China Wants an Export Licence for Your Two-Trillion-Parameter Brain

For years the main export control story in AI ran one way: Washington deciding which chips Chinese companies could buy. This week Beijing published its own list, and it is aimed at something you cannot load onto a container ship. It wants a say over where the brains go.

China’s Ministry of Commerce and Ministry of Science and Technology have released a draft revision of the country’s catalogue of technologies restricted for export. According to the draft, as reported by Reuters, AP and The Decoder, 40 technologies would be added, and 23 of them relate to AI. The list includes large language models, AI agents, chip designs and computing clusters of more than 10,000 accelerators.

What would need a licence

The key part covers frontier models. Exporting the core technology of a model would need a licence if the model meets any one of three tests: at least 2 trillion parameters, at least 10^26 FLOPs of training compute, or a place in the global top ten on capability rankings. “Core technology” here means model weights, checkpoints, source code, architecture and training or inference methods. In other words, the recipe and the finished cake.

There are notable exceptions. Selling access through an API, an app, a website or a software-as-a-service product is not covered, so a foreign customer can still use a Chinese model. They just cannot take it home. Open-source models that have already been published are exempt, and new open-weight releases would get automatic licensing after a filing. Deploying a controlled model on an overseas customer’s servers would need a licence, while a Chinese company running it on its own overseas cloud could use automatic licensing. The draft is open for public comment until November 15.

Why it matters

There is a certain symmetry here. The United States spent the past few years treating advanced chips as strategic assets. China is now treating the models trained on them the same way. Both sides seem to agree that the most valuable thing in AI is no longer the hardware alone but the weights, the huge file of numbers that holds what a model has learned.

For businesses, the practical effect is fairly narrow for now. Most companies outside China that use Chinese models do so through hosted services or through open weights that have already been released, and both of those routes appear to stay open. The firms most affected would be those hoping to license and run the newest, largest Chinese models on their own hardware abroad.

The exemption for published open-source models is the part that will draw the most smiles. Once weights are on the internet, they have been copied onto more hard drives than anyone could count. A rule that controls only future releases accepts reality: you cannot put toothpaste back in the tube, but you can make the next tube come with paperwork.

It also hints at where the competition is heading. If the top ten models in the world become controlled goods in both Washington and Beijing, frontier AI will start to look less like software and more like aerospace, with licences, filings and lawyers attached.

The draft can still change before it is final. For now the customs question has a new answer. Anything to declare? Just two trillion parameters, officer.