China proposes export controls on 23 AI technologies

For years the AI export control story has been told from one direction. Washington decides which chips, and lately which model weights, may leave the United States, and Chinese labs look for workarounds. On 16 October 2026 Beijing wrote its own chapter. China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology published notice No. 48 of 2026, a draft revision of the Catalogue of Technologies Prohibited and Restricted from Export. It adds 40 technologies, and 23 of them are in artificial intelligence: large models, AI agents, computing clusters and information processing.

The draft is open for public comment until 15 November 2026, so the final list can still change. But one part already applies. From the day the notice was published, companies that want to export any of the listed technologies must first register the transfer on the ministry's online platform. That makes this more than a consultation paper. It is the first time China has put its own frontier AI models, and the people who build them, inside a formal export control regime.

Here is what the draft actually says, what it leaves out, and why the shape of these rules matters more than the headline number.

What is on the list

The catalogue itself is not new. According to Global Times, it was first published in 2001 and revised in 2008, 2020, 2023 and 2025. Technologies in the prohibited category cannot be exported at all. Technologies in the restricted category need an export licence. The new draft adds 14 entries to the prohibited category and 26 to the restricted one. By sector, 23 are in AI, 12 in photovoltaics and 5 in biomedicine.

On the AI side, a MOFCOM spokesperson gave concrete examples in a question and answer statement published by Beijing News. The prohibited group includes AI driven generation and mutation of malicious code, which the ministry says could be used to attack critical information infrastructure, defence systems and core government networks, or to get around national security defences. The restricted group includes AI assisted vulnerability discovery, the kind of automated bug hunting that frontier labs have been showing off all year.

Global Times adds more detail. The draft would prohibit exports of core technology for large models trained on China's sensitive national data, and of AI agents that serve critical national information infrastructure, national defence and core government affairs. Design technology for 2.5D and 3D AI chips and for 3D flash memory would be restricted, as would AI information processing and voiceprint recognition.

The solar and biotech parts are not the focus of this article, but they show the logic. Perovskite and tandem solar cell production, a field where Chinese companies lead, moves to the prohibited list. Certain cell and gene therapy platforms and small interfering RNA technologies join the biomedicine section. In all three sectors, Beijing is trying to protect areas where it believes it is ahead.

The thresholds for large models

The most closely watched item is the one for large models themselves. The detailed technical scope sits in an annex to the notice, distributed as a word processor file. Several outlets that read it, including The Decoder, report the same three triggers. The core technology of a large model becomes restricted if it meets any one of these conditions:

• training compute of 10^26 floating point operations or more
• 2 trillion parameters or more
• a top 10 global ranking on public evaluations at the time of export, judged from major domestic and foreign leaderboards that look at overall performance, specific tasks, safety or usage

"Core technology" is defined broadly. It covers the model weights, key checkpoints, source code, architecture design, and the methods used to speed up training or inference. In other words, everything a foreign buyer would need to rebuild or run the model on its own.

There is a second large model trigger that has nothing to do with size. According to the same reporting, a model that was trained on China's "important data", as defined by the Data Security Law, and is then offered as a service would also be restricted. Lawfully public data is exempt, and a trace share of up to 0.1 percent is tolerated if it was desensitised, never appears in outputs and does not materially lift the model's capability. Models trained on the more sensitive category of "core data" land on the prohibited list. Neither category is defined in the notice itself, which leaves room for later rules.

The 10^26 figure will sound familiar to anyone who followed the American debate. It is the same order of magnitude that Western regulators and researchers have used for years as a rough line for frontier scale training. China has now written a comparable line into its own rules, and applied it to its own developers.

What does not count as an export

The most important part of the draft may be what it leaves out. The MOFCOM spokesperson said that "the vast majority" of AI technologies will remain freely exportable once the rules take effect. The key question, according to the ministry, is whether restricted technology physically moves from inside China to outside it.

Three categories matter here.

Outside the rules entirely. Using a restricted model through an API, an app or a web chat, through a SaaS product, or as a lightweight on device version does not count as a large model technology export, the ministry said. Reuters noted that this distinction could let developers such as DeepSeek and Alibaba keep expanding abroad. Technology that has already entered the public domain through open source releases or published academic papers is also excluded from the catalogue.

Automatic licence. Open sourcing AI technology, running a Chinese company's own compute centre abroad, taping out chips overseas, joint research with foreign partners and participation in international standards work all qualify for "automatic licensing" and other simplified measures. The ministry describes this as "report and export": file the paperwork and the transfer is cleared. It also promised easier exports between members of the World AI Cooperation Organization, the body China launched for international AI cooperation, and support for capacity building in the Global South.

Needs a licence. Deploying restricted model technology locally on servers abroad does fall under export management. If the deployment runs on a company's own overseas cloud, which the ministry considers relatively controllable, the simplified process can apply. Transferring ownership of the technology, licensing it for foreign use, or a foreign party becoming the owner or actual controller of a Chinese technology all count as export cases that must be registered.

Registration starts now

While the catalogue is still a draft, the registration duty is not. The notice says that from its publication date until the revised catalogue formally takes effect, technology exporters must register before exporting anything on the annex list. Exports that have started but are not yet completed are included. Companies must give details about themselves and the recipient, the destination country or region, the technology and its purpose, and the expected value of the deal. Exporters are expected to judge for themselves whether a planned transfer falls under the list.

Comments on the scope, the technical indicators, the control policy and the facilitation measures can be sent to the ministry by email, fax or letter until 15 November.

Why Beijing says it is doing this

The official explanation has two halves. The first is safety and control. The ministry says AI brings internal and derived risks that worry governments everywhere, and that tools like autonomous malware generation could cause serious harm if non state actors or malicious entities got hold of them. The stated goal is to "ensure AI always remains under human control" and to advance international AI governance.

The second half is openness. The same statement says China wants to support an open source innovation ecosystem, which is why open releases get automatic licences and public technology is excluded. The ministry also stressed that the revision is routine, follows the Foreign Trade Law and is "not aimed at any particular country, region or entity", and that experts, industry groups and major companies were consulted before the draft was published.

Reuters places the move in a wider context. In July it reported that Chinese authorities had discussed with companies such as Alibaba, ByteDance and Z.AI whether to restrict overseas access to the country's most advanced models, including unreleased ones. It also noted that the cyber capabilities of frontier models such as Anthropic's Mythos have pushed Beijing to think harder about which AI tools should leave the country. At the same time, the United States and China set up a dialogue and an incident communication channel for advanced AI last month, so the two rivals are building rules in parallel rather than in a vacuum.

Why it matters

For people who run open models. The good news first: nothing in this draft switches off Chinese open weight models. Anything already released is public domain under the rules, and future open releases get an automatic licence. But the framework gives Beijing a lever it did not have before. Dean W. Ball, quoted by The Decoder, put it this way: the move "does not mean 'China is shutting down open-weight models,' but it does indicate the Chinese government establishing the optionality to shut down future open-weight models at will." An automatic licence is still a licence, and registration means the state sees every transfer.

For companies outside China. The rules mostly bite on private deals: a foreign company that wants a Chinese lab's frontier weights on its own servers, a joint venture, or an acquisition. API access stays free under the draft. That pushes foreign users toward exactly the two channels that are hardest to control anyway: open releases and cloud access. For European firms that hoped to run strong open models in their own data centres as a path to sovereignty, the message is mixed. Today's models remain usable. Tomorrow's top models could come with conditions.

For the security debate. By putting AI malware generation on the prohibited list and AI vulnerability discovery on the restricted list, China is treating offensive cyber capability as the clearest case for control. That lines up with the concerns Western labs have raised about their own models. Both sides now agree, at least on paper, that automated hacking tools are a category of their own.

For the export control race. Until now, the most visible export controls in AI were American restrictions on chips and model weights going to China. This draft shows Beijing building a mirror image system for the technology it considers its own crown jewels, with compute, parameter and capability thresholds of its own. Export control is no longer a one way street.

Dany's take

I expected a headline like "China bans AI exports" and the actual document is much more careful than that. The ministry clearly wants to keep the open source story going, because Chinese open models are a huge part of its global influence right now. So open weights get a fast lane and APIs stay free. What gets controlled is the quiet stuff: private weight transfers, on premise deployments abroad, and the offensive cyber tools nobody should be shipping anyway.

The part I would watch is the "top 10 global ranking" trigger. Compute and parameter counts are numbers you can check. A leaderboard ranking at the moment of export is a moving target, and it means a model can become restricted simply because others fell behind. Combine that with registration for every transfer and Beijing has a very flexible switch. Whether it ever flips it is a political decision, not a technical one. The comment period runs until 15 November, and the final thresholds are the thing to look for.

Sources: [MOFCOM and MOST notice No. 48 of 2026](https://www.mofcom.gov.cn/zcfb/blgg/gg/2026/art/2026/art_05e2643a8ca04eb9adf9dd0542991fe4.html), [MOFCOM spokesperson Q and A via Beijing News](https://m.bjnews.com.cn/detail/1792137269129878.html), [Reuters via StreetInsider](https://www.streetinsider.com/Reuters/Beijing+seeks+tighter+grip+on+AI+exports+amid+US+rivalry/27194895.html), [Global Times](https://www.globaltimes.cn/page/202610/1372405.shtml), [The Decoder](https://the-decoder.com/china-tightens-export-controls-on-open-weight-models-and-23-key-ai-technologies/), [AP](https://apnews.com/article/china-export-restrictions-ai-chips-solar-us-5ac381c0c6d6628bbea2e0da7d855a5b).