Dashboard

China AI Export Controls: What Builders Should Know

China's draft export catalogue adds 23 AI technologies, 6 of them prohibited. For most builders outside China the practical effect is close to zero, and the reason is one specific clause.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
19 October 20261 min read

China's proposed AI export controls would add 23 artificial intelligence technologies to the country's restricted export catalogue, 6 of them prohibited outright and 17 requiring a licence. For almost everyone building with AI outside China, the practical effect is close to zero, because using a model through an API is explicitly not an export. The people it does touch are the ones who take Chinese open-weight models and do something with them beyond running inference.

Here is what the draft says and how to work out which group you are in.

What Announcement No. 48 actually does

On 16 October 2026 the Ministry of Commerce and the Ministry of Science and Technology jointly issued Announcement No. 48 (2026), a draft revision to China's Catalogue of Technologies Prohibited or Restricted from Export. It proposes adding 40 technologies across three sectors, according to a detailed reading by Geopolitechs.

Sector

Prohibited

Restricted

Total

Artificial intelligence

6

17

23

Photovoltaics

5

7

12

Biomedicine

3

2

5

Total

14

26

40

The AI categories cover foundation models, AI agents, chip design, computing cluster technology and cybersecurity applications. Public comment runs from 16 October to 15 November 2026, so this is a consultation draft, not law.

One part is already in force. From publication until the revised catalogue takes effect, anyone exporting a listed technology must pre-register the transaction through MOFCOM's Technology Trade Management Information System, giving the parties, destination, technology details, intended use and estimated value. That obligation reaches transactions that started before 16 October and have not completed.

The three carve-outs that decide whether this touches you

Most coverage of this story has flattened the carve-outs into "open source is exempt". That is not quite what the draft says, and the difference matters.

  1. Already public is out of scope. Technology lawfully in the public domain, through an existing open-source release, an academic paper or otherwise, is not subject to the controls. Weights already on Hugging Face under a permissive licence do not retroactively become controlled.

  2. API and SaaS access is not an export. Serving a model through an API, a website, an app or a SaaS platform does not count as a technology export. This is the clause that removes the overwhelming majority of builders from the picture.

  3. New open-source releases get facilitated treatment, not an exemption. Open-source releases of controlled AI technology are eligible for simplified handling, including automatic licensing on what MOFCOM's spokesperson described as a "register and export" basis. That is a lighter process, not an absence of process.

Read together, these say something fairly specific: China is not trying to stop the world from using Chinese models. It is trying to see, and in six categories to stop, the transfer of the methods behind them.

Who this actually reaches

Work through it honestly.

  • You call a Chinese model's API. Not an export. Nothing changes.

  • You download open weights published before this draft and run them. Already public, out of scope.

  • You fine-tune an existing open-weight model for your own product. Still working from published material. The draft is aimed at the transfer of technology from China outward, not at what a foreign company does with material it lawfully obtained.

  • You are a Chinese entity licensing model technology, training methodology or cluster design to an overseas party. This is the actual target. Register, and check whether the category is restricted or prohibited.

  • You have a joint development arrangement with a Chinese lab, or an acquisition that moves control of a Chinese AI asset overseas. The registration requirement explicitly covers changes of ownership or control, including deals already in flight.

If you are in the last two groups you need a lawyer, not a blog post. If you are in the first three, the honest answer is that this is a story to note rather than act on.

The second-order effect worth watching

The one real risk for ordinary builders is supply, not compliance. A lot of the cost advantage in the open-weight tier comes from Chinese labs publishing competitive models for free. If six prohibited categories make Chinese labs more cautious about what they publish, the open-weight frontier could thin out, and the gap between open-weight and closed models would widen again after two years of narrowing.

That would show up in your costs before it showed up in your news feed. It is the same shape as the broader question of what an AI slowdown means for small builders: the effect arrives as a price change, not an announcement.

Nothing in the draft forces that outcome. The facilitated-licensing route for open-source releases reads like a deliberate attempt to avoid it. But if you have a cost model that assumes a free competitive open-weight option will always exist, this is a reason to check whether self-hosting actually saves you money at your volume, before you need the answer in a hurry.

What to do this month

For most readers, three things, none of them urgent:

  • Note which of your dependencies are Chinese open-weight models, and under which licence you obtained them. The licence terms on the weights govern your use, and this draft does not change a licence you already hold.

  • If a model matters to your product, keep a local copy of the weights you depend on rather than assuming the download stays available.

  • If you are in a regulated sector, add this to the same watchlist as domestic rules like California's AI audit laws. The consultation closes 15 November and the final catalogue will differ from the draft.

The general pattern worth internalising is in our guide to AI risks for builders: regulatory risk rarely arrives as a rule that bans what you are doing. It arrives as a change in what your suppliers are willing to give away.

FAQ

Does this ban Chinese AI models outside China?

No. The draft governs the export of technology from China. Using a Chinese model through an API, or running open weights that were already published, is not covered.

Is open-sourcing a model now illegal in China?

No. Technology already in the public domain is outside the controls, and new open-source releases of controlled technology are eligible for simplified, automatic licensing. Six prohibited categories cannot be exported, including by publication.

When does this take effect?

The catalogue revision is a draft open for public comment until 15 November 2026. The pre-export registration requirement took effect on 16 October 2026 and runs until the revised catalogue does.

Do I need to register anything?

Only if you are exporting listed technology from China. Buying access to a model, or downloading published weights, is not an export under the draft.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.