China Drafts AI Export Curbs: Models, Agents, Registration
China has drafted export controls for 23 AI technologies, including large models and AI agents, and has already started requiring registration of pending deals.
Key takeaways
- On 16 October 2026 China's Ministry of Commerce and Ministry of Science and Technology published a draft revision of the catalogue of technologies banned or restricted from export: 40 items, of which 23 are AI [1].
- A temporary registration rule applies from 16 October to deals such as IP transfers, overseas licences and changes of control to foreign owners. Public comments close on 15 November [1].
- According to the ministry, open-source and published technology is outside the controls, and API access, apps and SaaS are not covered [1][2].
- This is a draft. The reports we read carry no statements from Chinese AI companies [3][4].
What happened
On Friday 16 October the two ministries issued Announcement No. 48, opening a consultation on adding AI, biomedicine and solar technologies to China's export catalogue [1]. The 40 items split into 23 AI, 12 solar and 5 biomedicine technologies [1]. According to the ministry's explainer, carried by Xinhua, 14 items would be prohibited and 26 restricted; Caixin reports the same split [2][5].
For AI, the ministry describes the scope as large models, agents, computing clusters and information processing [1]. It gives two examples: AI-driven malicious code generation would be prohibited, and AI-assisted vulnerability discovery restricted [2]. Technology already public through open-source release or academic papers is not covered [1].
Specific thresholds come from news reports, not from the pages we read. Bloomberg and the Associated Press (AP) report that models with at least 2 trillion parameters, or ranked in the global top 10, could be restricted, and Bloomberg lists weights, checkpoints, source code and architecture among covered items [3][4]. AP notes that Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max exceed 2 trillion parameters, and that the draft names no model [3]. We have not read the draft's attachment with the full scope.
What stays open, and what does not
The ministry says API access, app or web interaction, SaaS and lightweight on-device deployment are not covered by the large-model controls [2]. Running a restricted model locally abroad is covered, though it may receive an automatic licence if it runs on a controllable overseas cloud [2]. The ministry also lists open-sourcing, overseas compute centres run by Chinese firms and joint research abroad as eligible for automatic licensing [2].
One report points the other way. The Financial Times, as relayed by Startup Fortune, says regulators are considering limits on training data sent overseas and on whether foreigners keep downloading weights, after closed-door meetings with Alibaba, ByteDance and Zhipu. That is reported, not confirmed, and it sits uneasily with the open-source exemption [6].
Why it matters
Our analysis: until now, China's strongest model makers could ship weights and deployments abroad with few formal steps. A registration duty and a licence regime add friction and give Beijing visibility into who receives frontier-scale Chinese technology.
Strengths of the approach, as officials frame it: it aims at misuse and keeps hosted access open, so most overseas users of Chinese AI see no change [2]. Limits: the thresholds and definitions are unclear, "restricted" versus "prohibited" is untested in practice, and the effect on self-hosted deployments abroad is the open question [3][6].
Our analysis, not a report: overseas buyers who self-host a large Chinese model are the most exposed, while API customers are the least. The ministry says the move is not aimed at any country [2][4]. Omdia's Lian Jye Su reads it as a response to Western sovereignty rules, and Natixis's Gary Ng says technology access is fragmenting [3].
The global picture
China: the draft follows the Manus case, when Beijing ordered a Meta deal unwound; see our 10 October story.
Taiwan: Economic Affairs Minister Kung Ming-hsin said AI servers and advanced chips will be added to the controlled list, and that shipping them to China would become a crime under the Foreign Trade Act, Bloomberg reported on 16 October. It is a proposal with no date, and penalties are not specified in the reports we read [7][8].
China and Taiwan, chips: Bloomberg reported on 15 October that ByteDance is close to buying custom AI chips from Taiwan's MediaTek. Reuters could not verify it, and neither company commented in the report we read [9].
Gulf: the physical side of sovereign compute. Al Jazeera reported on 13 October that Iran's security chief Mohsen Rezaei warned that cables, data centres and AI facilities could be attacked if war with the US resumes. It also reports that Amazon Web Services said drones struck two UAE facilities in March [10]. This is a single-outlet account.
US, Europe, Japan, Korea, India: no matching new measure in our checks today.
What to watch next
- 15 November 2026: the consultation closes [1].
- Final text: the 2-trillion-parameter and top-10 thresholds, and the treatment of public weights.
- Companies: any statement from Alibaba, Moonshot, ByteDance or Zhipu.
- Taiwan: the Foreign Trade Act amendment and its penalties [7].
Our take
Export control used to be one-directional. China is now writing its own list, and the details will decide whether this is a narrow safety rule or a broad brake on open models. Until the thresholds are public, treat every reported number as provisional.
FAQ
Does this ban Chinese AI models abroad? No. It is a draft, and the ministry says API access, apps and SaaS are not covered [1][2].
Are open-source models affected? The ministry says technology already public through open-source release is outside the controls [1]. One report says weights could still be limited [6].
Is anything already in force? A temporary registration rule applies from 16 October to listed technologies in pending transfers [1].
Which models are named? None. AP notes two models above the size threshold [3].
AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. This is analysis, not legal or financial advice.
Sources
- Announcement No. 48 of 2026 on revising the Catalogue of Technologies Prohibited and Restricted from Export (Chinese), Ministry of Commerce and Ministry of Science and Technology, 16 October 2026. Official Chinese source.
- Ministry of Commerce explains the revision (Chinese), Xinhua, 17 October 2026. State media, used only for the ministry's position.
- China moves to curb tech exports, including AI, citing national security risks, Associated Press, 16 October 2026.
- China proposes AI export curbs to manage security, Bloomberg via Taipei Times, 17 October 2026.
- China Proposes Export Curbs on AI and Solar Technology, Caixin Global, 17 October 2026 (paywalled; top-line figures only).
- China moves to restrict exports of its own AI models, chips and agents, Startup Fortune, citing the Financial Times, October 2026.
- Taiwan Will Make It a Crime to Smuggle AI Chips Into China, Startup Fortune, citing Bloomberg, 16 October 2026.
- Taiwan to make AI chip smuggling to China a crime, APAC Signal, October 2026.
- ByteDance nears chip supply deal with MediaTek for data centres, Bloomberg News reports, Reuters via The Star, 15 October 2026.
- Iran threatens Gulf data centres: How exposed is the region's internet?, Al Jazeera, 13 October 2026.
AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 17 October 2026.