The story is not that China produces open-weight models. The story is that, when you look at actual usage, those models have become the default infrastructure for anyone working locally. The data Nathan Lambert presented in his written Congressional testimony make this clear: since August 2025, Chinese open-weight models on Hugging Face have recorded about 3.2 billion downloads, double the 1.6 billion American ones. Over the same period, weekly traffic for open models on OpenRouter went from roughly 1 trillion to roughly 80 trillion tokens, and more than 80 percent now comes from Chinese models.

For anyone evaluating on-premise inference, the most relevant figure is not the absolute volume: it is Lambert's estimate of the distance from the closed frontier. For Chinese models he cites a 2-5 month lag, against 6-9 months for American ones. The smallest gap is in agentic coding, where Qwen, GLM and Kimi have become the default local choices for tool-use. That is not a technical footnote: it is a signal that competition has moved from weight releases to concrete workloads.

Why does this happen? An open-weight model does not live in isolation. People using it for tool-use tend to build around it fine-tuning pipelines, quantization profiles, safety tests and serving framework integrations. When the majority of open inference tokens flow through one model family, that is where optimizations, fixes and field-tested use cases concentrate. It is a self-reinforcing advantage: more traffic, more feedback, more shared experience, more mature tooling. American models, even when they catch up on benchmarks, start with a less battle-tested ecosystem for local deployment.

This changes incentives for infrastructure builders. Inference engine and serving library vendors tend to prioritize architectures with the largest user base, because that is where they find feedback and real demand. If open model traffic remains dominated by Chinese architectures, it is reasonable to expect optimization at the kernel, memory and hardware support level to converge toward those architectures. This is not ideology, but technical gravity: optimization follows the token.

For European organizations and those evaluating on-premise deployment, the question is not just which model to download. It is which ecosystem comes with it: evaluation tools, quantization recipes, runtime support, documentation and real-world use cases. Data sovereignty can be guaranteed by a self-hosted instance, but technical supply chain sovereignty is a different game. A Chinese open-weight model can run entirely inside your own servers without leaking information; what remains to be assessed is how it gets updated, who maintains the tooling and how service dependencies behave over time.

Lambert's testimony offers no recommendations, but it outlines a shifting balance. American leadership in closed models does not automatically translate into open model leadership. If anything, the more strategic local deployment becomes for companies and public administrations, the heavier the Chinese open ecosystem weighs.