Oluwadi
Investigating the Overlooked
Region
What
Who For
Analysis
Oluwadi
Investigating the Overlooked
In a single stretch of September 2026, a senior member of Congress pressed for a binding US–China understanding to put guardrails on frontier AI — and the American labs cut the price of that same frontier AI, some reportedly by 40 to 70 percent, barely a week after those labs had publicly called for a slowdown. Rep. Ro Khanna, the ranking Democrat on the House Select Committee on the CCP, used the run-up to the Trump–Xi meeting to float an alternative playbook: a negotiated US–China agreement with real limits on the most dangerous capabilities.[1] Within days, OpenAI and Anthropic pushed out cheaper models to compete with China's low-cost rivals — the same firms that, ten days earlier, had asked the industry to slow down.[2] Read together, the two moves are the whole problem in miniature.
Arms control is a technology of scarcity. It works on things that are hard to make, expensive to hold, and possible to count: enriched uranium, centrifuges, warheads, missile silos. Every serious treaty of the last century — the test-ban regimes, SALT, the Non-Proliferation Treaty — rests on the same three assumptions. The weapon is scarce. The weapon is dear. And the weapon can be inspected, because it occupies physical space that a satellite or an inspector can find.
Frontier AI is moving against all three assumptions at once. It is getting cheaper every quarter — the price war is not an accident, it is the business model. It is getting more abundant, not less, as each generation ships to more users at lower cost. And, most fatally for the arms-control frame, it is going open-weight: the capability increasingly escapes into model files that anyone can download, run, and copy without permission. The centrifuge cascade that took a nation-state years to build has an analog that fits in a torrent.
That is where the nuclear analogy — which the entire vocabulary of "AI arms control" quietly borrows — breaks in half. The barrier to a bomb is materials and physics: hard to acquire, hard to hide, available to a handful of actors a treaty can name. The barrier to a frontier model is falling toward an API key, and past that, toward a file already on the open internet. You can verify a nation is not enriching uranium by walking inspectors through a facility. You cannot walk an inspector through a training run inside a private data center, and you certainly cannot un-release a set of weights that a hundred thousand people have already downloaded.
None of this makes Khanna wrong to try. The dangers he is pointing at — recursive self-improvement, a race with no brakes — are real, and a diplomatic channel between the two countries that matter is worth having on its own terms. But the instrument he is reaching for was built for a different physics. A treaty can freeze a stockpile because a stockpile sits still and can be counted. Software does not sit still, does not deplete, and does not disarm. Ship it once and it is everywhere; cut its price and you have made proliferation cheaper by policy.
So the tell is the timing. When the same labs asking for a pause are, the same month, cutting the price of the thing they want paused — and a congressman is drafting a treaty for it while its weights leak into the open — you are not watching an arms race that can be controlled by agreement. You are watching a technology behave like software, and software, once shipped, has never once been recalled by treaty.