"If we stop building superintelligence, someone else will." It is the conversation-ender in a hundred rooms. It sounds like geography and game theory. Often it is a way to keep building while calling the build inevitable. Try the same shape with a technology we already handled. "If we stop making CFC fridges, someone else will just keep making fridges."
Yes, they kept making fridges and stopped using the chemical that was opening a hole in the sky.
Two different nouns
A refrigerator is a service: cold food, vaccines, less spoilage. Chlorofluorocarbons were one way to get that service, and a disastrous one for the ozone layer once the chemistry was clear. The Montreal Protocol did not outlaw cold. It scheduled the phaseout of the substances doing the damage, funded alternatives for poorer countries, and backed the deal with trade pressure. Industry had said the phaseout was impossible; refrigeration continued, and CFCs did not.
Superintelligence rhetoric collapses two nouns into one. "AI" becomes a single blob that includes cancer models, route planners, and a hypothetical agent smarter than the human species. If you refuse the last item, accelerationists say, you have refused the first ones too, and someone less scrupulous will ship all of it. The move is to pretend the poison is identical to the product so that removing the poison sounds like a war on the product.
We did not save the ozone layer by using more CFCs than our rivals. We kept the fridge and retired the chemical.
What "someone else will" is for
A rival lab, or a rival state, might try to train through any line you draw. That possibility is real enough to plan for. It is also the possibility arms control was invented to handle. The answer to "someone else might build a nuke" was not "then we must build an even bigger nuke." It was verification, chokepoints on fissile material, and consequences for breakout.
Compute is messier than uranium, and no one serious pretends otherwise. It is still physical, expensive at the frontier, and routed through a short list of chipmakers, fabs, and cloud regions. That is enough structure to build inspections on, which is why compute governance shows up in every serious treaty sketch. "Someone else will" skips that work and goes straight to permission.
If restraint is framed as unilateral surrender, then the only adult posture is full speed. Investors hear destiny. Safety staff hear that slowing down is naive. Lab leaders do not need to win a debate about mechanisms. They only need to make the third option, shared restraint, feel childish.
The fridge was never the enemy
We want narrow AI that diagnoses, designs drugs, and finds faults in bridges. Say that plainly so the smear does not stick. The foundation's line is a refusal to build systems that exceed and replace human control while the control problem is unsolved.
That is closer to "use other refrigerants" than to "ban cold storage." The people who collapse the distinction are borrowing the popularity of useful AI as cover for a capability that does not stay a tool.
What the ozone fight actually teaches
Montreal worked because the science got loud, substitutes existed, and holdouts faced trade costs. Superintelligence is harder on substitutes for "strategic dominance," and easier in one respect AI leaders ignore: no major power has a coherent interest in a system that does not stay under anyone's command. Shared vulnerability is the raw material of a deal. Rival fear can fund a treaty as easily as it funds a sprint, once the hollow finish is visible.
Ask whether they mean useful AI or superintelligence: we keep the tools and stop the system no one can control. For superintelligence, defection risk is a reason to build enforcement, not a reason to pre-defect. The fridges kept running. The sky healed because people could tell the difference.