The Nuclear Fuel Cycle

Part VI of The Nuclear Fuel Cycle: A Complete Primer

Chapters 22–25 turn from the physical and industrial system to the decisions that shaped it. They explain how the country that led commercial nuclear power for its first two decades ended up with an atrophied industrial base, no permanent repository and a dependence on foreign enrichment, and what is now changing. The chapters move from history to regulation to arithmetic to demand: from the 1977 reprocessing deferral through three accidents, the enrichment exit, Yucca Mountain and Vogtle, to the new licensing and financing architecture, the scale argument for dispatchable non-fossil generation, and AI demand placed inside it. The anchors that carry forward are Vogtle’s cost baseline, the 2028 Russian import ban, Part 53, and the book’s verdict that AI is an accelerant, not the cause.

Chapter 22 — The US Fell Behind

Anatomy of a retreat. By the mid-1970s the United States operated the largest commercial nuclear fleet in the world. Westinghouse and General Electric PWR and BWR designs had been licensed, built and exported across Western Europe and Asia, and the fuel cycle behind them was a US-dominated industry. By 2010 that position had eroded. New construction had been essentially halted for three decades, enrichment had atrophied, and the supply chains for nuclear-grade components had contracted. The chapter’s premise is that the causes “are not mysterious”: specific policy decisions, regulatory responses to specific accidents, and the economics that followed.

The Carter deferral. In April 1977 President Carter indefinitely deferred commercial reprocessing on nonproliferation grounds. It was an executive decision, and there was no statutory ban. The Nuclear Non-Proliferation Act of 1978, which set conditions for exports and bilateral 123 Agreements, did not prohibit domestic reprocessing, and Reagan rescinded the deferral in 1981. By then the effect was set: the Barnwell, South Carolina plant, effectively ready to operate, never processed spent fuel commercially, and no comparable facility was built. France, the UK and Japan kept reprocessing regardless. What is undisputed is the domestic result: a structurally once-through US cycle.

Three Mile Island. On March 28, 1979, Unit 2 underwent a partial core meltdown, the most serious accident in US commercial nuclear history. A stuck relief valve, misleading instruments and reduced emergency coolant flow left about 45% of the core melted. Containment worked. The NRC estimated the maximum individual dose at less than 100 millirem, roughly a few weeks of background, and no evacuation was required. No new US commercial plant received a construction permit afterward. A boxed note widens the picture: the overnight cost of a plant had already risen several times over between the early and late 1970s, driven by in-construction design changes, regulatory backfits and inflation. TMI added uncertainty to an equation that was already stressed.

Chernobyl. The April 26, 1986 accident was qualitatively different. The RBMK-1000’s positive void coefficient, in which boiling raised reactivity, has no analog in Western light-water designs. Thirty-one workers died of acute radiation syndrome, and UNSCEAR attributes about fifteen confirmed thyroid cancer deaths, with up to sixty more possible. No US reactor shared the design, but “Nuclear accidents do not reach the public as engineering case studies.” The international response brought the 1994 Convention on Nuclear Safety, an expanded IAEA OSART program and the World Association of Nuclear Operators (WANO), founded in 1989. These were genuine improvements that also added compliance burden.

The enrichment exit. The government gaseous diffusion plants at Paducah, Portsmouth and Oak Ridge’s K-25 were effective but energy-hungry; Paducah at its peak used more electricity than some mid-sized American cities. USEC, created in 1992 and privatized in 1998, inherited plants made obsolete by European centrifuges. Its American Centrifuge Project at Piketon consumed more than $2 billion over more than fifteen years before cancellation. Portsmouth stopped enriching in 2001 and Paducah in 2013, leaving domestic capacity essentially zero. Utilities bought from Urenco or Russia’s Tenex, and by 2023 Russian enrichment supplied about 24% of US requirements. A boxed note covers Megatons to Megawatts (1993–2013): about 500 metric tons of HEU from roughly 20,000 warheads became about 15,000 metric tons of LEU, equal to about 350 million pounds of uranium oxide.

Fukushima. On March 11, 2011, a magnitude 9.0 earthquake and tsunami disabled the emergency diesel generators for four units, and three reactors suffered partial meltdowns. No member of the public died from radiation. About 2,200 people died from evacuation-related causes among roughly 154,000 displaced residents. Germany accelerated its phase-out and closed its last plants by April 2023, and Japan suspended most of its fleet. The NRC’s post-Fukushima requirements added several billion dollars of compliance costs, which, with cheap gas in the early 2010s, hastened the closure of marginal plants.

Yucca Mountain. The Nuclear Waste Policy Act of 1982 created a repository program, and the 1987 amendments narrowed it to Yucca Mountain. After billions of dollars and decades of characterization found the site technically suitable, DOE applied to the NRC in 2008. In 2010 the Obama administration directed withdrawal. The Atomic Safety and Licensing Board ruled the withdrawal improper, but DOE’s budget request fell to essentially zero. Spent fuel sits at seventy-five or more sites, and Nuclear Waste Fund fee collection was suspended in 2014.

Lead time and Vogtle. A plant takes ten to twenty years from site selection to grid connection under current US conditions, and neither components nor workforce can be reconstituted in response to a demand signal that arrives this year. The 2012 combined construction and operating license (COL) approvals for Vogtle Units 3 and 4 were the first new construction authorization since 1978. Planned for 2016 and 2017 at about $14 billion, the units entered service in 2023 and 2024 at about $35 billion for about 2.2 GW, among the highest per-unit capital costs in US nuclear history. The book treats Vogtle as “a documented baseline” for building after a three-decade gap. Its closing diagnosis: each step came from specific decisions by specific actors, and today’s configuration is their cumulative result.

Chapter 23 — The New Safety Regime

A regime built for one technology. US reactor licensing sits in Part 50 and Part 52 of 10 CFR, both written around light-water features: water cooling and moderation, specific containment configurations, specific emergency core cooling architectures. A design cooled by molten salt or helium, or relying on passive safety physics, does not fit those categories. Writing technology-neutral rules that specify safety outcomes means building evaluation methods for reactor types the NRC has never licensed. The pace of that work has drawn criticism from industry and from Congressional oversight.

Part 53. 10 CFR Part 53 is performance-based. Applicants show that a design meets specified safety outcomes, on whatever technical basis they can support. It is intended to apply to light-water and non-light-water designs, SMRs and large plants alike. The rule was finalized on March 25–26, 2026, published in the Federal Register on March 30 and effective April 29, 2026. It completes work under way since about 2019, when the NRC committed to the approach following NEIMA (January 2019). The first applications will set the precedents, which is why the chapter treats 2025–2026 as consequential.

Microreactors and Part 57. Microreactors, below about 20 megawatts thermal, are small enough for factory fabrication and transport. They target remote power, industrial heat, military installations and off-grid communities. The proposed 10 CFR Part 57 would scale licensing to the risk profile of a remote unit with passive safety and sealed, non-refuellable fuel cassettes. As of mid-2026 it remained a proposal with no established timeline to finalization.

Milestones of 2025–2026. In May 2026 the NRC approved the PDC (Principal Design Criteria) for Oklo’s Aurora, a sodium-cooled fast reactor on metallic high-assay low-enriched uranium producing about 15 MWe. It proceeded under the existing process because it was submitted before Part 53, and PDC approval is neither a construction nor an operating license. Also in May 2026, DOE approved a Documented Safety Analysis (DSA) for Aalo Atomics’ Aalo-X, a sodium-cooled, thermal-spectrum test reactor on UZrH fuel drawing on INL’s MARVEL heritage. The DOME microreactor test facility at INL entered fueled testing in April 2026. A boxed note makes the general point: the passive safety of molten salt and TRISO-fueled gas-cooled designs flows from their physics, and Part 53 creates room for that argument where Part 50/52 did not.

The ADVANCE Act. Signed July 9, 2024, the ADVANCE Act is, in the book’s words, the most significant US nuclear legislation “in decades.” It directed the NRC to cut fees for certain advanced reactor applications, set new review timelines, streamline bilateral cooperation agreements, create NEPA categorical exclusions for low-impact activities, and develop a licensing pathway for former fossil-fuel and other brownfield sites. The Prohibiting Russian Uranium Imports Act, technically a distinct statute in the same package, bans Russian imports no later than 2028, with waivers. Urenco’s New Mexico plant runs at commercial scale. Centrus’s Piketon cascade produces HALEU at demonstration scale. Neither can quickly replace Russian supply, which the chapter calls the Chapter 6 enrichment problem “now with a statutory deadline attached to it.”

Strategic capital. The Office of Strategic Capital (OSC) was established in the Department of Defense by the FY2024 National Defense Authorization Act. The One Big Beautiful Bill Act (2025) appropriated $1.5 billion in credit subsidy for its FY2026 lending. The chapter explains the credit subsidy appropriation: the government appropriates expected default losses, not face value. At a 1–2% subsidy rate, $1.5 billion supports a theoretical $75–150 billion of lending, subject to authorization, borrower creditworthiness and deal structure. The prior model, private capital with limited Title XVII loan guarantees under the Energy Policy Act of 2005, had not financed Vogtle without overruns and the near-insolvency of its contractor. The OSC treats nuclear supply chains as a national security asset.

What has and has not changed. A boxed note sets the limits. The import ban creates a Western supply requirement that current Western enrichment cannot meet on that timeline. Part 53 does not accelerate reactor development, qualification testing or construction. OSC financing does not resolve the workforce and supply chain gaps behind Vogtle’s overruns. A reactor licensed in 2026 still needs years of construction, a trained workforce and a reconstituted supply chain. The chapter’s conclusion is that the regulatory changes are “necessary conditions” for expansion, and that they are not sufficient ones.

Chapter 24 — The Scale Argument

A constrained claim. The chapter starts from documented demand trends and the operating characteristics of available technologies. From those it identifies a gap between what variable renewable generation can reliably deliver and what a modern grid requires. It does not argue that renewables are unimportant or for a single-technology system. Its narrower claim is that projected demand growth requires dispatchable baseload generation at a scale the non-fossil alternatives cannot supply collectively without nuclear.

Demand trajectory. The IEA’s World Energy Outlook runs Stated Policies, Announced Pledges and Net Zero scenarios. Across central and high-electrification scenarios, global demand roughly doubles between 2023 and 2050. Stated Policies alone goes from about 28,000 TWh to about 52,000 TWh, growth of more than 80%. Net Zero shows similar or higher totals, because electrification is itself a decarbonization strategy. US demand, flat for roughly two decades before 2022, has reversed with data centers, manufacturing onshoring and early transport electrification, and several major grid operators have revised near-term projections upward.

Variability. Utility-scale solar costs fell about 90% and onshore wind about 70% between 2010 and 2023 (IRENA and BloombergNEF). Neither can be dispatched on command. The chapter separates curtailment, when output exceeds what the grid can use, from intermittency, extended shortfalls against demand, and notes that they reinforce each other. Winter evenings with peak heating demand have no solar and, in many regions, below-average wind. Daily variation is manageable with gas peakers, interconnection and hourly batteries. Multi-day and multi-week low-generation periods are the harder problem.

The storage gap. The UK’s grid-scale batteries totaled about 4.5 GWh in 2024 (National Grid ESO), against daily consumption of about 950–1,000 GWh: less than 1% of one day. Bridging a multi-day low-wind period, which the UK sees several times each winter, would take on the order of 15,000–20,000 GWh or more, roughly 3,000–4,000 times installed capacity. The requirement compounds, since more renewables need a larger buffer: “The solution, pursued without adequate dispatchable backup, generates its own scaling problem.” A boxed note concedes that future storage costs will fall, but holds that lithium per megawatt-hour is set by cell electrochemistry and does not decline with scale.

Materials. Lithium-ion storage needs about 150–200 kg of lithium carbonate equivalent (LCE) per MWh, depending on chemistry. At high-renewable scale for a large economy that means millions of metric tons of LCE, against global production of about 240,000 t a year and known reserves of about 28 million t (USGS, 2024). Recycling recovers a small fraction of lithium today and cannot meet that throughput within any near-term horizon. The book classes these as material throughput constraints, separate from cost.

Dispatchable options. Hydro is fully dispatchable but geographically constrained: about 1,400 GW in 2023, with its feasible large sites substantially developed and remaining potential concentrated in Africa, South America and South and Southeast Asia. Geothermal stood at about 15 GW of nameplate capacity in 2023 (IRENA), concentrated in Iceland, the US, Kenya, the Philippines and New Zealand. Enhanced geothermal is undemonstrated at grid scale, and its significant contribution falls outside the 2030–2040 window. Nuclear runs at rated output for 18–24 months between refuelings regardless of weather, is not tied to geological features, scales by multiple units at multiple sites, and emits no greenhouse gases in operation.

The arithmetic. Assembled, the picture leaves nuclear as the only non-fossil, non-weather-dependent technology that can in principle be deployed at the scale a doubled system needs. The conclusion does not require nuclear to be cheap, fast or politically uncontroversial. The book names what would change it: storage resolving its material and capital limits at grid scale, or enhanced geothermal proving itself in time. Without non-fossil dispatchable capacity, a high-renewable grid in a long lull sheds load or draws on remaining backup, in most systems natural gas, which preserves fossil dependence. The Energiewende box gives the illustration: Germany closed its last three plants in April 2023, and in the winters of 2021–22 and 2022–23 kept its grid stable with gas, coal and imports. The chapter ends by noting that the argument does not depend on AI demand.

Chapter 25 — The Demand Signal

Cause or accelerant. A technology company powering AI with nuclear, or a plant shut for economic reasons being restarted for data center load, is a significant development. The chapter asks what it represents. If AI is the cause of nuclear’s renewed interest, nuclear’s prospects hinge on AI’s growth. If AI is an accelerant applied to a supply gap that was already there, the case stands regardless of AI’s trajectory. The chapter argues the second interpretation on documented grounds.

A trajectory that predates AI. The IEA had projected growing demand and a dispatchable gap for many years. Flat US demand before 2022 had made nuclear economically vulnerable, and several plants closed between 2013 and 2020 for economic reasons, unable to recover fixed costs against cheap gas. After 2022 the forecasts turned. PJM Interconnection, the world’s largest competitive electricity market, raised its five-year growth forecast in 2024 from about 1.4% a year to about 2–3%, mainly for data centers and industrial load. MISO also revised upward in 2024. These revisions preceded some of the largest AI announcements.

A baseload load. A large AI data center runs twenty-four hours a day, cannot tolerate interruptions and needs high power quality. A hyperscaler such as Microsoft, Google, Amazon or Meta builds campuses drawing 50–100 MW continuously, with some announced projects at 200–500 MW. That profile suits nuclear: the load is stable and needs reliability and continuity. Hyperscalers can also sign PPAs of ten to thirty years, which give a predominantly fixed-capital plant the revenue certainty the spot market does not.

The Crane template. The most documented example is Constellation’s September 2024 twenty-year PPA with Microsoft to restart Three Mile Island Unit 1. The unit had run safely for decades and closed in 2019 because PJM capacity prices could not cover its operating costs. Restart costs are about $1.6 billion, and a $1 billion DOE loan closed in November 2025. The reported price is about $100/MWh, above spot. Renamed the Crane Clean Energy Center, the plant targets commercial operation in 2027. Google, Amazon and others have followed. Some agreements involve advanced reactors not yet built, conditional commitments that still show creditworthy buyers will pay prices that support development financing.

Community acceptance. Data Center Watch, a project of 10a Labs, documented more than $18 billion in data center projects blocked and more than $46 billion delayed by local opposition across twenty-eight states through early 2025. Bisconti Research’s biennial survey finds favorable impressions among roughly 85–90% or more of nuclear plant neighbors. A large plant employs several hundred to over a thousand workers, substantially local and well paid. Pairing a data center with a plant may give it an acceptance profile it could not achieve alone.

Behind the meter and siting. Behind-the-meter generation feeds the consuming facility directly and exports any surplus. Texas House Bill 14 (June 2025) created a $350 million Texas Nuclear Development Fund. X-energy’s project at Dow’s Seadrift plant is designed for four 80 MW units supplying process heat and power, with surplus to ERCOT. In December 2023 the NRC finalized scalable, performance-based emergency planning zone (EPZ) requirements for SMRs and advanced reactors. In place of the fixed 10-mile radius, designs with smaller source terms can size the EPZ by consequence analysis, potentially to a site-boundary EPZ that opens industrial parks and former fossil sites. The ADVANCE Act’s brownfield pathway complements the rule. The UK’s National Policy Statement EN-7, in force December 2025, extends siting beyond the eight pre-approved large-reactor sites, and Wylfa is the expected first Rolls-Royce SMR site, with three initial units. A boxed note records the argument that a co-located SMR could power hospitals, water treatment and homes in emergencies, making the data center “a local resilience asset.”

What AI has done. The structural case was documented before the AI financing wave. AI has compressed the timeline, moving the baseload gap projected for the 2030s and 2040s into grid operator forecasts for 2025–2030. It has created a creditworthy buyer class with twenty-year planning horizons. It has raised the visibility of nuclear’s operating characteristics, which data center operators require but “did not create.” None of this changes the physics or the supply chain constraints. It changes the commercial and political environment, and whether that yields construction is the question Part VII takes up.

Supporting appendices

• Appendix IX — Regulatory and Policy Timeline. It runs from the Carter deferral (April 7, 1977) through TMI (with the Kemeny Commission’s finding that the fundamental problems were institutional and organizational), the Reagan reversal (October 1981) and the 1987 Yucca Mountain designation. It adds the Energy Policy Acts of 1992 (combined license process) and 2005 (loan guarantees), the NRC Fukushima Task Force’s 12 recommendations, Vogtle’s operation dates (July 31, 2023 and April 29, 2024; about 2,234 MWe) and IRA Section 45U. It closes with the ADVANCE Act (July 9, 2024), the Prohibiting Russian Uranium Imports Act (May 13, 2024; full ban January 1, 2028), DOE’s $2.7 billion in enrichment task orders (January 2026), the Part 53 dates and the Crane restart.
• Appendix X — Industry and Landscape. “Data Centers and Firm Clean Power” names both sides of the Chapter 25 pairing. “Regulation, Licensing, Insurance, and Security” covers the ADVANCE Act and Part 53, and “Specialized Construction and Skilled Labor” ties N-stamp certification and forging capacity to Vogtle’s overruns.
• Appendix XII — Physics and Mathematics. “The Negative Temperature Coefficient” and its RBMK contrast explain Chernobyl’s positive void coefficient. “Decay Heat — The Way-Wigner Approximation” gives the cooling problem behind TMI and Fukushima: for a 3,000 MWth reactor, about 38.5 MW at one hour, 20.4 MW at one day and 16.4 MW at the end of the AP1000’s 72-hour passive window.
• Appendix XIII — Optional Problem Set. Problem 2-2, Decay Heat and the Cooling Requirement, has the reader compute those figures. Problem 2-0, The Iodine-Xenon Chain, connects xenon poisoning to the Chernobyl sequence.
• Appendix VIII — Glossary of Terms gives working definitions, with the chapter of first use, for every term marked here.

Longer technical notes and Top Stories tagged to Part VI will appear on this page as they are published.