China Proposes AI Tech Export Bans, Impacting Open-Weight Models

China's Ministry of Commerce and Ministry of Science and Technology jointly issued a draft plan on Friday proposing export prohibitions and restrictions on 23 AI-related technologies. The technologies include models, agents and chip designs. Analysts say the proposals aim to tighten Beijing's control over the strategically important sector and create uncertainty about domestic model developers' global commercial prospects. The measures come amid growing global concerns about advanced AI risks following recent cyberattacks, underscoring AI's national security importance for Beijing and Washington. The draft also proposes export bans and restrictions on critical biotechnology and photovoltaic technologies.

The proposed restrictions on foundational AI models include thresholds covering models trained on China's important data, developed using computing power reaching 10 to the power of 26 FLOPs, sized at 2 trillion parameters or more, or ranked in a global top 10 leaderboard. The computing threshold is widely used to identify frontier AI models. Authorities said on Friday that the month-long public consultation would last until November 15.

Zhang Yi, founder and chief analyst at consultancy iiMedia, said the measures would inevitably introduce short-term adaptation costs for companies in China targeting international markets. However, Zhang said Beijing's increased scrutiny should not be interpreted as an effort to prevent companies in China from deploying AI services globally. Zhang said the key was to control core engineering technologies while keeping commercial services open.

A Ministry of Commerce spokesperson said on Friday that the scope of AI technologies covered was targeted and limited. The spokesperson said overseas AI services delivered through official application programming interfaces and mobile apps would not fall under the new export restrictions. The spokesperson said the vast majority of AI technologies would remain freely exportable once the measures formally took effect. The spokesperson said the controls aimed to ensure AI technologies remained secure and controllable amid risks including misuse by non-state actors and malicious entities.

The treatment of open-weight AI technologies under the final export regime remains uncertain, with China leading globally in this area. Growing numbers of developers in China, including Kimi maker Moonshot AI and GLM maker Z.ai, also known as Zhipu AI, are entering revenue-sharing agreements with major US hyperscalers such as Amazon Web Services. The agreements deploy open-weight models on hyperscalers' cloud infrastructure and represent a major new revenue source. Several models from China, among the world's leading open-weight systems, would meet the draft's technical thresholds. However, the Ministry of Commerce spokesperson said China continued to encourage open-source AI and would expedite new licensing requirements through "immediate clearance upon filing".

Alex Colville, an Australian Strategic Policy Institute analyst specialising in China's AI sector, said compliance requirements might delay or prevent partnerships with global cloud providers to deploy companies' most powerful open-weight models. Colville said such deployments would involve exporting model weights outside China. Colville said much depended on the final wording, including what automatic approval for open-sourcing AI technologies would mean in practice. Colville said China's regulators had sought flexibility and could change export requirements if the field changed dramatically within the next six months.

AI has become increasingly central to China-US relations as both countries seek to address risks amid commercial and geopolitical competition. China's President Xi Jinping and US President Donald Trump agreed during last month's Washington summit to establish a bilateral dialogue to manage AI risks. AI is expected to be a major agenda item at the leaders' two remaining meetings this year, at the Asia-Pacific Economic Cooperation summit in Shenzhen next month and the G20 summit in Miami in December.

You Yunting, senior partner at Shanghai DeBund Law Offices, said the measures appeared to be a targeted response to Meta Platforms' aborted acquisition of AI agent developer Manus earlier this year. One proposed measure would require prior registration with the Ministry of Commerce before restricted technologies from China could pass to an overseas owner or controller. Manus' parent company, Butterfly Effect, was founded in mainland China in 2022. Butterfly Effect relocated to Singapore last year, allowing the company to sidestep US technology export and investment controls and paving the way for Meta's $2 billion acquisition. However, Beijing halted the transaction soon afterward on national security grounds. Manus announced last month that Manus had formally resumed independent operations.

You said compliance costs would rise in the short term for domestic AI companies expanding abroad. However, You said export controls could deliver long-term competitive advantages for China. You argued that reasonable controls were necessary to protect the substantial domestic investments in the sector.