- China's draft export plan would restrict 23 AI-related technologies. A ministry spokesperson says official APIs and mobile apps would not be covered.
- Analyst Alex Colville says open-weight models are a key uncertainty, because running them on foreign clouds means exporting the weights.
- Leaders should map which AI models they rent and which they host, and watch the draft before consultation ends November 15.
An AI model can reach your company in two ways. You can rent it as a service. Or you can hold it as a file. A Chinese ministry spokesperson's remarks suggest the new draft export rules may treat those two paths differently. How they treat the second path, downloaded weights, is not yet settled. Do you know which path each of your AI tools depends on?
What China has proposed
On Friday, China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology issued a draft plan. According to the South China Morning Post (SCMP), it would ban or restrict exports of 23 AI-related technologies. These include models, agents and chip designs. The draft also covers some biotechnology and photovoltaic technologies.
The measures are a draft. A public consultation runs until November 15. Alex Colville, an analyst at the Australian Strategic Policy Institute, says much depends on how the final text words its key provisions.
Where the line falls
For foundational models, the draft sets several thresholds. A model could be restricted if it was trained on "China's important data". It could also be restricted if it was built with computing power of 10²⁶ FLOPs, has 2 trillion parameters or more, or ranks in a global top 10 leaderboard.
FLOPs count the arithmetic a computer performs while training a model. Parameters are the adjustable numbers inside it. The SCMP describes 10²⁶ FLOPs as a widely used marker for frontier models.
Rented services versus downloaded weights
A MOFCOM spokesperson said some ways of selling AI abroad would not be covered. These include official APIs and mobile apps. An API is the connection a customer uses to send requests to a model that stays on the provider's servers. The spokesperson said that once the measures take effect, "the vast majority of AI technologies will still be freely exportable".
Open-weight models are the harder case. The weights are the learned values that make up the model. Releasing them lets others run the model on their own hardware. The SCMP calls their treatment under the final rules a key uncertainty.
Colville, who specialises in Chinese AI, sees a possible snag. Putting a top Chinese open-weight model on a foreign cloud means moving its weights out of China. New compliance steps could slow or stop that for the most powerful models.
This matters commercially. The SCMP reports that Moonshot AI, maker of Kimi, and Z.ai, maker of GLM, are signing revenue-sharing agreements with US hyperscalers such as Amazon Web Services. The deals put their open-weight models on those clouds. The report says several Chinese models would meet the draft thresholds as currently worded.
MOFCOM says it continues to encourage open-source AI. It said new licensing would be expedited as "immediate clearance upon filing". Colville said it is unclear what "automatic approval" would mean in practice.
One lawyer links the draft to the Manus deal
The draft also covers ownership. One proposed measure would require registration with MOFCOM before restricted technologies pass to an overseas owner or controller.
Lawyer You Yunting of Shanghai DeBund Law Offices told the SCMP the measures seemed to be a targeted reaction to Meta's aborted acquisition of Manus. The draft itself does not say so. This is one lawyer's reading.
The SCMP reports that Manus's parent company, Butterfly Effect, moved to Singapore last year. That let the firm avoid US export and investment controls and cleared the way for a $2 billion purchase by Meta. Beijing called a halt soon after, citing national security. Manus announced last month that it had resumed independent operations.
The idea behind the rules
The lesson here is that AI supply chains may now have a legal layer, not only a technical one. A model you call through an API and a model you download carry different dependencies. The first rests on a provider's service. The second rests on a transfer of files that a government may license.
Colville noted that China has been flexible in AI regulation. If the field changes sharply within six months, he said, it can alter the export requirements. For anyone planning around these models, that points to rules that may move.
Questions to put to your team
These questions apply whether or not the draft passes in its current form.
First, which AI models do we use, and which come from Chinese developers? Second, do we call each one through an API, or do we host the weights ourselves or through a cloud provider? Third, what do our contracts say if a supplier can no longer deliver weights or updates? Fourth, could we switch to another model without rebuilding the product around it? Fifth, who owns the job of reading the final text once the consultation closes?
A model is easy to adopt and hard to replace. The draft is a reminder to know which of your AI tools you rent and which you hold.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: South China Morning Post.
