← Back to feed

Beijing is drafting AI export controls that would reverse the posture that made Chinese open-weight models a global infrastructure layer. The Ministry of Commerce has held consultations with Alibaba, ByteDance, and Zhipu, mapping out rules across four categories: chip design IP, model weights, training data, and foreign acquisitions.

The rules have not been enacted. But their scope represents a structural shift: China spent three years publishing open-weight frontier models while the US tightened chip export controls. That strategy closed the intelligence gap to single digits. Now Beijing may adopt the same closed posture the US used against it.

The Four Proposed Controls

1. TSMC chip design restriction. The most architecturally novel element: the draft would bar TSMC from manufacturing chips designed using Huawei or Alibaba architectures. Chinese chip IP would stay on Chinese silicon. The US has spent years preventing China from accessing TSMC’s advanced nodes — this is Beijing doing the reverse, preventing Chinese chip designs from being manufactured abroad where the IP could be extracted or studied.

2. Model weight export restrictions. Weights for frontier open-weight models — currently downloadable globally — would require state approval to export. The practical effect: Kimi K3-class models could remain open within China while being restricted for download from foreign infrastructure. This targets the Hugging Face distribution model that has made Chinese open-weight releases globally accessible within hours of launch.

3. Training data export controls. Data used to build frontier Chinese models would be classified as a controlled export. This includes synthetic datasets, RLHF annotation sets, and filtered pretraining corpora. The concern is that the data pipeline, not just the weights, conveys competitive advantage.

4. Foreign acquisition blocks. Following the forced unwind of Meta’s $2B Manus acquisition after Chinese regulators intervened, the new draft would codify deal-blocking for foreign buyers pursuing Chinese agentic AI startups.

Kimi K3 as the Tipping Point

China’s open-weight strategy was an underdog play. DeepSeek, Kimi, and Qwen released capable models at a fraction of Western frontier costs because openness expanded distribution and accelerated adoption without requiring the user base that OpenAI and Anthropic have built.

Kimi K3 changed the calculation. At 93.4% SWE-bench Verified and 78.5 on LiveBench, the frontier intelligence gap closed from 13 index points in January 2025 to roughly 2-3 points in July 2026. Beijing no longer sees itself as playing catch-up. The draft rules reflect that: when you have the best or near-best model, openness is a gift to rivals, not a competitive lever.

“The real shift is psychological,” one observer noted in Rohan Paul’s newsletter. “Beijing now guards AI like it once guarded rare earths. Openness was the underdog’s strategy, and China no longer sees itself as one.”

Industry Pushback

Chinese companies consulted in the process have warned that restricting open-weight distribution would slow their own progress. The feedback loop between global developer adoption and model improvement is real — community fine-tunes, use-case discoveries, and external red-teaming all feed back into subsequent training runs. Closing weights removes that signal.

The chip design restriction has a separate problem: Huawei and Alibaba fabs depend on a domestic semiconductor ecosystem that is still maturing. Locking chip IP to Chinese manufacturing capacity may slow commercial deployment of designs that would otherwise be fabricated faster at TSMC.

No timeline for formalising the rules was disclosed. The Ministry of Commerce has not published the draft publicly.