China's AI Strategy Isn't Copying America. It's Playing a Different Game

China is the only country competing with the United States across the full artificial intelligence stack, from chips to models to applications. Export controls on advanced semiconductors were designed to prevent exactly that, and the Chinese response has reshaped the global picture in ways the restrictions did not anticipate.

The open-weight pivot:

The most consequential development is China's embrace of open-weight models. Chinese laboratories have released capable large language models with downloadable weights, and these have been widely adopted internationally, including by developers in countries with no particular connection to China.

The strategic logic is clear. A country that cannot guarantee access to the most advanced chips can still shape the global ecosystem by making the software layer free. If developers worldwide build on Chinese open-weight models, Chinese technical standards, tooling and design assumptions propagate regardless of who controls the hardware.

This has already changed pricing expectations globally. When capable models can be downloaded and self-hosted at no licence cost, closed commercial providers face pressure that has little to do with their own competitive dynamics.

The chip problem:

Export controls restrict China's access to the most advanced accelerators and to the lithography equipment needed to manufacture them domestically. The restrictions have real effect: training the largest models requires enormous quantities of high-end chips, and China's access is constrained.

The response has been substantial domestic investment. Chinese firms design AI accelerators, domestic foundries have advanced further than many expected, and memory manufacturers have expanded capacity. The remaining bottleneck is manufacturing equipment, particularly advanced lithography, where no domestic alternative currently matches the restricted imports.

Chinese engineering has partially compensated through efficiency: better training techniques, more effective use of available hardware, and architectures that reduce compute requirements. Necessity has driven genuine innovation in efficiency, which is one reason Chinese models compete despite hardware disadvantages.

Regulation and control:

China regulates AI more actively than most countries, though with different priorities than Western frameworks. Rules govern recommendation algorithms, synthetic media and generative AI services, requiring algorithm registration with authorities, labelling of AI-generated content, security assessments before public release, and alignment of outputs with content requirements.

Compliance is a licensing prerequisite rather than a post-market obligation. The system is designed around control of information as much as safety of systems.

Industrial deployment:

China's largest AI advantage may be application rather than research. Manufacturing, logistics, retail, transport and public administration deploy machine learning at enormous scale, supported by dense sensor infrastructure and mobile payment systems that generate extensive behavioural data.

Electric vehicles and autonomous driving are a particular focus, with Chinese manufacturers deploying driver assistance broadly and testing autonomous services in multiple cities. Surveillance applications are extensive and internationally contentious.

Constraints:

Beyond chips, China faces a slowing economy, cautious enterprise spending and constraints on international collaboration as research links with Western institutions have weakened.

The regulatory environment cuts both ways: it provides clarity and it limits what consumer products can do, which shapes what companies build.

The outlook:

The assumption that export controls would halt Chinese AI progress has not held. They have slowed access to the largest training runs while accelerating domestic chip investment, efficiency research and a strategy of ecosystem influence through open weights.

The competition is not a straight race on the same track. It is two systems optimising for different constraints, and predictions based on measuring one by the other's metrics have repeatedly proven wrong.