China Proposes Export Controls on 23 AI Technologies
China has proposed export bans and restrictions on 23 AI related technologies, including models, agents and chip designs, under a draft plan issued by the Ministry of Commerce and the Ministry of Science and Technology. The same document proposes controls on technologies tied to biotechnology and solar manufacturing. According to SCMP, the stated aim is to tighten Beijing's grip on a sector it treats as strategically vital.
The plan sets thresholds for foundational models. Controls would bite where a model was trained on what the document calls China's important data, was developed with computing power of 10^26 FLOPs or more, reaches 2 trillion parameters, or ranks in the global top 10 on a leaderboard.
Officials describe the scope as narrower than it looks. A commerce ministry spokesperson called it "targeted and limited" and said services delivered overseas through official programming interfaces and mobile apps would not be caught. And once the measures take formal effect, "the vast majority of AI technologies will still be freely exportable", the spokesperson said.
Analysts see a near term cost. Zhang Yi, founder and chief analyst at the consultancy iiMedia, said the measures would inevitably bring short term adaptation costs for Chinese companies at the point when they are chasing international markets. He argued the intent is to keep core engineering controlled while commercial services stay open.
The unresolved question is open weight AI, an area where China leads. Several Chinese models would meet the thresholds as drafted. Moonshot AI, maker of Kimi, and Zhipu AI, which builds the GLM family, have been signing revenue sharing deals with US cloud providers such as Amazon Web Services.
Alex Colville of the Australian Strategic Policy Institute said the compliance burden could delay or block partnerships with global cloud providers, because deploying those models abroad means exporting their weights. But much depends on the final wording, he said, including what "automatic approval" for open sourcing would mean in practice.
The commerce ministry says it still encourages open source AI. New licensing requirements are to be expedited as "immediate clearance upon filing", its spokesperson said. The measures are also meant to "ensure the security and controllability" of AI technology against misuse by "non state actors and malicious entities".
Lawyers read the plan as a response to a specific deal. You Yunting, senior partner at Shanghai DeBund Law Offices, said the measures appeared to be a targeted reaction to Meta's aborted acquisition of the Chinese AI agent developer Manus. One proposed provision would require registration with the commerce ministry before restricted technology changes hands to an overseas owner or controller.
Manus was founded in mainland China and its parent, Butterfly Effect, moved to Singapore, which let the company sidestep US export and investment controls. That cleared the way for a $2 billion sale to Meta. But Beijing halted the deal on national security grounds, and Manus said it had resumed independent operations.
You said compliance costs for Chinese firms going abroad will rise in the short term, while the controls could bring long term advantages. "Given the substantial domestic investments in the sector, implementing reasonable controls is necessary", he said. A month long public consultation on the draft now follows.
The draft covers 23 AI related technologies, among them models, agents and chip designs
Model thresholds include 10^26 FLOPs of computing power or 2 trillion parameters
The commerce ministry says overseas services delivered through interfaces and apps are not covered
Open weight models are the key uncertainty as a month long consultation begins
Source: SCMP
