AI Pipeline Control: China, US, and a New Model

Researchers at MIT and Hugging Face found that Chinese AI models have surpassed American models in global downloads, accounting for roughly 17 percent of the total in 2025, compared to just under 16 percent for the US. This shift isn’t solely about capability, but also distribution.

Chinese models offer full weights, permissive licenses, and significantly lower pricing. The author argues the debate over AI control isn’t about open versus closed, or US versus China, but a competition between three control architectures.

China embeds control in model weights and updates them with each release. The US is attempting control through silicon and jurisdictional enforcement.

A third model proposes embedding control in the ownership structure of AI companies. The author contends the US cannot suppress the Chinese pipeline but can focus on building verifiable provenance, a clear record of a model’s origins, and allowing regulated markets to enforce standards.

This approach would leverage American market strength to incentivize transparency and accountability. The piece outlines a program to establish a standard for AI model provenance, creating a competitive advantage for models that adhere to it. The author predicts that liability insurance and government procurement will increasingly require proof of model provenance by 2027, and that China will maintain its rapid release cadence through 2027 unless hampered by compute limitations.