China has proposed adding 23 artificial intelligence technologies, from frontier model weights to methods for running large chip clusters, to its list of technologies that are banned or restricted from export, and companies have until Nov. 15 to comment. The Ministry of Commerce and the Ministry of Science and Technology published the draft on Oct. 16 as Announcement No. 48, alongside 17 additions covering biomedicine and solar manufacturing.

For investors in Chinese AI, the draft draws a line between selling access to a model and selling the model itself. Reuters reported that overseas customers would broadly still be able to use Chinese AI through websites and cloud platforms, while some transfers of the underlying technology, and some deployments abroad, would need government approval. That matters for model makers such as DeepSeek, Moonshot AI and MiniMax, which serve users outside China, and for any foreign buyer eyeing a Chinese AI startup. For Chinese tech companies raising money in a Hong Kong IPO market already on a record pace, the rules add a new question about what they can license or sell abroad.

What China's AI export draft covers

The announcement lists 40 technologies in three fields. The 23 AI items cover large models, AI agents, computing clusters and information processing, according to the ministry's notice. The biomedicine list has five items, covering cell and gene therapy, small nucleic acid drugs and pathogenic microbes, and the solar list has 12, covering polysilicon, wafer and cell manufacturing.

Technology that has already entered the public domain through open-source releases or academic papers would not be covered, the ministry said. That carve-out is central for Chinese labs, most of which publish model weights.

For frontier models, ChinaTalk, a newsletter on Chinese technology policy, reported that the draft would require licenses to export model weights and core private know-how, such as key training checkpoints, source code, architecture designs and methods for speeding up training and inference. A model is covered if it crosses any one of three tests: training compute of at least 10^26 operations, at least 2 trillion parameters, or a capability level ranked among the world's best. Reuters described the third test as a top-10 ranking in certain performance, use or efficiency assessments.

ChinaTalk counted 14 of the 40 additions as outright export bans and 26 as restrictions that would require a license. Citing Epoch AI, it said no released Chinese model was estimated to have crossed the compute test as of Oct. 20, with the largest at roughly 2x10^25 operations, but that four Chinese models released since mid-2024 cross the 2 trillion parameter mark. Reuters said the controls also reach methods for coordinating more than 10,000 AI chips.

Registration rule already in force

The part with immediate effect is an interim registration regime. From Oct. 16 until the revised catalogue takes effect, anyone exporting a listed technology must first register it on the commerce ministry's online platform, the notice said. That includes exports already under way but not finished.

Registration is triggered by four situations: transferring ownership of a technology abroad, such as an intellectual property sale; licensing it for use abroad; a change in the owner or actual controller of the technology to a foreign entity; and any other case the ministry specifies. Filers must give the buyer's details, destination, intended use and the expected value of the export.

The third trigger targets deal-making. ChinaTalk said it effectively writes into law the outcome of the Manus case, in which a planned sale of the Chinese-founded agent startup to Meta was reversed after a government review. Reuters also noted that Beijing had ordered Meta to unwind that acquisition. Under the draft, a similar deal could need an export license, and moving a company's domicile abroad would not by itself take its technology out of Chinese jurisdiction, according to ChinaTalk.

What companies and lawyers are asking

Lawyers told Reuters that the coverage tests leave open questions. Han Kun, a Chinese law firm, questioned how model size would be counted for mixture-of-experts systems, which activate only part of their parameters for each task, and said the chip-design provision lacked a minimum performance threshold. Liu Jing of Beijing Dacheng Law Offices said companies want more clarity on how the model tests will work.

Alfredo Montufar-Helu of the consultancy Ankura told Reuters that the structure Beijing chose was "the one the market was already converging on," meaning wide access to services with tighter control of core assets. ChinaTalk noted an incentive for labs to release large models publicly before the rules are final, since published weights fall outside the scope.

The draft adds to other recent export control announcements from Beijing, including ones on port dredgers and tunnel borers, according to ChinaTalk. It also lands while Washington and Beijing work under the terms of a U.S.-China trade truce. The comment period closes Nov. 15. Reuters reported that China is expected to adopt the revised catalogue by the end of 2026, with implementation guidelines to follow.