China's Draft AI Export Controls: What the List Covers, and What It Leaves Alone

On 16 October, China's Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48, a draft revision of the catalogue of technologies whose export is banned or restricted. Forty technologies would be added: 23 in AI, five in biomedicine and twelve in solar. Public comment runs until 15 November. The same day, a temporary pre-export registration requirement took effect for transfers, licences and changes of control that put the listed technology in foreign hands, and it reaches transactions already underway but not completed.

The AI entries carry the weight. Of the 40 proposed technologies, 14 would be prohibited outright and 26 restricted to licensed export. Inside AI the split is six prohibited and seventeen restricted. The prohibited side includes technology for AI-driven malware generation and mutation. The restricted side includes AI-assisted vulnerability discovery above defined benchmark scores, chip design work at 14 nanometres or finer, DRAM design at 18 nanometres or finer, cluster know-how around high-bandwidth interconnects and clusters above 10,000 accelerators.

For anyone using AI to make images and video, the question is narrower than the headline. Does this draft reach the models in your pipeline?

The thresholds are written for frontier language models

The restricted large-model entries are defined by three numbers. Core model technology means weights, checkpoints, source code, architecture and training or inference acceleration. Per the draft annex, a model lands in the restricted class if any one of these is true: training compute of at least 10 to the 26th floating-point operations, at least 2 trillion parameters, or a global top-10 ranking on mainstream evaluations at the time of export. Models trained on what the annex calls important data are listed on a separate track.

Those are frontier scale thresholds. Two trillion parameters and 10^26 training FLOPs describe the largest language models in current production, not the diffusion and video backbones that creative work runs on. No image or video model anyone would actually prompt comes close to that scale. So the draft, as written, is not aimed at the tool layer most creators touch. It is aimed one layer down, at the assets and designs that make a frontier model a frontier model.

Le Monde's framing put it in strategic terms: Chinese AI had been gaining ground abroad partly because it was offered as an accessible, cheaper alternative to American systems, and now Beijing wants to keep hold of the technologies while trying not to smother its own sector. The ministry's own summary leans the same way, describing the AI scope as precise and limited and saying most AI technology would remain freely exportable.

The carve-out that decides how this lands

The most consequential lines in the draft are the exclusions, because they draw the boundary between selling access and moving the thing itself.

Technology already released as open source or published in academic papers sits outside the control list. Serving a restricted model to users abroad through an API, an app or web interface, a SaaS product, or a lightweight on-device deployment is not treated as a technology export. Overseas on-premise deployment is, though a company deploying on its own overseas cloud can qualify for automatic licensing and other facilitation measures.

Read together, that means calling a hosted Chinese model from anywhere in the world is not the activity this draft regulates. Shipping the weights, the architecture or the training recipe across the border is. That distinction protects the workflow most creators actually depend on: API prompts, credits and subscriptions, including the practice of routing between video models by cost, which only works because providers are easy to switch between.

It also tells you where a change would first be felt. Because the exemption covers what has already been published, the existing open-weight checkpoints that reached third-party interfaces and local pipelines stay where they are. The exposure is forward-looking. A future flagship checkpoint that has not been released yet could need a licence before it can be published abroad, which is the single plausible path by which a creator's toolbox gets narrower rather than just more expensive.

Where this draft does touch a creator

Three situations move from background policy to your problem.

The first is self-hosting. If a studio or an agency runs inference on hardware it controls outside China, the draft classes that deployment as a technology export. The own-overseas-cloud carve-out may soften it, but the line is worth knowing before infrastructure plans assume that open weights today mean open weights everywhere tomorrow. The same tension runs through the wider argument about keeping confidential work on a self-managed stack, where the constraint has always been who else holds the model.

The second is a covered technology transfer. The registration rule applies to transferring ownership of listed technology abroad, authorising its use overseas, or any change that puts the technology's owner or actual controller in foreign hands. That is narrower than it sounds, and worth reading precisely: the ministry's own FAQ says API, app and web access, SaaS and lightweight on-device deployment are not technology exports, so a service arrangement alone does not trigger it. Those are the cases the pre-export registration requirement already covers, and it is live now, before the catalogue is final.

The third is timing. Comment closes on 15 November, and the numbers are draft numbers. The training-compute figure, the parameter count and the top-10 criterion could survive, tighten or move. Anyone building a pipeline that assumes a specific model will always be obtainable should read the final annex rather than the reporting around the draft.

What to do with this

Nothing urgent, if your image and video work runs through hosted APIs. That path was explicitly left out, and the thresholds sit far above the models behind it.

Pay attention if you are buying rather than renting: technology licences, overseas deployments, weight transfers, or a deal that puts ownership or actual control of listed technology in foreign hands. The registration step applies to those today.

And keep one habit from the last two years of churn. Treat the model as a replaceable component of a pipeline you own, not as the foundation. The argument for that habit has usually been about model churn and pricing. This draft is a reminder that it can also be about availability, decided by a ministry in a comment period you cannot vote in.