China locks down AI model weights with immediate export controls
China made AI model weights a controlled export. Deploying a frontier Chinese model on your own servers now requires a government license.
Policy · Source: TechTimes
What happened
China's Ministry of Commerce and Ministry of Science and Technology just dropped Announcement No. 48. It proposes adding 40 technologies to the national export control catalogue, with 23 specifically targeting the artificial intelligence stack. The draft covers model weights, source code, and underlying architecture. Under the new rules, 14 technologies face outright bans. Another 26 require specific government licenses before they can leave the country.
The restrictions target massive frontier systems. Models with over two trillion parameters, models trained with at least 10^26 operations, or those ranking in the global top ten on performance benchmarks are restricted. High-performance AI agents designed for national defense, critical information infrastructure, or core government systems are completely banned. The rules also capture hardware infrastructure. Computing clusters of 10,000 or more accelerator cards and chip designs at or below the 14-nanometer node are caught in the net.
The formal rules remain open for public comment until November 15. But the government is not waiting around. An interim pre-export registration requirement took effect immediately on October 16. Any overseas licensing agreement, ownership transfer, or foreign control change involving these technologies must be registered right now. Deals already underway are covered. Exporters must disclose basic information, intended use, and transaction value to the Commerce Ministry.
Key facts
- Announcement No. 48 — The joint ministry proposal adding 40 technologies to China export control list
- 2 trillion — The parameter threshold that triggers licensing requirements for AI models
- October 16 — The date the interim pre-export registration requirement took effect
- 10,000 — The number of accelerator cards in a computing cluster that triggers restrictions
Why it matters
The line between software and sovereign asset is now drawn at the server rack. If you call a Chinese model like DeepSeek or Alibaba through an API, you are fine. But if you try to download their restricted weights to host on your own infrastructure, you trigger a technology export. Builders relying on cheap, powerful Chinese models for on-premise deployments must now navigate Beijing's compliance maze. The cost gap that made Chinese models attractive still applies to API users, but the infrastructure game has changed entirely.
This directly impacts the sovereign cloud rush in the Middle East, Asia, and Africa. Governments and massive enterprises have spent the last two years trying to run capable models on their own hardware to avoid oversight from Washington. Now, China is mirroring that exact strategy. By controlling the weights, Beijing ensures it retains leverage over anyone building critical infrastructure on Chinese artificial intelligence. Model weights are no longer just math. They are a controlled geopolitical commodity.
For builders
API users dodge the compliance bullet
Software-as-a-service and API access are explicitly exempt from these controls. Cloud providers and developers using Chinese models via API pay nothing extra and face no new legal friction. This preserves the cloud revenue streams Chinese AI companies have built internationally.
Open source models remain untouched
Technologies already in the public domain are carved out of the restrictions. If a model is already released as open-source or published in academic papers, you can keep building with it. Models already in global circulation remain beyond Beijing's legal reach.
Deals face immediate registration hurdles
The interim rules are already active. Startups and investors involved in cross-border acquisitions, hardware development, or IP licensing of Chinese AI assets will lose time and money dealing with immediate compliance checks. Deal timing and ownership diligence are now massive bottlenecks.
My take
I have been saying AI models are the new geopolitical borders. Washington tried to control the chips, and now Beijing is controlling the weights. If you build infrastructure that relies on downloading frontier models from another country, your business is at the mercy of their foreign policy. You cannot build a sovereign product on someone else's permission slip.
Original reporting: TechTimes. This is my rewrite and opinion.