Nuclear Energy — Jon Moshier

Nuclear supplies about 9% of the world’s electricity and has since roughly 2020. That number is the whole puzzle in one figure: it is low relative to what the technology can do, it has been falling for 30 years, and it is falling while nuclear posts one of the best safety and carbon records of any energy source. This note lays out the usage numbers, the mechanism behind the stall, and the reasons the trend may be reversing right now.

How much we use

Nuclear generated about 9% of global electricity in 2024, down from a peak of 17.5% in 1996, per the World Nuclear Industry Status Report. Absolute output has been roughly flat for two decades and is only now ticking up to a new high, with the world’s ~420 reactors set to produce more power than ever in 2025. The share still fell, because the denominator, total electricity demand, grew far faster than nuclear did.

The global average hides enormous spread:

France: about 65-70% of its electricity from nuclear, the highest of any large country, the result of the “Messmer Plan” crash program after the 1973 oil shock.
European Union: about 23% (IEA).
United States: about 18%, from 94 reactors, still the largest fleet in the world by output.
China: a small share of its own huge grid but the center of gravity for new construction, with more reactors under construction than any other country.

The number of countries running power reactors has been stuck at about 32 since 1996. Phase-outs have matched new entrants one-for-one.

Why it works well on the metrics people say they care about

The stall is not explained by nuclear performing badly on safety or carbon. It performs near the top.

Deaths per unit of energy. Using Our World in Data’s figures in deaths per terawatt-hour, including accidents and air pollution: coal ~24.6, oil ~18.4, natural gas ~2.8, hydropower ~1.3, wind ~0.04, nuclear ~0.03, solar ~0.02. Nuclear’s number already includes Chernobyl and Fukushima. It is roughly as safe as wind and solar and hundreds of times safer than coal, because everyday particulate pollution, not rare accidents, dominates the fossil comparison.

Accident death tolls are far lower than the mental image. Chernobyl killed 31 people directly; the WHO’s contested upper estimate for eventual cancer deaths is around 4,000. Fukushima had one confirmed radiation-attributed death; the roughly 2,000 additional deaths came from the stress and disruption of the evacuation itself, not radiation. This is the core of the “nuclear fear” argument advanced by writers like Michael Shellenberger: the policy response to accidents has sometimes caused more harm than the accidents, for instance when post-Three Mile Island shutdowns shifted generation back to coal.

Carbon. IPCC lifecycle median emissions are about 12 gCO2/kWh for nuclear, roughly the same as onshore wind (~11) and well below utility solar PV (~40+) and coal (~820). On the two metrics most climate advocates foreground, safety and carbon, nuclear scores well. So the explanation for the decline has to lie elsewhere.

Why we’re not using more: the cost mechanism

The dominant explanation is economic, and it has a specific and strange shape. Most technologies get cheaper as you build more of them, a learning curve. Nuclear’s learning curve inverted around 1970: each new plant cost more than the last. Costs rose several-fold in real terms over the following decades, a pattern analysts have called almost unprecedented in industrial history. See Negative Learning Curve.

What drove the inversion is debated, and the debate matters because different causes imply different fixes:

Regulatory ratcheting. After Three Mile Island (1979), safety requirements tightened repeatedly, each change adding cost and, more importantly, schedule. Critics argue much of this bought little marginal safety. Defenders argue the industry’s early cost estimates were never realistic.
Loss of standardization and serial construction. France built a few standardized designs many times and stayed cheap. The US built many bespoke designs and did not. A first-of-a-kind plant is a mega-project; the tenth identical unit is manufacturing. Most Western programs never reached the tenth.
Financing structure. A reactor is an all-or-nothing bet that ties up billions for a decade before producing a single revenue-generating kilowatt-hour. High interest rates compound over that decade. Solar goes from permit to operation in one to three years, which cuts financing risk sharply.

The result shows up in levelized cost. New nuclear runs roughly $90-150+/MWh; utility-scale solar runs roughly $30-50/MWh (Forbes / Lazard). The UK’s Hinkley Point C, first estimated at £16 billion, is now projected around £46 billion and delayed past 2030, a live example of the schedule-and-cost trap. Cheap batteries have begun to erode nuclear’s clean advantage in daily firm output by letting solar-plus-storage cover more hours, though storage still does not solve multi-day or seasonal firmness.

There is a counter-case worth stating fairly. LCOE ignores system costs: renewables need overbuilding, long-distance transmission, and storage to deliver firm power, and some full-system studies find that deep decarbonization is cheaper with nuclear in the mix than without it (ScienceDirect). The economic verdict depends heavily on how much you value firm, land-light, weather-independent generation.

Why we’re not using more: acceptance, phase-outs, and waste

Cost is not the whole story. Much of the decline was a deliberate political choice, driven by public opinion rather than economics, and it operates as an independent cause.

Phase-outs. Germany is the clearest case. After Fukushima it legislated a full exit and shut its last three reactors in April 2023, replacing much of that low-carbon output with coal and gas in the near term. Italy voted nuclear out entirely by referendum in 1987 and again in 2011. These plants were not closed because they were uneconomic to run; they were closed because electorates decided against them. This is the “nuclear fear” mechanism from the safety section acting through the ballot box, not the balance sheet.

Waste. The technical case is that spent fuel is small in volume, containable, and deep-geological disposal is a solved problem in principle: Finland’s Onkalo repository is the first built and is set to begin operation this decade. The political case is that almost no other country has sited a permanent repository. The US spent decades and billions on Yucca Mountain before cancelling it in 2010, and spent fuel still sits in dry casks at reactor sites. Waste is less a physics problem than a siting-and-trust problem.

Proliferation. Enrichment and reprocessing technology overlaps with weapons pathways, which is a real constraint on which countries can build unsupervised and a genuine reason for caution, distinct from reactor safety. It shapes the geopolitics of who gets to expand.

The load-bearing point: even if the cost mechanism above were fully solved, acceptance, waste siting, and proliferation would still cap how fast nuclear could grow. The stall is economic and political.

Why the trend may be reversing

For the first time in a generation, demand is pulling toward nuclear, and the driver is AI. Data centers already draw an estimated 4% of US electricity, a figure widely projected to more than double by 2030. Hyperscalers want power that is carbon-free, firm, and available around the clock, which is precisely nuclear’s profile. See Data Center Externalities and LLM Energy Use.

The 2024-2026 deal flow is concrete:

Three Mile Island Unit 1, shut in 2019, is being restarted under a 20-year Microsoft power deal, targeting 2028. A US commercial reactor has never been restarted from retirement before.
Amazon is investing over $20 billion to build data-center capacity around the Susquehanna plant.
Google signed the first US corporate small modular reactor fleet deal with Kairos Power. See Small Modular Reactors.
Meta issued a request for 1-4 GW of new nuclear.
Tech firms signed 10+ GW of US nuclear capacity in roughly a year.

Two cautions. First, restarts and uprates of existing plants are cheap and fast; genuinely new build still faces the cost mechanism above, unproven for SMRs at commercial scale. Second, routing a retired-then-restarted plant’s output to a single private data center rather than the public grid raises the question of who pays for grid infrastructure and who gets the clean power. See Ratepayer Cost-Shifting. Whether this is a durable revival or a demand-driven blip depends on whether SMRs can finally deliver the serial, standardized construction that made France cheap and America expensive.

Try it

Compare death rates and carbon side by side (1-2 hours, browser + spreadsheet). Pull the per-TWh death-rate and lifecycle-CO2 tables from Our World in Data’s energy pages into a sheet and plot safety against carbon for coal, gas, solar, wind, hydro, and nuclear. What you are looking for: nuclear clustering with wind and solar in the safe-and-clean corner, far from coal and gas. Then look up the public-perception surveys that rank nuclear as feared and sit with the gap between the two charts. That gap is the note’s whole subject.

Model the financing trap (a weekend, Python or a spreadsheet). Build a toy discounted-cash-flow model for two plants: one with a 2-year build and one with a 10-year build, same total overnight cost and same lifetime output. Discount at 4%, then at 8%. Watch the long-build project’s levelized cost blow up as the rate rises while the short-build one barely moves. You will have reproduced, in ten rows, the single biggest reason new nuclear struggles against solar.

Learn more

Books

Stewart Brand, Whole Earth Discipline (2009). The environmentalist case for nuclear, from a founder of the modern green movement. The book that flipped several prominent anti-nuclear writers. Start here for the pro-nuclear-as-climate-tool argument.
Vaclav Smil, Power Density / Numbers Don’t Lie. Smil is the antidote to hype in any direction. Rigorous, quantitative, unsentimental on the physical limits and land-use of every energy source. He is skeptical of easy stories on all sides.
Charles Perrow, Normal Accidents (1984). The classic argument that tightly-coupled, complex systems like reactors will fail in unforeseeable ways no matter how careful the design. Read it against Brand for the strongest safety-skeptic case.
Richard Rhodes, The Making of the Atomic Bomb (1987). Pulitzer-winning history of the physics and the people. Not about power generation, but the definitive account of how the technology came to be.
Gwyneth Cravens, Power to Save the World (2007). A former nuclear skeptic’s guided tour of the fuel cycle, waste, and risk. Accessible narrative nonfiction.

Film and audio

Pandora’s Promise (2013, dir. Robert Stone). Documentary following environmentalists who changed their minds on nuclear. The visual companion to Whole Earth Discipline.
Chernobyl (2019, HBO miniseries). Dramatized but meticulous on the accident and the Soviet response. Watch it aware that its death-toll framing is more dramatic than the epidemiological consensus; pair with the safety numbers above.
World Nuclear Association and IEA reports. For current data rather than narrative: the IEA’s The Path to a New Era for Nuclear Energy and the (skeptical) World Nuclear Industry Status Report. Reading the industry and the critics side by side is the fastest way to calibrate.

Consider adding one of these to To Read, Watch, and Listen if you want to go deeper.