Return to the one idea that clarifies the whole AI safety argument, that a closed model can be opened later but an open one can never be closed. Apply it to China, and an uncomfortable question follows.
If open weights are the single irreversible act in artificial intelligence, the one choice that permanently strips the West of any future brake, then a country flooding the world with frontier-grade open models is not competing.
It may be foreclosing its rival's options, and doing so in a way that cannot be undone.
Pattern hard to ignore
The releases are no longer occasional. DeepSeek detonated the debate last year, Alibaba's Qwen family has passed hundreds of millions of downloads with tens of thousands of derivative models built on top, and Moonshot's latest Kimi system now sits at the capability frontier.
Each arrives openly, cheaply and from outside American and European jurisdiction, which means every US or EU rule, kill switch or pacing agreement covers a steadily shrinking slice of what actually exists.
Washington and Brussels can legislate a shutdown capacity for cloud-served models, and it is powerless over weights already copied onto thousands of hard drives worldwide.
The effect, whether designed or not, is to make the Western instinct to retain a brake almost pointless.
A deliberate act of sabotage?
There is a coherent strategic logic here, and it does not require a conspiracy to work.
China trails the United States in the most advanced chips, thanks to export controls, so commoditising the models themselves neatly attacks the value of the very hardware it cannot buy.
If frontier intelligence is free to download, the economic moat protecting OpenAI, Anthropic and their investors thins, and with it the returns underwriting American AI spending.
Analysts have described open weights as a way to trigger a slow unwinding of Silicon Valley's expensive proprietary bet, and Beijing has openly promoted domestic open platforms as instruments of national self-reliance.
A policy that simultaneously lifts China's global influence, saddles its rival with an unsolvable control problem, and devalues the West's hardware advantage is not obviously an accident.
It is, at the very least, a strategy Beijing would have every reason to choose even if the outcome were merely convenient.
Hyperbole?
The trap reading can also be overstated. Chinese labs open their models partly because it is the fastest way to win global developer mindshare when they cannot match American marketing or cloud distribution.
Openness builds ecosystems, and an engineer in Europe or Africa building on Qwen today is a customer, a standard and a dependency captured for years.
There is also genuine ideological conviction, echoed by Mark Zuckerberg, that the real danger is concentration of power in a few labs rather than diffusion of capability.
By that logic China is not laying a trap so much as backing the side of the argument that happens to suit it, which is what every actor in this debate is doing.
West's belated scramble
What is clear is that the response has arrived late and looks increasingly defensive.
Proposals like a kill switch act, or exemptions for open-source models under new frameworks, are attempts to regain a grip on a frontier that is partly beyond reach.
The awkward truth is that a brake restraining only domestic developers does not slow the technology, it handicaps America's share of it while Chinese weights circulate freely.
Motive may be impossible to prove, and it may not matter.
Deliberate strategy or fortunate by-product, the result is identical, a Western control problem that grows more irreversible with every release, and a rival with no reason at all to stop.
