China's Draft Export Rules Would Cover 23 AI Technologies, Exempting Cloud and App Access to Models

Since China's Ministry of Commerce and Ministry of Science and Technology released draft revisions to the Catalogue of Technologies Prohibited and Restricted from Export on Oct. 16, the details have come into focus. So have the questions from the companies that would live under them.

The draft covers 40 technologies across three sectors. Twenty-three are AI, 12 are photovoltaics and five are biopharma. Fourteen would be banned from export outright. The other 26 would need a license.

What the AI rules cover

The AI items include large language models, AI agents, intelligent computing clusters and AI-chip design. Banned items would include AI-enabled malware generation and mutation tools, high-performance agents built for highly sensitive uses, and large models trained on China's state "core data" as defined under laws such as the Data Security Law.

Licensing would apply to core large-model technology, advanced agents, chip-design technology, AI-assisted vulnerability discovery and large-scale computing-cluster technology.

Three tests decide which models are covered. One is training compute of at least 10²⁶ calculations. Another is a model with at least 2 trillion parameters. The third is a ranking among the world's top 10 on measures of performance, usage or efficiency.

Bloomberg cited Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max as examples of models above the parameter line. For those, the restricted items would include weights, checkpoints, source code and architecture.

Access stays open. Ownership does not.

The central distinction is between using Chinese AI and owning it. According to the commerce ministry's interpretation, as described by law firm White & Case, serving AI through APIs, mobile or web apps, SaaS products or lightweight local edge deployment generally falls outside the controls. That holds as long as the restricted technology itself does not leave China.

Transfers of weights, training data or the know-how to run or rebuild a system independently would need government approval. So would changes in ownership or control and some overseas deployments.

Technology already disclosed through open-source release or academic publication is not affected, according to White & Case's reading of the draft. The draft also signals that overseas data-center construction, overseas chip tape-out, cross-border joint R&D and participation in international standard-setting could qualify for automatic licensing or other facilitation.

Alfredo Montufar-Helu, Beijing-based managing director at consultancy Ankura, said the structure fits what the market was already doing. "The structure the leadership prefers, with IP and R&D kept onshore and operations and access sold abroad, is the one the market was already converging on," he said.

Where the labs want answers

DeepSeek, Moonshot and Z.AI sell models to overseas customers through websites and cloud platforms. Z.AI's models are also available on U.S. cloud providers such as Amazon Web Services, according to Bloomberg.

Some Chinese developers also use computing facilities abroad to reach advanced chips. Under U.S. restrictions, Nvidia's most advanced processors cannot be exported to China.

A person briefed on the matter, speaking anonymously because the discussions were private, said researchers at major Chinese labs want to know whether they can train or run their most advanced models on foreign-owned servers overseas. In the worst case, that person said, developers could have to maintain separate domestic and overseas versions of the same downloadable model.

Liu Jing, a senior partner at Beijing Dacheng Law Offices who joined consultations with policymakers, said companies want clearer criteria for deciding which models are covered. The ranking test could sweep in smaller models that excel at specialized tasks, but the draft does not say which rankings count.

Law firm Han Kun questioned how regulators would measure models that use only part of their components for each task.

Beijing's stated rationale

The commerce ministry rejects the idea that this is a trade weapon. Its spokesperson said the controls aim to ensure "the safety and controllability" of AI. The spokesperson cited the risk of misuse against critical infrastructure, defense industries and core government systems by non-state actors or malicious entities.

The ministry has also described the exercise as a targeted, dynamic update to its export-control framework, not a measure aimed at any particular country. In its statement, it said the management of these technologies is meant to "solidify the security baseline, prevent misuse and abuse, ensure that AI remains under human control, and promote international governance of AI."

The ministry did not respond to a request for comment on the industry's questions.

Timing and the market backdrop

A temporary registration requirement for certain transfers took effect on the day of the announcement. A source familiar with the discussions, who asked not to be named because the schedule is not public, said China is expected to adopt the proposals by the end of the year and issue implementation guidance afterward.

The draft lands while Chinese AI and semiconductor stocks are in a sharp sell-off. That adds regulatory uncertainty to a sector already under U.S. chip restrictions, content controls and a licensing regime.

In practice, Beijing wants foreign customers paying for Chinese AI, and it wants the weights, chip designs and engineering knowledge to stay inside China. The U.S. has been tightening the same screws from the other side through chip controls.

Public consultation closes Nov. 15. Whether the final text lets Chinese labs keep training on foreign-owned servers abroad, and which rankings define a top-10 model, are the details still unanswered.