*Published July 24, 2026* --- In April I saved a tweet from Danielle Fong: society never found out how much abundance nuclear energy could deliver, because the nuclear club "proposed to scale to mutually assured destruction, and close this off forever." I filed it without questioning it. It matched a story I already believed: a technology promised to everyone, then captured by a security race and locked away. Swap enriched uranium for compute and it reads like a forecast for AI. ## The script everyone knows The club-conspiracy version is easy to tell, and I'm not the only one telling it. Frontier models are already under export control, and Andrew Curran predicts training compute will be "treated like uranium: monitored, licensed, tracked, guarded." A two-power race with no off-ramp, abundance repurposed as a strategic asset, the club closing this off the way it closed off the atom. ## What the record actually says Go looking for the moment the nuclear club foreclosed civilian nuclear power and you find the opposite. Eisenhower's 1953 "Atoms for Peace" address launched a transfer program: reactors, fuel, and training exported worldwide, plus the IAEA to spread the peaceful side. Several recipients used the transfers to advance weapons programs; the nonproliferation regime reacted to too much openness, not too little. The club's own charter says so. Article IV of the Non-Proliferation Treaty guarantees "the inalienable right" of all parties to develop nuclear energy "for peaceful purposes without discrimination." The treaty restricts weapons; civilian abundance is a guaranteed right, contested only at the enrichment edge in cases like Iran, never the blanket foreclosure I'd filed away. Civilian nuclear didn't die among the excluded. It died inside the club. American utilities stopped ordering reactors in 1978 and stayed out for three decades. What stopped them was cost, not any decree, and the cost wasn't physics either. Across 349 reactors in seven countries, American costs climbed while South Korea's *fell* for twenty-five years straight, on the same technology. Korea standardized its designs, kept one regulator, and carried each build's lessons into the next. Two acronyms did the ratcheting. Linear no-threshold dosimetry treats every dose of radiation as harmful in proportion to its size, which turns any release into a catastrophe. ALARA ("as low as reasonably achievable") pegs required safety spending to whatever the industry can afford, so every saving just clears room for another rule. That, Jack Devanney argues, is the whole regime: regulate right up to what the industry can bear. Fear did the rest. Three Mile Island killed no one and froze the industry for a generation. The killing was cultural: bomb-born dread, secrecy, a safety ratchet tuned to fear instead of measured harm. The club policed the weapons layer and let the civilian dream die at home. That culture works like fog. It settles everywhere, including on the countries that never got a vote. ## The rerun So is the AGI race rerunning the nuclear script? The club machinery is assembling: export controls on model weights, access revocable by one company or one government. But the decree layer is leaky in a way enrichment never was. Open-weight models trail the frontier by months, not the decades that separate a country from an enrichment cascade; Epoch AI puts the average lag near three months. The direction is contested, with Stanford's 2026 AI Index showing the top-end gap widening in reasoning and agentic tasks. But no lab's moat is measured in half-centuries. If AI abundance dies, it probably won't die by decree. Which is why the nuclear precedent should worry the people cheering the race. The part of the script that actually killed abundance requires no club: panic after a visible accident, harm models that score any capability as proportional risk, compliance costs indexed to what incumbents can afford, a public taught to dread the technology before it touched them. All it takes is one Three Mile Island and a regulatory reflex, and AI starts growing its own LNT. Its dread is already arriving in advance of any real accident, inherited from fiction and the race rhetoric itself, and I can't tell yet whether that is prudence or preemptive panic. The strongest objection is the one I can't dismiss. Nuclear's fear was miscalibrated: radiation harm is measurable, and the ratchet spent billions against risks near zero. AI's worst case (Bostrom's, roughly) is that a capable enough misaligned system is a failure you can't correct after the fact, which would make the ratchet not a tax on abundance but the price of surviving it. That disanalogy is real. But the lesson still cuts: a ratchet indexed to fear rather than measured harm ate the abundance and delivered little safety in return. The choice is how the ratchet gets set, anchored to demonstrated harm or left to compound on dread. Eisenhower proposed something real: share the civilian technology, cage the weapons side in an institution. The AI equivalent would separate the abundance layer from the weapons layer and defend the first while regulating the second. None of the labs setting the pace is building it, and the culture forms on its own. The tell will be the day incumbents brag about how much compliance costs them. That spending isn't buying safety anymore. It's buying a wall around the market. The tweet I saved in April named the wrong killer. The decrees are easy to watch, so everyone watches them. I'm watching the fog. ## Sources - Eisenhower, "Atoms for Peace," UN General Assembly, 1953-12-08 — https://www.iaea.org/about/history/atoms-for-peace-speech - Treaty on the Non-Proliferation of Nuclear Weapons, Article IV — https://www.un.org/disarmament/wmd/nuclear/npt/text - Lovering, Yip & Nordhaus, "Historical construction costs of global nuclear power reactors," *Energy Policy* 91 (2016) — https://www.sciencedirect.com/science/article/pii/S0301421516300106 - Eash-Gates et al., "Sources of Cost Overrun in Nuclear Power Plant Construction," *Joule* (2020) — https://www.sciencedirect.com/science/article/pii/S254243512030458X - Jack Devanney, *Why Nuclear Power Has Been a Flop* — https://gordianknotbook.com/ - Epoch AI, "Open-weight models lag state-of-the-art by around 3 months on average" — https://epoch.ai/data-insights/open-weights-vs-closed-weights-models - Stanford HAI, 2026 AI Index, Technical Performance — https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance - Nick Bostrom, *Superintelligence: Paths, Dangers, Strategies* (Oxford University Press, 2014), esp. Ch. 8, "The treacherous turn" — https://global.oup.com/academic/product/superintelligence-9780199678112 - Andrew Curran (@AndrewCurran_), X, 2026-06-15 (the "treated like uranium" prediction) — https://x.com/andrewcurran_/status/2066332670817456584 - Danielle Fong (@daniellefong), X, 2026-04-20 (the seed tweet) — https://x.com/daniellefong/status/2045908526602875217