China Proposes AI Export Controls While Keeping Model Services Accessible

China proposed new export restrictions covering 23 categories of artificial-intelligence technology on Friday, while setting out a distinction between transferring restricted technology abroad and letting overseas customers use a service. The draft is part of a wider review covering 40 technologies across AI, biotechnology and solar manufacturing.

The commerce and science ministries are accepting comments through November 15. Alongside the draft, they introduced temporary pre-export registration for specified technology transfers, including relevant transactions already underway. The consultation and the registration requirement therefore have different timing: the proposed catalogue is not yet final, while registration begins with the announcement.

Access to a model is different from transferring it

The commerce ministry’s explanation says overseas deployment involving restricted technology would fall within the proposed controls. Providing access through APIs, apps, websites and software-as-a-service would sit outside the large-model technology export framework described in the proposal. It also sets out facilitated treatment for certain activities and excludes technology already in the public domain through open-source release or academic publication.

For a customer, those distinctions can produce very different purchasing choices. Calling a hosted service means sending a request to a provider-operated system. Acquiring technology for a locally operated deployment gives the customer a different degree of possession and control. A business cannot assume that permission for one arrangement establishes permission for the other.

A sourcing issue for AI buyers

The practical question for an overseas buyer is which part of its planned arrangement constitutes a transfer. A subscription, a technology licence and an acquisition of intellectual property are different transactions. Contracts that appear similar commercially can require different treatment when the underlying technology and deployment model differ.

The draft should not be read as a blanket withdrawal of Chinese AI services from international markets. Nor does continued service availability establish that every form of local deployment will remain equally accessible. The final wording and its application will determine that boundary.

Deployment choices also change business continuity risk

A hosted service can reduce the buyer's need to operate computing infrastructure, while making continued access dependent on the provider and the service agreement. A locally operated system can provide a different degree of operational control while requiring equipment, maintenance and permission to obtain the relevant technology. The draft's distinction therefore intersects with ordinary procurement economics; it is not only a question of where an invoice is sent.

A buyer assessing a planned deployment needs a documented description of what crosses the border and what remains with the supplier. Model access, software delivery, technical support and transfer of know-how are not interchangeable descriptions. The applicable review depends on the actual arrangement and the final rules. A commercial label such as partnership cannot resolve those questions by itself.

Timing is equally important for transactions already in progress. A consultation deadline gives a date for comments, not an automatic assurance that a proposed transfer can proceed under unchanged conditions until then. The temporary registration requirement needs separate attention. Companies should distinguish what the announcement says is already operating from what remains proposed.

The broader financial implications would emerge through approved transactions, changes in project schedules or disclosed alterations to contracts. A draft list alone does not quantify lost export revenue or establish how many overseas deployments will be affected. The material new issue is the proposed boundary between technology transfer and service access, with the final commercial effects still uncertain.

Why the distinction matters beyond China

TTN’s coverage of Germany’s planned industrial AI agreement concerns another part of the same procurement problem: where computing is operated, who supplies it and what a customer actually receives. Access to capacity and ownership of technology are separate commercial assets.

For companies planning multi-year AI projects, the draft adds a reason to document deployment assumptions before signing. An architecture built around a hosted service may have different operational dependencies from one built around technology installed locally. Friday’s announcement brings that distinction into policy, with the consultation still open and no basis yet for treating the proposed catalogue as final.