Kai-Fu Lee has restated a case that has become the standard argument from Chinese industry: American restrictions on advanced chips pushed Chinese labs toward efficiency and open release, open weights spread faster than licensed APIs, and cost-sensitive markets outside the US and Europe will default to whatever is free and good enough. He extends it to labour effects and to the adoption pattern in developing countries. The argument is worth taking seriously because much of it is correct, and worth examining because the parts that are correct do not support the conclusion as cleanly as they appear to.
Where the argument holds
Compute restriction did impose real cost, and the response was real adaptation. Training efficiency work, aggressive distillation, and architecture choices that reduce the compute required per unit of capability were pursued in China with an urgency that a firm with unlimited access to accelerators does not feel. Open release then solved a distribution problem that the restrictions had also created, since a lab denied the compute to serve global inference at scale can still ship a file.
The adoption claim is also sound. A ministry, a bank, or a startup in a low- or middle-income market weighing a metered frontier API against downloadable weights it can run on modest hardware is making a straightforward budget decision. Free wins a large share of that market, and the models are good enough for most of what is being asked of them.
Where it strains
Openness here is a distribution strategy rather than a governance property. Weights are released; training data, filtering decisions, and post-training procedure generally are not. That distinction matters for anyone treating a downloaded model as auditable. It also means the advantage being described is a commercial position, not a transparency commitment, and commercial positions get revised when the strategic logic changes.
The compute argument also proves less than it claims. Training constraints bind hardest at the frontier, and every reported efficiency gain still sits on top of accumulated access to hardware acquired before or around the controls. Inference at national scale remains a compute problem that efficiency work reduces without removing. The controls did not fail. They shifted the competition to a layer where the US position is weaker.
The part that is genuinely a policy failure
US policy has no equivalent instrument at the distribution layer, and until this week the main hub for open weights was independent. Washington restricted the export of accelerators while the artefact that actually diffuses capability moved freely, hosted on infrastructure it did not control, into markets it was not competing for. Chinese labs did not need to smuggle anything. They uploaded.
The durable asset in that exchange is not the model. It is the developer base, the fine-tuning ecosystem, the tooling, and the documentation language, all of which create switching costs that outlast any particular release. A ministry that has trained its staff on one family of open models and built its evaluation pipeline around them is not going to re-platform because a better closed model appears at a price it cannot pay.
Assessment
Treat the open-model advantage as real, bounded, and partly self-inflicted by the country it disadvantages. The variables that would change the picture are the pace at which Chinese labs continue releasing at the frontier rather than one tier below it, the appearance of any US restriction on hosting or downloading models of Chinese origin, and whether cost-sensitive adopters begin paying switching costs to move. None of those has resolved.