China Proposes Export Controls on 23 AI Technologies
China's Ministry of Commerce proposed sweeping new export controls on Friday, October 16, 2026, covering 40 strategic technology categories, 23 of them tied directly to artificial intelligence. The draft would give Beijing tighter oversight of AI chips, large language models, and the techniques used to build them, with specific thresholds that decide when a model's underlying technology needs a license to leave the country. China says the goal is curbing misuse, not a blanket ban, and most of the listed technologies can still be exported, often with simplified licensing. Public comment on the draft stays open until November 15, 2026, before Beijing issues a final version.
⚡ Quick facts
- What's proposed: draft export controls on 40 strategic technology categories, 23 of them tied to artificial intelligence
- Who announced it: China's Ministry of Commerce, in a draft unveiled Friday, October 16, 2026
- Key threshold: large AI models face controls once they hit 10^26 FLOPs of training compute, 2 trillion or more parameters, or a global top-10 ranking
- Comment window: public feedback accepted through November 15, 2026, before Beijing issues final rules
- Not a blanket ban: most AI exports can continue, often under simplified licensing, and using a restricted model through an API or cloud app stays unaffected
What the draft rules cover
China's Ministry of Commerce published draft export controls on Friday, October 16, 2026, covering 40 strategic technology systems. Of those, 23 are tied to artificial intelligence, spanning AI chips, large language models, and the techniques used to build them. The same draft adds similar oversight for solar energy and biotechnology exports, giving Beijing broader control over cross-border technology transfers generally. China said the goal is to curb misuse of the listed technologies, not to cut off exports outright, and most of the 40 categories remain exportable, in many cases through a simplified licensing path. The ministry opened a public comment period that runs through November 15, 2026, after which it plans to issue a final version of the rules.
How a large AI model gets flagged
Buried in the draft is a specific, numeric test for when a large language model's underlying technology, meaning its weights, checkpoints, source code, architecture, or training and inference methods, becomes subject to export control. A model trips the threshold if it was trained using data that counts as "important data" under China's Data Security Law, unless that data was already public or fell under a narrow allowance for tiny amounts used under specific conditions. Separately, a model becomes controlled once it crosses any of three lines: 10^26 FLOPs of training compute, 2 trillion or more parameters, or a spot in the global top 10 on public rankings that weigh performance, safety, and real-world use. Cross any one of those lines, and the model's core technology needs registration or a license to leave the country.
Two categories get a harder line regardless of size. Autonomous AI tools built to write malware targeting critical infrastructure, military systems, or government networks face a total export ban. AI tools designed to hunt for software vulnerabilities face new limits too, though those fall short of an outright ban.
What stays open
Beijing built in several carve-outs. Chinese companies open-sourcing a model, running overseas data centers, sending chips out for fabrication, or working with foreign researchers and international standards bodies can qualify for fast-track "automatic licensing" rather than a full review. The controls also do not reach into how people actually use these models day to day: accessing a restricted model through its API, a hosted app, a website, software-as-a-service, or directly on a device is untouched by the draft. The license requirement only kicks in if a company wants to install the model on its own servers outside China, and even then the process is meant to move faster if those servers sit on the company's own overseas cloud.
Why the timing is awkward
The draft lands at a delicate moment. Washington and Beijing agreed last month to cut tariffs on $30 billion of goods and open a dialogue on AI, following President Xi Jinping's visit to the White House. China's AI-driven exports had just jumped more than 36 percent year-over-year in September, well above the 25.5 percent analysts had expected, according to IndexBox data. Leonhard Xu, a senior consultant at the Berlin-based consultancy Sinolytics, called the announcement "extremely expansive," adding that its registration requirement starts immediately, even for exports already shipping, and cautioned that "export control does not equal an export ban in most cases." Wendy Cutler of the Asia Institute was blunter about what the move signals, saying it shows "the fragility of our trade truce cannot be overstated" and that it is "only a matter of time" before Washington answers with restrictions of its own. Separately, Erin Browne of the US Treasury told Reuters that Washington is pushing its partners at the International Monetary Fund to rename AI "superintelligence," a term the Trump administration has favored.
Who else the rules could touch
Ting Peng, a trade compliance partner at BZW Law Firm in Shenzhen, pointed out that the registration requirement reaches beyond typical export paperwork. It also covers certain cross-border intellectual property transfers, licensing arrangements, and changes in ownership or control of domestic technology moving to overseas entities. She flagged the biopharmaceutical sector in particular, where cross-border mergers and acquisitions have been active over the past two years, as an industry that should check whether its deals now trigger registration.
What's confirmed and what isn't
Confirmed: China's Ministry of Commerce has published a draft covering 40 technology categories, 23 of them AI-related, with specific compute, parameter, and ranking thresholds for large language models, a public comment period running to November 15, 2026, and carve-outs for open-source releases, API access, and cloud-hosted use. Not confirmed: the final wording of the rules, which could still change during the comment period, and how the US will respond. No US countermeasure had been announced as of this article's publication, though several analysts expect one.
Primary reporting from The Decoder, sourced from China's Ministry of Commerce draft factsheet, with additional reporting from The National. The draft remains open for public comment until November 15, 2026, and its final wording may change; confirm current details before relying on them.
Frequently asked questions
What did China propose?
China's Ministry of Commerce unveiled draft export controls on October 16, 2026, covering 40 strategic technology categories, including 23 linked to artificial intelligence. The draft would give Beijing tighter oversight of AI chips, large language models, and related technology transfers, alongside similar controls on solar energy and biotechnology exports.
Does this mean China is banning AI exports?
No. Most of the listed technologies can still be exported, and officials say licensing will often be simplified rather than restricted. A small number of categories face a full export ban, including autonomous AI tools built to write malware targeting critical infrastructure, military, or government systems. Everything else, including most large AI models, would need registration or a license rather than an outright ban.
Which AI models would actually be affected?
The draft targets a model's core technology, meaning its weights, checkpoints, source code, architecture, or training methods, once it is trained on data classified as important under China's Data Security Law, or once it crosses one of three thresholds: 10^26 FLOPs of training compute, 2 trillion or more parameters, or a ranking in the global top 10 by public benchmarks. Public datasets and small amounts of flagged data are exempt under published conditions.
Can people outside China still use restricted Chinese AI models?
Yes, for now. Using a restricted model through its API, a hosted app, a website, software-as-a-service, or on a device stays unaffected by the draft. Installing the model on a company's own servers located outside China would need a standard export license, though that process is meant to be streamlined if the servers sit on the company's own overseas cloud infrastructure.
When do the rules take effect, and how is the US responding?
Beijing is taking public comment on the draft until November 15, 2026, before issuing a final version. Analysts note the move lands weeks after the US and China agreed to cut tariffs and open an AI dialogue following President Xi Jinping's White House visit, and several expect Washington to respond with its own new restrictions. US Treasury officials have separately pushed to rename the technology "superintelligence" in international forums.