A nuclear weapon is a design wrapped around a materials problem. You need fissile material, usually highly enriched uranium or plutonium from a reactor route, in quantities and purities that do not fall out of the sky. Enrichment plants are large, power-hungry, and full of kit that intelligence services have spent seventy years learning to spot. That hardness is not the whole non-proliferation story, but it is the part the physics contributed for free.
Now run the thought experiment the slogan skips. Suppose the exact configuration of atoms that makes a bomb were easy. Suppose the inputs were cheap, dual-use, and sold by the pallet. Suppose a finished design could be copied like a file. How many years does the species last?
Not many. Treaties would still matter. So would norms and fear. They would be asked to do all of the work that materials difficulty currently does, at a scale no enforcement system has ever managed. The twentieth century's partial success with nuclear weapons is unintelligible without the friction in the supply chain. We got lucky that uranium is picky.
What hardness bought us
Nine states are known to have nuclear weapons. That number is a scandal and also a kind of mercy. Hundreds of states have the scientific talent to understand the physics. Far fewer have completed the industrial marathon under sanctions, scrutiny, and cost. The Non-Proliferation Treaty, export control regimes, and the IAEA safeguards system all lean on a simple fact: the bottleneck is visible enough to watch.
Remove the bottleneck and you do not get a slightly harder diplomatic problem. You get a world in which the marginal actor, a thin state, a well-funded group, a splinter command, can hold cities hostage. Deterrence math that barely works among a handful of sober powers does not work among dozens of players with shorter time horizons. The bomb is terrifying. Easy bombs are a different species of problem.
We do not survive easy nuclear materials with better speeches. We survive hard nuclear materials with inspection on top. Superintelligence is drifting toward easy.
Which way is AI moving?
Frontier training still takes serious money, serious chips, and serious engineering. A cluster large enough to push the frontier is a physical object in a building with a power bill.
The trend line is not uranium's trend line. Models get more capable per dollar. Techniques improve. Weights, once trained, can be copied at near-zero cost. Open releases and leaks spread capability sideways. Talent is global. None of that makes superintelligence free tomorrow morning. All of it points away from the accidental gift that made nuclear non-proliferation feasible: a stubborn materials gate.
Compare the two stacks honestly. To get a second nuclear arsenal you still need plants, feedstock, and time under telescopes. To get a second copy of a trained model you need a drive. If the dangerous object is the agent rather than the warhead, replication favors the software pattern, not the isotope pattern. We take the metaphor's limits seriously elsewhere.
The wrong lesson and the right one
The wrong lesson is fatalism: the inputs are getting easier, so governance is hopeless. It is the mirror image of the lab slogan that only more capability can save us. Both skip the interval we are still in, where frontier runs are visible and the number of serious actors is small.
The right lesson is urgency about hardness we have to invent. If physics will not keep the gate narrow, law and industry standards must. Know-your-customer rules on advanced chips. Reporting thresholds on training runs. Audit rights. Hardware features that make large covert training harder to hide. A treaty that says superintelligence is not an allowed product. Superintelligence cannot be controlled. The NPT and the model treaty sketch are attempts to put that hardness on paper before the easy path finishes opening.
Close the thought experiment
Imagine briefing a president in 1960 with a proven design and a world in which bomb material sold like fertilizer. Every instinct in the room would scream for control at the source. Nobody would call that instinct anti-science. Nobody would say the answer was to invent a smarter bomb to police the first one.
We are building toward a capability whose natural materials gate is weaker than the bomb's, while telling ourselves that speed is safety. The thought experiment is not subtle. If the dangerous thing is easy to make, the world does not get a mulligan. Make it hard on purpose, while hard is still available, or find out what easy feels like without a fossil record of survivors to argue from.