from mass ultrasound screening "has the potential to cause considerable anxiety among some of those screened and to lead to unnecessary treatment, the detrimental effects of which may outweigh those of the radiation exposure itself."^[14] That is a UN scientific committee saying, in the most careful language available to it, that the screening did more harm than the accident.

Workers: more than 20,000 emergency workers received an average effective dose of about 13 mSv to March 2012; 174 workers (0.8%) exceeded 100 mSv; no worker has received more than 50 mSv in a year since April 2013.^[14] An increase in cancer incidence "is unlikely to be discernible amongst workers."

And yet. There were 2,313 disaster-related deaths among evacuees from Fukushima prefecture, predominantly among people over 66 — deaths from the disruption of moving frail people out of hospitals and care homes, not from radiation.^[17] That number is larger than the entire radiological death toll of the civil nuclear industry's history outside Chernobyl. It points in a direction that neither side of the nuclear debate finds comfortable: the response to the accident killed people and the accident did not.

The consequences, which had nothing to do with the dose

Japan shut down its entire fleet. Before the accident, 54 reactors and 47.5 GWe supplied about 30% of Japanese electricity; by 2013 all were offline pending review against new Nuclear Regulation Authority standards.^[42] Fossil imports rose by ¥3.6–4.0 trillion a year, and the carbon intensity of Japanese electricity jumped from 350 g/kWh in 2011 to 487 g/kWh in 2012.^[42] As of April 2026, 15 reactors have restarted, 10 more are in the approval process, and 27 have been permanently shut down.^[42] Fifteen years to restart fewer than a third of the fleet.

Germany went further and faster. Before Fukushima it drew a quarter of its electricity from 17 reactors; a 2000 phase-out agreement had been partially reversed in 2010 with eight- and fourteen-year licence extensions. Within days of the accident Chancellor Merkel ordered the eight oldest reactors — 8,336 MWe, about 6.4% of German generating capacity — shut immediately, and legislated an exit. The last three units, Isar 2, Neckarwestheim 2 and Emsland, closed in April 2023.^[41] German nuclear capacity is now zero, down from 20,400 MWe. Coal supplied 120 TWh (24%) of German electricity in 2024 and gas 89.1 TWh (17%).^[41]

An earthquake and tsunami off Japan killed more than 20,000 people. The nuclear accident that followed produced no documented radiation deaths. In response, a country eight thousand kilometres away, at no meaningful tsunami risk, closed a fleet of reactors that were generating about 140 TWh a year of zero-carbon electricity, and burned more coal. Whatever else this is, it is not a proportionate response to a radiological hazard, and treating it as one obscures what actually happened: Fukushima did not change the risk of German reactors, it changed the political price of defending them. That distinction is the mechanism by which the industry's fortunes actually move.

112020s: the pivot back, and why this time is differently motivated

Three forces reopened the case, and none of them is the one that closed it.

Climate targets that need firm capacity. Nuclear's lifecycle emissions are among the lowest of any source: the IPCC's median across peer-reviewed studies is 12 gCO₂-equivalent per kilowatt-hour, comparable to wind and lower than all types of solar; a March 2022 UNECE assessment put nuclear at 5.1–6.4 g, the lowest of any low-carbon technology.^[51] Grids with very high shares of wind and solar discover that the last 10–20% of decarbonisation is the expensive part, and that firm, weather-independent capacity has a value that levelised cost comparisons do not capture. At COP28 in Dubai on 2 December 2023, twenty-five countries launched a declaration committing to "work together to advance a global aspirational goal of tripling nuclear energy capacity from 2020 by 2050."^[35]

Energy security after February 2022. Russia's invasion of Ukraine turned an abstract dependency into an operational one. On 13 May 2024 the United States enacted the Prohibiting Russian Uranium Imports Act, banning imports of unirradiated low-enriched uranium produced in Russia or by a Russian entity from 90 days after enactment, with waivers available only where "no alternative viable source" exists, subject to declining annual caps of 476,536 kg in 2024 falling to 459,083 kg in 2027, all waivers terminating by 1 January 2028 and the prohibition itself running to 31 December 2040.^[29] Two months later, on 9 July 2024, the ADVANCE Act became law, restructuring NRC fees for advanced reactor applicants, adding provisions for micro-reactors, brownfield sitingBuilding on land that already carried heavy industrial use — most usefully a retiring coal plant, which comes with a grid connection, cooling water, a road, a permit history and a workforce. The attraction is everything that does not have to be built again., non-electric applications and fusion regulation.^[30]

Demand. This is the genuinely new element, and it is the subject of §21. For twenty years electricity demand in the OECD was flat or falling, which meant that no new baseloadThe steady, round-the-clock floor of electricity demand — and, by extension, the plants built to serve it: large, cheap to run, designed to sit at full output rather than to chase the daily peaks. capacity was needed at any price. That ended. Datacentre construction produced a class of buyer with an unusual set of requirements — very large blocks of power, delivered continuously, carbon-free on an hourly basis, contracted for twenty years — and an unusual willingness to pay for them. In September 2024 Constellation announced a twenty-year power purchase agreementA contract to buy a plant's output at an agreed price for an agreed term. It is usually the thing that makes a capital-heavy project financeable at all, because it converts uncertain future revenue into a promise a lender will lend against — which is exactly the problem §13 shows nuclear has. with Microsoft to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center, restoring about 835 MWe that had been shut for economic reasons exactly five years earlier.^[33]

What has actually been delivered so far is modest, and the arithmetic belongs on the table before Part IV gets enthusiastic. Global nuclear capacity ended 2025 at 420 GW, with 3 GW added and 3 GW retired during the year.^[2] The 2020 baseline the declaration uses was 369.8 GW across 414 reactors.^[3] Tripling that by 2050 means about 1,109 GW, which requires adding roughly 28 GW a year for twenty-five years. The industry has managed 3.^[2]

12The grid: what firm power is worth, and what nuclear actually contributes

A grid is a machine that must balance supply and demand continuously, second by second, and its cost is driven less by average generation than by the hardest hours. That is where nuclear's characteristics matter, and they are unusual in one specific way: a reactor's output is essentially independent of weather, season, time of day and fuel delivery.

That property has a name in the trade, and the rest of this primer leans on it: firm power. A firm source is one the system operator can count on for a particular hour and plan the rest of the grid around — not one that produces a great deal on average, but one whose output is there when it is called for. The difference is not a detail. A grid has to be built for its worst hour rather than its typical one, so a megawatt that shows up on demand and a megawatt that shows up on average are two different products that happen to share a unit.

Capacity factor is the measurable expression of it. US utility-scale nuclear generators ran at 91.0% of nameplate capacity across calendar 2025; geothermal managed 65.9%, hydroelectric 35.3%, solar photovoltaic 24.4% and wind 23.6%.^[9] Globally, the IAEA reports a median load factor of 86.4% across the fleet in 2024, with the best quartile averaging 93.7%.^[3]

Two qualifications. First, capacity factor is not the same as value: solar's 24% arrives in daylight when demand is high, which is worth considerably more than 24% of randomly distributed hours. Second, a high capacity factor is partly a choice. Reactors are run flat out because their marginal cost is near zero and their capital cost is enormous, so every idle hour is pure loss. Boiling water reactors "can operate in load-followingVarying a plant's output through the day to track demand instead of running flat out. Technically ordinary; economically painful for a machine whose costs are all fixed, because every megawatt-hour not sold is revenue that never comes back. mode more readily than PWRs", and the French fleet, which is too large a share of its own grid to be run as pure baseload, load-follows routinely.^[52] Nuclear plants can follow load. They simply prefer not to, for reasons of economics rather than physics — and as wind and solar shares rise, that preference becomes an increasingly expensive one, because a plant whose economics require 90% utilisation is poorly matched to a market that periodically prices power at zero.

Where does nuclear sit in decarbonisation? In 2025 it supplied 8.9% of world electricity and 20.9% of the low-carbon total.^[8] In the European Union it remained the largest single source of electricity at 23% of generation, and eight of the ten countries with the highest nuclear shares were EU members.^[8] In the United States nuclear provided 18.2% of electricity in 2024 and over 40% of low-carbon power.^[3][38] In China, despite the world's largest construction programme, nuclear is only 4.6% of generation and has been roughly unchanged since 2019 — not because the reactors are not being built, but because total Chinese generation is growing as fast as the reactors are added.^[8]

That Chinese figure is the clearest available demonstration that even an aggressive, uninterrupted, state-financed nuclear build cannot outrun electricity demand growth in a fast-industrialising economy. Nuclear's share is not falling because nuclear is failing. It is falling because the denominator is winning.

13The economics: why almost-free fuel produces expensive electricity

A paid-off American reactor produces the cheapest firm power on the continent at $30–38 per megawatt-hour. A new one produces some of the most expensive at $141–220. Same machine, same fuel, same physics. The difference is capital and time.

The fuel is what makes everything else strange. Lazard's 2025 analysis assumes a nuclear fuel cost of $0.85 per million BTU, against $3.45 for natural gas and $1.47 for coal.^[22] Georgia Power's own 2025 disclosures put its cost of nuclear fuel at 0.83 cents per net kilowatt-hour against 3.37 cents for gas and 3.75 for coal — nuclear fuel is a quarter the cost of the alternatives per unit of electricity produced.^[21] A nuclear plant is close to being a machine that turns capital directly into electricity with a rounding error of fuel in between.

Which is precisely the problem. When almost all of a technology's lifetime cost is incurred before it earns anything, the price of that electricity is dominated by two variables: how much the plant cost to build, and what rate of return the money demanded while it was being built.

The capital cost, and the gap between the model and the build

The US Energy Information Administration's engineering estimates, prepared by Sargent & Lundy in 2023 dollars, put the overnight capital costWhat a plant would cost if it could be built overnight — equipment, labour and materials, with no interest charged for the years of waiting. A clean engineering comparison and a misleading commercial one, because on a nuclear project the interest is a large part of what makes it expensive. of a two-unit AP1000 plant on a brownfield site at $7,861 per kilowatt and a six-module small modular reactor plant at $8,936 per kilowatt. A combined-cycle gas plant is $868, onshore wind $1,489, ultra-supercritical coal $4,103.^[23] Nuclear is roughly nine times gas per kilowatt of capacity installed.

Georgia Power's share of Plant Vogtle Units 3 and 4 was certified by the Georgia Public Service Commission at $4.418 billion. Its final net investment, disclosed in Southern Company's 2025 Form 10-K filed 19 February 2026, was $10.670 billion, excluding approximately $440 million of capitalised financing costs accrued through Unit 4's in-service date.^[21] Georgia Power owns 45.7% of the units. Lazard, working from public estimates, puts total project capital cost at approximately $32.3 billion for about 2.2 GW — roughly $14,500 per kilowatt, against the EIA's engineering estimate of $7,861.^[22]

Flamanville 3 tells the same story in euros: €3.3 billion estimated in 2005, €13.2 billion delivered, and seventeen years from first concrete to grid instead of five.^[25] Hinkley Point C, EDF's British project, was re-estimated in January 2024 at £31–34 billion in 2015 money with unit 1 operational between 2029 and 2031^[25] — and revised again in EDF's 2025 annual results on 20 February 2026 to £35 billion in 2015 sterling with unit 1 in 2030, the delay attributed to productivity in electromechanical installation rather than to civil works, with EDF stating that a further year would add about £1 billion more.^[80] Three estimates in twelve years, each higher than the last, on a plant whose design was certified before the first concrete was poured.

Which makes the British response to that record the most interesting policy experiment now running. On 22 July 2025 the UK government took a final investment decision on Sizewell C, a two-unit EPR station of the same design as Hinkley, at a stated capital cost of around £38 billion in 2024 prices — presented as roughly a 20% saving against Hinkley on a like-for-like design — financed for the first time in nuclear under the Regulated Asset Base model, with the government holding 44.9% alongside La Caisse at 20%, Centrica at 15%, EDF at 12.5% and Amber Infrastructure at 7.6%.^[79]

The RAB mechanism is the direct institutional answer to the discount-rate problem. Under a contract for difference — Hinkley's structure — the developer carries the entire construction risk and is paid nothing until electricity flows, so the cost of capital reflects a decade of single-asset construction exposure. Under RAB, consumers pay a regulated charge during construction (the government's own figure is about £1 a month on an average bill over the build), which shifts a defined portion of that risk onto the ratepayer base and, by construction, lowers the return investors must demand. Investors remain exposed to overruns, but they are no longer exposed to the whole of them.^[79] Whether Sizewell C's ~20% claimed saving materialises is the single most informative Western experiment of the next decade — not because the reactor is different, but because the financing is, and this primer's argument is that the financing is what the cost actually was.

The cost of capital is the whole argument

Two pieces of vocabulary carry everything that follows. The levelised cost of electricity is the attempt to reduce a power plant to one comparable number: everything it costs across its life — building it, borrowing for it, fuelling it, running it, taking it down — divided by every megawatt-hour it is expected to produce. The discount rate is the price the money charges for waiting: the annual return investors require while capital sits in a half-finished asset earning nothing. For a gas plant, cheap to build and expensive to feed, the second number barely disturbs the first. For a reactor it decides it.

Because nuclear's cost is almost all up-front, its levelised cost is more sensitive to the discount rate than any other generating technology's — and the discount rate is not a property of the technology. It is a property of the political and regulatory environment the technology is built in.

The OECD Nuclear Energy Agency and the IEA put it in a single sentence in Projected Costs of Generating Electricity 2020: "The more capital-intensive a technology, the more sensitive its LCOE are to changes in the discount rate. Among baseload plants, this means that in particular the costs of nuclear new build depend on the discount rate. With a low discount rate of 3%, reflecting a stable market environment with high investment security, the LCOE of new nuclear plants is lower than for new coal and gas plant. With higher discount rates at 7% or 10%, which would reflect riskier economic environments, the costs of a newly built nuclear plant would exceed those of fossil fuel-based plants."^[24] The IEA, joint publisher of the same study, states the finding in identical terms on its own report page, and the underlying country results show the range the sentence compresses: nuclear LCOE from $27/MWh in Russia to $61 in Japan at 3%, and from $57 to $146 at 10%.^[85]

At 3%, nuclear is the cheapest baseload option available. At 10%, it is the most expensive. Nothing about the reactor changed. What changed is whether investors believe the project will be finished on the schedule promised, under the regulatory regime that existed when it was ordered. The report is explicit that "in the real world the question of who bears the risk is important: Government support, in the form of price guarantees for example, would shift the risk from the investor to the public… Although the overall risk remains the same, the investment would thus become cheaper from an investor perspective."^[24]

This is why South Korea, France in the 1970s, and China today build reactors at a fraction of Western costs, and why the explanation is rarely about labour rates. Those programmes had a single buyer, a standard design, a continuous order book and a state that absorbed the schedule risk. Korea's fleet reaches capacity factors up to 96.5%, among the highest in the world; the four APR1400 units at Barakah in the UAE — a $20.4 billion contract signed in December 2009 — all reached commercial operation by September 2024, an on-schedule delivery of a first-of-a-kind export project that the Western industry has not matched in fifty years.^[43] China's Hualong One is targeted at $2,800–3,000 per kilowatt, with construction typically 50–60 months from first concrete to grid connection.^[37]

The other side of the same coin: existing plants are extraordinarily cheap

If nuclear's cost is capital, then a nuclear plant whose capital has already been paid off should be exceptionally cheap to run. It is. Lazard puts the marginal cost of operating a fully depreciated US nuclear plant, including decommissioning provision, at $30–38 per megawatt-hour — below gas combined cycle at $24–39 on its low end and far below coal at $31–114.^[22] The NEA's assessment of long-term operation — the refurbishment programme that extends a reactor from forty years to sixty or eighty — computes levelised costs of $26–49 per megawatt-hour across the plausible range of refurbishment costs, discount rates and capacity factors, and concludes that lifetime extension "remains not only the least cost option for low-carbon generation… but for all power generation across the board."^[24][26]

The two facts together make the shape of rational nuclear policy uncomfortably clear. Extending an existing reactor costs $26–49/MWh. Building a new one in the West costs $141–220/MWh, and Lazard's illustrative figures for Vogtle 3 and 4 specifically are $169–228/MWh.^[22] Utility-scale solar is $38–78 and onshore wind $37–86 on the same basis, though neither is firm.^[22] Every megawatt of existing nuclear capacity that a government allows to close and then replaces with new nuclear has performed one of the most expensive substitutions available in the energy system. Germany closed 20.4 GW. Japan has 27 units permanently shut. The United States has spent a decade arguing about how to build reactors while, until very recently, letting operating ones retire on economics.

14Radiation: the dose, the risk, and the gap between them

Ionising radiation is measured in sieverts, a unit that weights absorbed energy by the biological damage different radiation types cause. The numbers that matter are all in the same units, so they can be compared directly.

Natural background radiation delivers a global average of about 2.4 millisieverts per year, typically in the range of 1–10 mSv depending on geology, altitude and radon.^[50] That range is not narrow. In Ramsar, Iran, natural background reaches up to 250 mSv a year with no identified health effects; in Kerala, India, gamma dose alone exceeds 15 mSv a year.^[50] An abdominal CT scan is about 10 mSv; a long intercontinental flight about 0.03 mSv; aircrew receive up to about 5 mSv a year.^[50] Occupational limits for nuclear workers are 20 mSv a year averaged over five years with 50 mSv in any single year; the public limit is 1 mSv a year.^[50] Acute radiation sickness begins around 1,000 mSv delivered in a short period; 4,000–5,000 mSv short-term kills about half of those exposed.^[50]

What is genuinely uncertain is what low doses do. Radiation protection standards use the linear no-threshold model, which assumes that any dose carries proportionate risk with no safe floor. UNSCEAR is careful about the model's status: it assesses risk by asking whether a theoretically calculated increase in disease "would be detectable compared to the normal statistical variability in the baseline incidence" — a formulation that concedes the model produces numbers which cannot be observed.^[14] The ICRPThe International Commission on Radiological Protection — the standing expert body whose recommendations national radiation-protection rules are generally written from. It sets the convention; it does not run the epidemiology. recommends LNT "for optimising radiation protection practices" rather than for predicting actual low-dose health effects.^[50]

Apply LNT to a very small dose spread across a very large population and you get a non-zero number of theoretical deaths — and the same arithmetic applied to natural background would attribute hundreds of thousands of cancers a year to granite and radon. The model is a conservative regulatory convention. It is not a measurement, and treating its output as a body count is a category error that has done real damage to public reasoning about nuclear risk, in both directions.

The strongest published challenge to that reading

The screening-effect conclusion is not unanimous, and a primer that presented it as settled would be overstating. Kato, Yamada and Hongyo, writing in Cancers in September 2023, analysed Fukushima's thyroid ultrasound data — 300,473 examinees in the first screening round and 270,511 in the second — and reported a linear dose-response relationship between UNSCEAR's own estimated thyroid doses and annual incidence rates across four prefecture areas. Their argument turns on a point that deserves to be taken seriously: an overdiagnosis artefact should not correlate with radiation dose, because screening sensitivity does not vary with exposure. "Overdiagnosis can never be a substitute for radiation effect," they write. To reconcile their finding with the observed cases they conclude that UNSCEAR underestimated thyroid doses by a factor of fifty to a hundred.^[73]

The dose-response logic is sound in principle and is the correct test to apply. The difficulty is the size of the correction it requires: a fifty- to hundred-fold error in dose reconstruction by a committee that had direct thyroid measurements, whole-body counting and food-monitoring data available to it, and whose estimates were reviewed by thirteen independent critical reviewers.^[14] The authors themselves note that age- and sex-stratified data were unavailable and that fine-needle aspiration rates differed materially between screening rounds (26.5% against 11.4%), which is precisely the sort of procedural variation that can generate an apparent gradient across areas screened at different times.^[73]

The screening explanation remains the weight of expert opinion and the position of the UN scientific committee, and it is contested in the peer-reviewed literature by an argument that is methodologically legitimate rather than fringe. Note also what the dispute does not touch. Nobody claims deaths from thyroid cancer at Fukushima; the disease's survival rate is very high and no fatalities have been attributed. The comparison that carries this primer's fourth conclusion — zero documented radiation deaths against 2,313 disaster-related deaths among evacuees — is unaffected either way.

The uncomfortable synthesis

Three Mile Island: no radiation deaths, an expected 0.7 statistical cancers, mental stress as the dominant health effect.^[20] Fukushima: no documented radiation-attributable health effects and none expected to be detectable; a thyroid cancer increase that UNSCEAR attributes to screening rather than radiation; 2,313 disaster-related deaths among evacuees.^[14][17] Chernobyl: 28 acute radiation deaths, more than 6,000 thyroid cancers largely attributable to radioiodine, no clearly demonstrated increase in solid cancers or leukaemia in the wider exposed population, and psychological reactions "due to fear of the radiation, not to the actual radiation doses."^[15]

The pattern is consistent and it is not the one in public circulation. In the two accidents at reactors with containment, radiation harmed essentially no one and the response harmed many. In the accident at a reactor without containment, radiation harmed thousands — and the largest single component of that harm, childhood thyroid cancer, was preventable by handing out iodine tablets and stopping the milk supply.

None of this argues that reactor accidents are acceptable, and it should not be read that way. It argues something narrower and more useful: the dominant hazard of a severe reactor accident in a country with a functioning public-health apparatus is the disruption, not the dose. That has direct operational implications — for how evacuation zones are drawn, for whether frail patients are moved, for whether mass ultrasound screening is a kindness — and those implications are still not fully reflected in emergency planning anywhere.

15Regulation and the bomb: the bargain at the centre of the system

Two regulatory systems govern nuclear power and they answer to different constituencies. National regulators — the US Nuclear Regulatory Commission, France's ASNR, Finland's STUK, Japan's NRA, Canada's CNSC — license and inspect facilities for safety within their own territory. The International Atomic Energy Agency verifies, on behalf of the international community, that declared nuclear material is not being diverted to weapons.

The legal architecture rests on the 1968 Treaty on the Non-Proliferation of Nuclear Weapons, whose bargain is straightforward: non-weapon states forgo weapons and accept IAEA safeguards on all their nuclear material; weapon states pursue disarmament; and all parties retain the right to develop nuclear energy for peaceful purposes. Article III requires each non-weapon state party to bring a comprehensive safeguards agreement into force with the Agency, and the IAEA reports each year on which states have not.^[31] The Additional Protocol, developed after Iraq's clandestine programme was uncovered in 1991, extends the Agency's access beyond declared facilities to the question of whether undeclared ones exist.

The scale of the verification enterprise is documented annually. In 2024 the IAEA applied safeguards for 190 states with agreements in force. One hundred and thirty-seven had both a comprehensive safeguards agreement and an additional protocol; for 75 of those the Secretariat concluded that "all nuclear material remained in peaceful activities" — the so-called broader conclusion, which requires the Agency to find no indication of undeclared material or activities anywhere in the state. For the other 61, evaluation of the undeclared question "remained ongoing", so the conclusion was limited to declared material. Forty-five states had a comprehensive agreement but no additional protocol; for 31 of those the Agency could conclude only that declared material remained in peaceful use, and for 14 with small-quantities protocols on the original text it could draw no conclusion at all. Three NPT parties had still not brought a safeguards agreement into force; for them no conclusion was possible. The Agency did not implement safeguards in the Democratic People's Republic of Korea and could draw no conclusion there.^[31]

The safeguards system is not a global surveillance net. It is a graduated set of assurances whose strength depends entirely on what each state has agreed to let the Agency see, and for a substantial minority of states the assurance is materially weaker than the headline suggests. That is not a criticism of the IAEA, which cannot verify what it is not permitted to inspect. It is a description of the instrument.

The Iranian case shows where the dual-use tension actually bites. The Agency reported that during 2024 Iran "continued to accumulate enriched uranium well beyond the limits agreed under the JCPOAThe Joint Comprehensive Plan of Action — the negotiated agreement that set ceilings on Iran's enrichment level and stockpile in exchange for sanctions relief. It is invoked here as the benchmark those ceilings are measured against, not as a live constraint. and expanded considerably its enrichment capacity", remains "the only NPT non-nuclear-weapon State that is producing and stockpiling high enriched uranium", and held an estimated stockpile at the end of 2024 of 3.9 significant quantities of UF₆ enriched up to 60% U-235.^[31] The Agency also recorded that it had "lost continuity of knowledge in relation to the production and current inventory of centrifuges, rotors and bellows, heavy water and uranium ore concentrate, which it will not be able to restore."^[31]

None of that involved a reactor. It involved centrifuges — the same machines, in the same physical arrangement, that Urenco operates in New Mexico to fuel American power plants. This is the dual-use problem in its exact form, and it explains a great deal about why nuclear power is regulated as it is: the international community cannot supply the fuel-cycle technology that makes civil nuclear power possible without also supplying most of the capability that makes a weapon possible. Export controls, the Nuclear Suppliers Group, fuel-supply assurances and multinational enrichment consortia are all attempts to sell the electricity without selling the bomb. None of them fully succeeds, and the fact that this bargain has held for fifty-eight years with fewer than a dozen weapon states is a better outcome than most analysts in 1963 would have predicted.

The export-control half of that architecture is where the dual-use tension is actually administered. The Nuclear Suppliers Group — 48 participating governments, with the European Commission as observer — publishes its rules through the IAEA as INFCIRC/254, in two parts. Part 1, the "Trigger List", covers items designed or prepared for nuclear use: reactors, reprocessing plants, isotope-separation equipment, heavy-water production, conversion facilities. Its conditions of supply are the operative clauses — IAEA safeguards on the recipient's material, an assurance of non-explosive use, physical protection to a stated criterion, controls on retransfer, and special restraint on enrichment and reprocessing technology specifically.^[87] Part 2, most recently issued as INFCIRC/254/Rev.12/Part 2 on 29 July 2022, covers dual-use equipment — items with ordinary industrial applications that would also "make a major contribution to an unsafeguarded nuclear fuel cycle".^[88] The structure is the same insight as §2's, written as trade law: the goods themselves are innocent, and it is the enrichment and reprocessing steps that carry the special controls, because those are the steps where electricity and weapons stop being distinguishable by the equipment involved.

16The actors: who controls which part, and what breaks if they stop

Laid out along the chain, the dependencies become obvious. What follows maps where the load sits; it does not rank companies.

The criticality test, applied

Ask of each step: who controls it, how replaceable is it, how long would a replacement take, and what fails downstream?

After February 2022 Western utilities discovered they had a Russian dependency, and it was not in uranium — Russia mined only 4.5% of world output in 2024.^[7] It was in enrichment, where Rosatom held 44% of world capacity, and in conversion, where it holds 12,500 tonnes of the roughly 62,000 tonnes of licensed capacity.^[6][45] Centrus, the largest American supplier of enrichment services, still describes its Russian TENEX contract as "our largest supply contract" in its FY2025 annual report, and notes that a market-related price reset in that contract "has significantly lowered our cost of sales and contributed to improved margins since 2019."^[32] The company also states baldly that "the United States has not had domestic uranium enrichment capability suitable to meet U.S. national security requirements since the Paducah GDP shut down in 2013."^[32]

The reactors were never the vulnerable link. Neither were the mines. The vulnerability was two chemical and mechanical processing steps that nobody talks about, performed at a handful of sites, on assets that take the better part of a decade and a licensing process to replicate.

17The two branches most explanations leave out

Navies: the lineage, and the enduring entanglement

Naval nuclear propulsion is not a footnote to civil power. It is the parent programme, and it remains larger in units than most national civil fleets. More than 200 small reactors have powered over 160 vessels, mostly submarines. The US Naval Nuclear Propulsion Program — jointly a Navy and Department of Energy organisation, reporting to Congress annually — states that in more than seventy-five years of operation "there has never been a reactor accident, nor any release of radioactivity that has had an adverse effect on human health or the quality of the environment", and that no person in the programme has ever exceeded the federal annual or lifetime radiation exposure limit; the current cumulative record is over 7,400 reactor-years and more than 171 million miles steamed.^[81][82] Russia claims more than 6,500 reactor-years to 2015.^[48] Naval cores are designed for exceptional life — ten years minimum between refuellings, with modern cores intended to last thirty to fifty years, effectively the life of the boat.^[48]

The enrichment is where the entanglement is visible. US submarine fuel was originally around 97% U-235 and is approximately 93% in the latest designs; the UK has used comparably high enrichment; Russia runs 20–45% depending on generation; India's Arihant uses 40%. France has moved the other way, transitioning to roughly 5–7%.^[48] A nuclear submarine reactor running on 93% enriched uranium is, materially speaking, carrying weapons-grade material through international waters under a naval-fuel exemption that the NPT's safeguards architecture does not cover. Nobody designed it that way as a loophole. It is a loophole nonetheless, and it becomes more consequential as more states — most recently Australia under the AUKUS arrangement — acquire nuclear-powered submarines without possessing nuclear weapons.

The technical lineage runs the other way too. Rickover's PWR became the world's civil reactor (§6); the small modular reactors of §19 are, in several cases, explicitly derived from naval-scale engineering; and the workforce and forging capacity that a state maintains for its navy is the same workforce and capacity a civil programme draws on. Countries that gave up naval nuclear propulsion generally lost civil nuclear manufacturing capability with it, and countries that kept it — the United States, France, Russia, China, and now the United Kingdom through the Rolls-Royce lineage — retained a supplier base that is very difficult to recreate from a standing start.

Isotopes: the branch with no substitute and a fragile supply

Over fifty million nuclear medicine procedures take place worldwide each year, and technetium-99m accounts for roughly 80% of them and about 85% of diagnostic scans.^[49] Tc-99m has a six-hour half-life, which means it cannot be stockpiled; it is generated on site from molybdenum-99, which has a 66-hour half-life and therefore cannot be stockpiled either. Every Tc-99m scan performed anywhere on earth this week depends on a reactor having irradiated a target within the last few days.

Six reactors supply essentially the entire world's Mo-99: BR-2 in Belgium (1961), the HFR at Petten in the Netherlands (1961), OPAL in Australia (2006), LVR-15 in the Czech Republic (1989), Safari-1 in South Africa (1965) and Maria in Poland (1974).^[49] Five of the six are more than fifty years old. This is not an incidental observation: the OECD Nuclear Energy Agency has run a standing High-level Group on the Security of Supply of Medical Radioisotopes since April 2009, created specifically because the 2009–10 shortage exposed how few and how old the producing reactors are, and its most recent demand-and-capacity projection — the first since 2019 — models supply security through 2027 on the explicit premise that the fleet's age and the sector's inability to recover full costs are the binding problems rather than any technical one.^[83] Canada's NRU, long a mainstay, ceased Mo-99 production in October 2016 and closed in March 2018.^[49] The 2008–09 shortage, when two of the ageing reactors went down simultaneously, cancelled diagnostic procedures across North America and Europe for months.

There is a proliferation dimension here as well, and it was resolved rather elegantly. Mo-99 was historically produced by irradiating targets of highly enriched uranium — weapons-usable material, shipped internationally, for medical purposes. As of March 2023 all Mo-99-producing reactors use low-enriched targets instead, at a cost of roughly 20% more per unit.^[49] That is what a solved proliferation problem looks like: a technical substitution, a modest cost penalty, and two decades of patient multilateral work.

The therapeutic side is growing fast and has its own bottlenecks. Lutetium-177 and actinium-225 target tumours directly rather than imaging them; Ac-225 in particular has been supply-constrained for years, with commercial production only recently expanding — TerraPower Isotopes began production in October 2024, Cardinal Health started weekly production in December 2024, and NorthStar began routine weekly production in late 2025.^[49]

The isotope supply chain is the clearest small-scale demonstration of every structural feature described in this primer: extreme concentration, ageing assets, no substitute for the physics, a dual-use input that had to be engineered out, and a public that has no idea any of it exists until it fails.

18Who is actually building

Over the past decade, 94% of reactors that started construction anywhere in the world were of Chinese or Russian design. Everything else in this section is a comment on that sentence.

The figure comes from the IEA's Global Energy Review 2026.^[2] At the end of 2024 the IAEA counted 62 reactors under construction across 15 countries totalling 64.5 GWe, of which China accounted for 28 units and 29.6 GWe — nearly half the world's expansion by capacity.^[3] By early 2026 the count had risen to roughly 74–80 units and 78–79 GW, "one of the highest levels seen in the last 30 years", with China expected to reach 100 GW of installed capacity by around 2030.^[2][1]

China's programme is a serial programme in the sense §7 described, and it delivers what serial programmes deliver. Sixty-three reactors operating and 38 under construction; a 2030 target of 110 GWe under the 15th Five-Year Plan approved in March 2026; construction typically 50–60 months from first concrete to grid connection; Hualong One targeted at $2,800–3,000 per kilowatt.^[37] The Shidaowan CAP1400 demonstration unit entered commercial operation in late 2024 at an expected first-unit cost of about $2,450 per kilowatt.^[37] Set those against the EIA's $7,861/kW engineering estimate for an American AP1000 and Lazard's ~$14,500/kW implied figure for Vogtle, and the gap is not marginal — it is a factor of three to five.

Chinese cost figures should be read with appropriate scepticism: state financing, state land, state labour arrangements and a cost of capital that no private developer faces all flatter the number, and the accounting is not audited to Western standards. But the schedule data is harder to dismiss, because grid connection dates are observable. China connected 38 reactors and 39.0 GWe to the grid in the decade from 2014, and its median construction time in 2021–24 was 80 months against a worldwide median of 102 and a Korean median of 124.^[3]

Russia's model is different and arguably more strategically potent. Rosatom is a state corporation of over 400,000 employees spanning mining, conversion, enrichment, fuel, reactor supply, construction, operation and waste, and it is involved in the construction of about twenty reactors abroad.^[36] Its export offer is not a reactor; it is a package — design, construction, financing, fuel supply for the life of the plant, operator training, and in the Akkuyu case in Türkiye, ownership of the plant itself under a build-own-operate structure. A country that accepts that package acquires electricity and simultaneously acquires an eighty-year relationship with the Russian state covering the only fuel its reactors can burn. Egypt has four VVER units under construction, Türkiye four, Bangladesh two, and Russia began construction of Leningrad 2-3 at home in March 2024.^[3]

South Korea is the one Western-aligned exporter with a demonstrated record. The $20.4 billion Barakah contract signed in December 2009 delivered four APR1400 units, all in commercial operation by September 2024; Korean reactors reach capacity factors up to 96.5%; and KHNP was selected as preferred bidder for two units at Dukovany in the Czech Republic in July 2024, signing in June 2025 at approximately $18.6 billion with first unit online expected by 2036.^[43] Even the industry's most credible Western-aligned builder is quoting eleven years from contract signature to first power.

The West's own record is the one already examined: Vogtle at a final Georgia Power net investment of $10.670 billion for a 45.7% share against $4.418 billion certified; Flamanville 3 at €13.2 billion against €3.3 billion estimated; Hinkley Point C re-estimated at £31–34 billion in 2015 money with unit 1 between 2029 and 2031.^[21][25] Britain has two units under construction, Slovakia one, Ukraine two.^[3] Britain has since added a third and fourth by taking final investment decision on Sizewell C in July 2025 under the Regulated Asset Base model (§13), and pushed Hinkley unit 1 to 2030 at £35 billion in 2015 sterling in February 2026.^[79][80]

India belongs in this picture and is routinely left out of it, because its programme is neither an export business nor a stalled one. It is a domestic serial build of an indigenous design. Sixteen 700 MWe Indian-designed pressurised heavy-water reactors are under construction or sanctioned across the country; Rajasthan unit 7, the third of them to get there, reached full 700 MW output on 10 February 2026 after first criticality in September 2024 and grid connection in March 2025, and further units are sanctioned at Kaiga, Gorakhpur, Chutka and Mahi Banswara. NPCIL's stated national objective is at least 100 GW of nuclear capacity by 2047.^[74] The IAEA counted India among the countries building at end-2024, and the IEA puts India alongside Egypt and Türkiye at roughly 5 GW under construction each.^[3][2] India fits §7's pattern exactly: one design family, one utility, a continuous order book, and no requirement to satisfy a private capital market — the same conditions that produced France's fleet, arrived at by a different political route. What India does not yet demonstrate is speed; the 700 MWe units have taken roughly a decade each.

19Small modular and advanced reactors: what has actually happened

The SMR argument runs as follows. Nuclear's cost problem is a construction problem; construction problems are worst on large, bespoke, site-built projects; therefore build small reactors in factories, ship modules, and substitute manufacturing learning for construction craft. Add passive safety that works without power or operator action, which permits smaller emergency planning zonesThe area around a plant within which evacuation, sheltering and iodine distribution have to be planned for, drilled and paid for. Its radius follows from how much material a severe accident could release, so a design that can release less can in principle be sited closer to the customer it is powering. and cheaper sites. More than a hundred SMR designs are at some stage of development worldwide.^[27] The logic is sound. The evidence is early, mixed, and worth separating carefully into what has been licensed, what has been financed, and what has been built.

Licensed. NuScale's original 50 MWe module received US design certification, the first SMR to do so, and on 29 May 2025 the NRC approved its upratedUprating is licensing a design or an existing plant to run at higher power than it was originally approved for — more output from substantially the same hardware, bought with analysis rather than with construction. 77 MWe (250 MWt) module, scalable to twelve modules and 924 MWe.^[61] TerraPower received an NRC construction permit for the Natrium reactor on 9 March 2026 — "the first construction permit ever issued by the NRC for a commercial non-light-water power reactor" and the first such application docketed in over forty years. The application was submitted in March 2024, accepted in May 2024, and the safety review completed in December 2025, ahead of schedule and 11% under budget.^[57] Kairos Power received test-reactor construction permits for Hermes in December 2023 and for Hermes 2 in November 2024.^[62]

Built, or being built. TerraPower began nuclear construction at Kemmerer Unit 1 in Wyoming on 24 April 2026: a 345 MWe sodium-cooled fast reactor with molten-salt heat storage that can boost output to 500 MWe for periods, targeted for completion in 2030, with Bechtel as EPC contractorEngineering, procurement and construction — the firm that takes responsibility for designing the works, buying the equipment and building the plant, rather than merely supplying labour to somebody else's plan. and about 1,600 construction workers.^[58] Non-nuclear site work had begun in June 2024.^[57] The Canadian Nuclear Safety Commission issued Ontario Power Generation a licence to construct one BWRX-300 — a 300 MWe water-cooled, natural-circulationCoolant moved by heat alone — hot water rises, cool water sinks — rather than by pumps. Fewer pumps means fewer things that stop working when the electricity does, which is the same instinct as passive safety. small modular reactor — at the Darlington New Nuclear Project site in April 2025, valid to 31 March 2035; the first regulatory hold point was removed on 30 March 2026, clearing foundation installation, and OPG applied for a 20-year operating licence in March 2026.^[54] Kairos broke ground on Hermes 2 in Oak Ridge in April 2026, targeting operation by 2030, and separately requested an extension of the Hermes construction completion date from 31 December 2026 to 30 April 2029.^[62]

Made critical, by a different route. The most striking development of the past year did not go through the NRC at all. Executive Order 14301, signed on 23 May 2025, directed the Department of Energy to reform reactor testing and stand up a pilot programme with the goal of at least three advanced reactor concepts reaching criticality outside the national laboratories by 4 July 2026; eleven projects were selected on 12 August 2025.^[68] Four made it. Antares Nuclear's Mark-0 achieved first criticality in early June 2026, Valar Atomics' Ward 250 shortly after, Deployable Energy's Unity on 1 July, and Aalo Atomics' Critical Test Reactor at Idaho National Laboratory at 00:20 local time on 4 July — under eight months from groundbreaking to criticality.^[67]

That is a genuine result and it should be read precisely. These are DOE-authorised test reactors under the department's own authority, not NRC-licensed commercial power plants; the pathways are legally distinct.^[68] None of the four supplies electricity to anything — Aalo's first power-producing unit is planned for 2027.^[67] What the programme demonstrates is not that advanced reactors are commercially ready. It is that the calendar time between deciding to build a small reactor and making it critical can be eight months rather than eight years when the regulatory pathway is different. Whether that speed survives contact with a commercial licence, a public utility commission and an insurer is the open question, and it is a different question from the one the licensing-reform debate has been asking.

Cancelled. On 8 November 2023 NuScale and Utah Associated Municipal Power Systems terminated the Carbon Free Power Project — the flagship American SMR deployment, six modules at Idaho National Laboratory, supported by roughly $1.4 billion of DOE funding — because it "appears unlikely that the project will have enough subscription to continue toward deployment."^[63] The proximate cause was cost: the target price per megawatt-hour rose far enough that the municipal utilities would not subscribe. That is the single most informative event in SMR history so far, and it happened to the most advanced project with the most regulatory progress.

The steelman for SMR scepticism, argued properly

The strongest version of the case against deserves to be beaten rather than dismissed.

Nuclear power has always exhibited economies of scale for a physical reason: the reactor vessel's cost scales roughly with surface area while its output scales with volume, so a bigger core is cheaper per kilowatt. That is why reactors grew from 60 MWe at Shippingport to 1,600 MWe at Flamanville. An SMR deliberately abandons that advantage and proposes to recover it through factory learning — which requires an order book large enough to run a factory, which requires customers, which requires a price, which requires the factory to already exist. The EIA's engineering estimate quantifies exactly this: $8,936 per kilowatt for a six-module SMR plant against $7,861 for a two-unit AP1000, in the same study on the same basis.^[23] On the analysis, small is more expensive per kilowatt, not less. And the one flagship American SMR project with regulatory approval, subsidy and a customer consortium was cancelled because the price went the wrong way.^[63]

Add the fixed costs that do not shrink with the reactor. A control room, a security force, an operating licence, an emergency plan, a regulatory relationship and a spent-fuel programme cost roughly the same whether the reactor is 77 MWe or 1,100. TerraPower expects about 250 full-time employees at a 345 MWe plant.^[58] A 1,100 MWe unit does not need four times as many. Per megawatt-hour, small plants carry heavier overhead.

The reply is real and should be stated with equal force. First, factory production genuinely does change the cost structure of anything ever built that way, and nuclear has simply never tested it. Second, the binding constraint on Western nuclear is not cost per kilowatt in an engineering study; it is that no private investor will absorb a ten-year, ten-billion-dollar, single-asset construction risk. A 300 MWe unit that costs more per kilowatt but can be financed, built in four years and repeated may deliver more capacity than a 1,600 MWe unit that is cheaper on paper and never gets ordered. Third, several advanced designs are not competing on electricity price at all: high-temperature gas reactors produce process heat above 700 °C for industrial applications that electricity cannot serve, and microreactors compete against diesel at remote sites where the alternative costs many hundreds of dollars per megawatt-hour.

Which side is right is not yet knowable, and the useful move is to say so and name the evidence that will settle it. The test is not a licence, an approval or a groundbreaking. It is the delivered cost and schedule of the second, third and fourth units of a single design — because a first unit tells you about the design and the fourth tells you about the factory.

20HALEU and the re-shoring of enriched fuel

Most advanced reactor designs will not run on the fuel the world currently makes. High-assay low-enriched uranium — HALEU, defined as uranium enriched between 5% and 20% U-235 — is required by the Natrium, Xe-100, Aurora and Kairos designs among others, because higher enrichment permits smaller cores, longer intervals between refuellings, simpler designs and lower waste volumes.^[32] The trouble is that essentially none of it exists commercially outside Russia.

Centrus began enrichment operations at its HALEU facility in Piketon, Ohio on 11 October 2023 under contract to the Department of Energy, and made its first HALEU delivery on 7 November 2023, completing Phase 1 of the demonstration.^[32] Production is running at roughly one metric tonne of HALEU a year from the pilot cascade, with a stated objective of up to six tonnes a year by the first quarter of 2029.^[64] Congress, in the FY2024 National Defense Authorization Act, directed DOE to make 21 metric tonnes of HALEU available on a schedule: 3 tonnes by 30 September 2024, 8 tonnes by 31 December 2025, and 10 tonnes by 30 June 2026; on 9 April 2025 DOE announced conditional commitments in the first allocation round to five developers — TRISO-X, Kairos Power, Radiant Industries, Westinghouse and TerraPower — with three of them requiring fuel delivery during 2025.^[64][72]

Twenty-one tonnes. A single Natrium first core is expected to require several tonnes; Oklo has been awarded five tonnes for one Aurora.^[60] The entire national HALEU programme, as legislated, supplies first cores for a handful of demonstration units and nothing beyond. Meanwhile the Prohibiting Russian Uranium Imports Act closes the only large existing source, with all waivers terminating by 1 January 2028.^[29]

Enrichment capacity is fungible between products: SWU spent making HALEU is SWU not spent making ordinary reactor fuel. A Western enrichment industry with about 25 million SWU a year of capacity — Urenco's 17.9 million plus Orano's 7.5 million — is being asked simultaneously to replace 44% of world capacity that has become politically unavailable, to fuel a fleet whose life is being extended, to fuel new large reactors, and to stand up an entirely new high-assay product line.^[6] The WNA projects world capacity rising to 70.3 million SWU a year by 2030, a 14% increase over eight years.^[6] Demand is not rising by 14%.

The dated project record turns that arithmetic from an assertion into a constraint, and it is more sobering than the aggregate. Orano is extending Georges Besse II at Tricastin by four modules on top of the existing fourteen: +2.5 million SWU, a 30% increase, for a total investment of nearly €1.7 billion part-financed by a €400 million European Investment Bank loan signed on 10 March 2025, with first production in 2028 and full commissioning in 2030; as of December 2025 the concrete works were over 70% complete.^[78][77] Urenco is adding 700,000 SWU at Eunice, New Mexico by 2027, four of those cascades already online ahead of schedule — and on 2 June 2026 announced a far larger step: 2.1 million SWU across 24 new cascades, a near-50% expansion of a plant whose present 4.3 million SWU covers about a third of US demand, with the initial cascades not producing until 2032 and the remainder installed through 2036.^[75][76]

Set those three dates beside the statutory one. Every waiver under the Russian import ban terminates on 1 January 2028.^[29] Orano's new modules begin producing in 2028 and finish in 2030; Urenco's large American expansion begins producing in 2032 and finishes in 2036. The West's answer to the enrichment problem is real, funded, under construction — and arrives four to eight years after the deadline it was meant to meet. That gap is not a forecast. It is the published schedule of the two companies doing the work.

Enrichment is the binding constraint. A binding constraint that takes a decade to relieve means the price of SWU rises and stays high, which transfers value from reactor operators to enrichers — the opposite of the usual assumption that the reactor vendor captures the revival. It also means advanced reactor developers are competing for the same scarce commodity as each other, which is why fuel supply agreements have become the milestone that separates credible programmes from press releases. And it means that the ADVANCE Act's licensing reforms, useful as they are, address a constraint that was not the binding one. A reactor licence with no fuel is a permit to wait.

21The datacentre surge: a new buyer with unusual requirements

For two decades OECD electricity demand was flat, which meant no new firm capacity was needed at any price and existing nuclear plants competed against gas on marginal cost alone. Several lost. That regime has ended, and the agent of its ending is a buyer that did not exist in the last nuclear cycle.

The deals are specific and dated. Constellation announced on 20 September 2024 a twenty-year power purchase agreement with Microsoft to restart Three Mile Island Unit 1 as the Crane Clean Energy Center, restoring approximately 835 MWe; the unit had shut for economic reasons exactly five years earlier, having run at 96.3% capacity in its final year.^[33] As of June 2026 the NRC had issued a draft environmental assessment finding no significant impact, with a final assessment expected in September 2026.^[38] Talen Energy announced on 11 June 2025 a long-term agreement to supply Amazon Web Services with 1,920 MWe from the Susquehanna plant, ramping to full volume by 2032 at the latest and running to 2042 with extension options — restructured from a behind-the-meterA customer wired straight to the plant, taking power before it reaches the public grid — which is why regulators take an interest: that load pays nothing for transmission and the grid loses a generator it had been counting on. Front-of-the-meter means the power goes onto the grid first and the buyer draws it off in the ordinary way. co-location to a front-of-the-meter arrangement that does not require FERC approval.^[65] Google contracted for up to 500 MWe of Kairos capacity by 2035; Meta signed a twenty-year agreement covering 2,176 MWe at Perry and Davis-Besse plus uprates in January 2026; Walmart signed its first nuclear PPA with Constellation for about 176 MWe at Dresden in June 2026.^[38]

These buyers do not want what a utility wants. They want large indivisible blocks. They want them carbon-free on an hourly rather than annual matching basis, which wind and solar cannot supply without storage that does not yet exist at the required scale. They want twenty-year price certainty, because a datacentre is a twenty-year asset. And they have balance sheets that make them better credit counterparties than most utilities. That combination is close to a perfect match for nuclear's cost structure, which needs exactly a long-dated, high-utilisation, creditworthy offtakeThe buyer's binding commitment to take the output — the contract that turns a plant's future electricity into a revenue stream a lender can underwrite. to make the cost of capital problem in §13 tractable.

Two cautions belong here. First, almost every deal so far involves existing reactors — restarts, uprates and life extensions — not new construction. That is entirely rational, given Figure 20: securing an existing megawatt-hour costs $26–49 and a new one costs $141–220. It is also, on the arithmetic, a redistribution of firm low-carbon power rather than an addition of it. Every megawatt-hour Microsoft buys from Crane is a megawatt-hour that would otherwise have gone to the Pennsylvania grid.

Second, restarts are proving harder than announced. Palisades in Michigan, the first US plant ever approved to return from decommissioning, received its NRC exemption in June 2025 and was formally transitioned from decommissioning status to an operating licence in July 2025. Holtec targeted October 2025, then end of February 2026. As of 8 July 2026 the major projects were complete and new fuel was on site, more than 5,000 individual work activities remained, the CEO expected restart "this year, ahead of its contract to supply power by March 2027", and the company would commit only to restarting "once all restart work is complete."^[34] No firm date. A targeted search on 1 August 2026 found no report of criticality or grid connection. That is a plant that already exists, with a licence, with fuel on site, and it has slipped by more than a year.

22Fusion: real progress, on a schedule that is not yet an energy schedule

Fusion is the other reaction. Rather than splitting a heavy nucleus, it joins light ones: deuterium plus tritium yields helium, a neutron and 17.6 MeV, which on a mass basis is more than four times the energy release of uranium fission.^[59] The fuel is effectively unlimited, there is no chain reaction to run away, the waste is short-lived activated structure rather than long-lived actinides, and the reaction stops the instant conditions fail. Every one of those advantages is real. None of them is available yet.

The physical problem is confinement. To fuse, nuclei must approach closely enough for the strong force to act, which means overcoming electrostatic repulsion, which means temperatures above 100 million degrees — and no material vessel can touch that. Magnetic confinement holds the plasma in a field, most successfully in the doughnut-shaped tokamak. Inertial confinement compresses a fuel capsule so fast that it fuses before it can fly apart.

The inertial line has produced the field's landmark result. On 5 December 2022 the National Ignition Facility at Lawrence Livermore delivered 2.05 megajoules of laser energy to a target and recovered 3.15 megajoules of fusion energy — the first time in a laboratory that fusion output exceeded the energy delivered to the target.^[28] Critically, it has been repeated: as of 20 June 2026 NIF has achieved ignitionIn fusion, the point at which the energy the reaction releases exceeds the energy delivered to the fuel, so the burn begins to heat itself. It is a threshold in the physics. It says nothing about the electricity a plant must draw from the wall to reach it. eleven times, with a record yield of 8.6 MJ from 2.08 MJ delivered on 7 April 2025, a target gain of 4.13 in a 456-terawatt peak-power pulse.^[28]

Target gain is fusion energy out divided by laser energy on the target. It is not fusion energy divided by the electrical energy drawn from the wall to fire the lasers, which is larger by a factor in the tens to low hundreds for NIF's flashlamp-driven design. NIF is also a weapons-physics facility that fires a few shots a day; a power plant would need several per second. The result is a genuine scientific milestone — a burning, self-heating plasma in a laboratory — and it is not a demonstration of net electrical power.

The magnetic line runs through JET and ITER. The Joint European Torus set a record of 69 megajoules of fusion energy sustained over five seconds in its final deuterium-tritium campaign in October 2023.^[59] ITER, under construction in southern France as an international collaboration, is designed to produce 500 MW of fusion power for at least 400 seconds from less than 50 MW of plasma heating — a plasma gain of ten. Its 2024 revised baseline moved the start of substantial research operation to 2034 and full deuterium-tritium operation to 2039, adding about €5 billion to the estimate.^[59]

ITER will begin its scientific mission proper roughly eighty years after Obninsk connected the first reactor to a grid, and ITER is explicitly not a power plant — it produces no electricity by design. A demonstration plant would follow, and a commercial fleet after that.

The private wave is genuinely new and deserves neither dismissal nor credulity. Commonwealth Fusion Systems is building SPARC, a compact tokamak using high-temperature superconducting magnets that permit much stronger fields in a much smaller machine, with operation planned for 2027 and a 400 MWe ARC plant intended to supply the grid in the early 2030s.^[59] Helion pursues magnetised target fusion with direct electrical recovery rather than a steam cycle, and reported reaching 150 million degrees Celsius in February 2026.^[59] Tokamak Energy's spherical ST40 reached 100 million degrees in March 2022 and the company plans an 85 MWe pilot for the mid-2030s; General Fusion reported plasma heating to about 8.4 million degrees in June 2026.^[59] TAE Technologies, which has raised more than $1.2 billion and pursues a field-reversed configuration aimed ultimately at hydrogen–boron fuel, announced on 15 April 2025 that its "Norm" device had formed and sustained an FRC plasma using neutral-beam injection alone — a result published in Nature Communications that removes the machine's plasma-formation section and cuts size, complexity and cost by up to half; TAE says it validates the components for its next prototype, Copernicus, intended to show net energy before the end of the decade.^[89] Zap Energy, pursuing sheared-flow-stabilised Z-pinch without magnets or lasers, exceeded gigapascal plasma pressures on FuZE-3 in November 2025 after reaching 37-million-degree electron temperatures in April 2024, on total funding above $330 million.^[90]

One datum from that group separates what is underway from what is promised. In April 2026 Zap — among the better-funded private fusion companies — announced that it was adding nuclear fission to its business alongside fusion.^[90] That is not a technical failure and should not be read as one. It is a commercial signal, and it says something about the gap between a plasma milestone and a revenue line that no plasma milestone can say: the company closest to its own physics goals concluded that its route to a paying customer ran through the eighty-year-old technology first. Treat every fusion date in this section against that.

Three claims need separating. That fusion is physically possible: settled, and NIF settled the ignition part of it. That a compact high-field tokamak can reach net plasma gain: plausible, untested, and SPARC is the test. That fusion can deliver electricity at a price that competes with fission, gas or solar, having also solved tritium breedingTritium is half of the deuterium–tritium fuel, barely exists in nature and decays away, so a fusion plant has to manufacture its own — by catching its own neutrons in a lithium blanket wrapped round the chamber. No machine has yet run as a closed loop on tritium it made itself., neutron damage to first-wallThe innermost surface of the vessel, the one facing the plasma and absorbing its neutron flux directly. Those neutrons displace atoms in the metal over years, and no material has yet been proved to survive a power plant's worth of it. materials, and duty cycle: entirely unproven, and the engineering problems there are not lesser than the physics ones. Fusion is closer than it was in 2010, in a way that is measurable rather than rhetorical. It is not close enough to appear in any 2040 electricity supply plan, and a primer that suggested otherwise would be doing the reader a disservice.

23What is genuinely contested, what is not, and what would change the picture

The strongest case against nuclear power, stated to persuade

Any primer that ends by finding for its subject has probably not tried hard enough to lose. The case against, then, argued on its best evidence rather than its worst.

Nuclear power is a technology that has had eighty years, unlimited state sponsorship, wartime-scale engineering talent, and a captive regulated market for most of its life, and it has arrived at a delivered cost of $141–220 per megawatt-hour in the West against $37–86 for onshore wind and $38–78 for utility solar.^[22] Its share of world electricity is 8.9% and falling — the lowest since the early 1980s — while solar and wind are expected to overtake it outright in 2026.^[8] Its two most recent Western projects overran by factors of 2.4 and 4.0 respectively.^[21][25] Its fuel cycle is 44% controlled by a state that invaded its neighbour, and no amount of policy has yet changed that.^[6] It has never, anywhere, closed its own back end: seventy years in, not one permanent repository is operating, and the country that collected $51.4 billion to build one has instead paid $9 billion in damages for not building it.^[39] Its flagship modular project was cancelled for want of subscribers.^[63] And the arguments for it change every decade — cheap power, then energy independence, then climate, now datacentres — which is what an answer looking for a question sounds like.

The strongest form of the argument is not about safety at all. It is about opportunity cost. Every billion dollars and every year of political attention spent on nuclear is a billion dollars and a year not spent on transmission, storage, demand response and renewables that are already cheaper and deploy in eighteen months rather than eighteen years. If the objective is tonnes of carbon abated per dollar per decade, nuclear loses, and it loses by a wide margin.

Why that case does not fully hold

It fails on three points, and each of them is quantitative.

First, the cost comparison is between different products. Wind at $37–86 and solar at $38–78 are unfirmed energy; nuclear at $141–220 is firm capacity available in every hour.^[22] Lazard's own solar-plus-storage figure of $50–131 is only partly firm and does not cover multi-day still, dark weather.^[22] The honest comparison is between a firmed renewable system and a nuclear one, and nobody has yet built the former at scale in a large temperate grid, so the comparison rests on models rather than on outcomes. Nuclear's number is expensive and it is observed; the alternative's number is cheaper and it is projected. That asymmetry does not settle the argument. It should discipline how confidently anyone asserts it.

Second, the opportunity-cost argument applies with far greater force to new build than to the existing fleet, and it inverts entirely for life extension. Keeping an existing reactor running costs $26–49 per megawatt-hour, which is cheaper than any new-build anything.^[24] Germany shut 20.4 GW of it and burned more coal.^[41] Japan shut 47.5 GW and its electricity carbon intensity rose from 350 to 487 grams per kilowatt-hour in a single year.^[42] Whatever one concludes about building reactors, closing working ones has been demonstrably, measurably, expensively wrong, and the demonstrations are recent.

Third, the safety premise underneath the political case does not survive the evidence. UNSCEAR found no radiation-attributable health effects at Fukushima, documented or expected; the Kemeny Commission projected 0.7 statistical cancers from Three Mile Island; and the accident that did cause mass harm did so through a reactor design that was never built outside the Soviet Union and a public-health failure that would not be repeated.^[14][20][52] A policy built on a hazard the evidence does not support will misallocate capital regardless of which direction it points.

Five conclusions a specialist may not already hold

The debates that are genuinely open

Is nuclear cheaper or dearer than firmed renewables? Unresolved, and the reason is that the comparison depends on a system nobody has built. At low renewable penetration, wind and solar are unambiguously cheaper. At very high penetration, the cost of the last increment of firmness rises steeply and nobody knows how steeply, because no large temperate grid has yet run at 90%-plus variable renewables through a multi-day winter anticyclone. The answer is empirical and the experiment is running now, in Germany, in California, in South Australia.

Will SMRs actually be cheaper, or merely smaller? Unresolved, and §19 sets out both cases. The evidence to date leans negative — the EIA's engineering estimate puts SMRs at a higher cost per kilowatt than a large AP1000, and the flagship US project was cancelled on cost.^[23][63] The evidence has also not yet had its real test, which is the fourth unit of a design rather than the first.

Once-through or closed cycle? Unresolved, and probably not resolvable on technical grounds. Reprocessing cuts high-level waste volume to a fifth and recovers 25–30% more energy, at the cost of expense and separated plutonium.^[47] Which side wins depends on the price of uranium decades out and on how much proliferation risk a state is willing to import in exchange for resource efficiency. Those are preferences, not facts.

How dangerous is low-dose radiation, really? Genuinely unknown, and this is the most consequential open question in the field. The linear no-threshold model is a regulatory convention, not a measurement; UNSCEAR's own language concedes that its risk estimates concern effects that would not be "detectable compared to the normal statistical variability in the baseline incidence."^[14] If LNT overstates low-dose risk, then decades of remediation spending, evacuation policy and design margin have been misallocated. If it understates it, the opposite. The experiment that would settle it cannot ethically be run.

Is fusion decades away or structurally closer? Both, depending on the claim. Ignition is achieved and repeated eleven times.^[28] Net electrical power from a plant that breeds its own tritium and survives its own neutron flux for years has not been demonstrated at any scale, and ITER's own baseline puts full deuterium-tritium operation in 2039.^[59] The private high-field tokamak route may compress that materially, and SPARC is the test.

The limits of what this primer can claim

Several things here are less certain than the prose may make them sound, and they should be named.

Cost figures across countries are not comparable in the way readers want them to be. Chinese and Russian capital costs are not audited to Western standards, do not carry a market cost of capital, and often exclude items a Western developer must fund. The schedule data is more reliable than the cost data, because grid connection dates are observable and cost accounting is not.

Capacity-factor comparisons between technologies measure utilisation, not value. Nuclear's 91% and solar's 24% are not two grades of the same thing.^[9]

The IAEA's 2024 fleet data excludes Ukrainian operating and outage data for 2022–24, which were not reported by publication.^[3] Any statistic here that involves Ukraine is incomplete for that reason and is flagged where it matters.

Levelised cost is a flawed instrument and it is the best available one. It embeds an assumed capacity factor, an assumed lifetime, an assumed cost of capital and no valuation of when the electricity arrives. Lazard's nuclear figure in particular is not a fresh engineering estimate but an inflation-adjusted v14.0 result anchored to one project, Vogtle — a first-of-a-kind build by a supply chain that had not built a reactor in thirty years, which is close to a worst case rather than a central one.^[22]

And the frontier section describes intentions as well as facts. Every SMR and fusion date in Part IV is a target stated by a party with an interest in stating it, several have already moved once, and the primer's own record of what happened to previous such dates — Flamanville's 2012, Onkalo's mid-2020s, Palisades' October 2025, the CFPP's entire existence — is the best available prior for how to read them.

One closing thought, and it is the one the physics keeps pointing at. Nuclear power is the only energy source humanity has ever developed whose central problem is not getting enough energy out of the fuel but managing the consequences of getting so much. That is an unusual kind of problem to have, and it is not principally a scientific one. The physics was settled in 1942. Everything since has been institutions — regulators, treaties, utilities, insurers, publics, capital markets — working out how to live with a machine whose density they never quite believed. On the evidence of eighty years, the physics has been the easy part.