Export controls · China
Beijing locks down its AI… but not its APIs
On 16 October, China proposed banning or licensing the export of 23 AI technologies, from the weights of giant models to chip designs. Beijing calls the text “precise and limited” and says it “does not target any country”. We read the annex.
Translated from the French original. In case of discrepancy, the French version prevails.
Friday 16 October, 4 p.m. in Beijing. The Ministry of Commerce posts an official notice and an annex in WPS format. Inside: 40 technologies that China wants to ban or subject to licensing for export, 23 of them in artificial intelligence.
The same day, a new obligation takes effect: every exporter must declare transfers before carrying them out. Including those already under way.
The consultation runs until 15 November. The text may still change. But its logic is already set down in black and white.
The vault: weights, code and designs
The target is the “core” of large models: weights, checkpoints, source code, architecture, training and inference methods. A single criterion is enough: at least 10²⁶ operations of training compute, at least 2 trillion parameters, or a place in the global top 10 of public benchmarks at the time of export.
The same logic applies to hardware. It is not the chips themselves that are covered, but their design: architectures, instruction sets, intellectual-property blocks. And the know-how behind clusters of more than 10,000 accelerator cards.
« A clear, precise and limited scope »
The door left wide open: the API
The spokesperson is explicit. Offering a restricted model via API, app, web page, SaaS or a lightweight on-device version does not fall under the controls. What matters is that the technology does not leave the country.
Models already released as open source or published in scientific papers are excluded. Open-sourcing a model, building one's own data centre abroad or having one's chips fabricated there will go through an “automatic” licence. The official formula: “declared, then exported straight away”.
According to Reuters, this dividing line would allow DeepSeek or Alibaba to keep expanding internationally, while giving Beijing oversight of sensitive transfers.
Precise and limited, really?
For models, the thresholds are quantified. For chips, it is another story. An AI chip design is covered if it meets just one of six criteria. One of them: “having floating-point or integer computing capability”.
To its credit, the text only covers chips dedicated to AI processing. But on a plain reading, this criterion taken on its own appears to cover most of that category. The ministry does not say how it will interpret it.
Another restricted item, with no official comment: voice-based identity recognition optimised for minority languages or regional dialects, with accuracy above 90%.
Counter-reading
What the official narrative doesn't tell you
Beijing presents this as “routine” work that “does not target any country”. Several major headlines talk about “chips”. The text says something else, or more.
- The threshold is not neutral: 10²⁶ operations is the threshold the Biden administration set in January 2025 to control the weights of closed models. Washington dropped it on 13 May 2025, two days before the compliance date.
- These are not chips: AP refers to technologies “covering AI, chips and solar”, and the FT headline is about “chips”. The annex covers designs; the spokesperson specifies that manufactured products are not affected.
- The cyber angle flies under the radar: the ban covers the autonomous generation and mutation of malicious code by AI, with quantified thresholds (attack success rate ≥ 60% against multi-engine detectors). The text does not say whether such capabilities already exist.
- Europe, absent from the debate: its AI Act sets the “systemic risk” threshold at 10²⁵, but to regulate models, not their export.
We analyse texts and headlines, not people.
- How many companies have already declared a transfer since 16 October?
- Who will draw up the “global top 10” of models, and as of what date?
- Will the “automatic” licences be written into the final text?
Beijing keeps the engine and rents out the car. The question is who will be at the wheel on 15 November.
Fact-check
The fact-check, point by point
Each official or media claim, checked against the text of Announcement No. 48 and its Annex 1. Click on a point to read the analysis.
What the official version says
China's Ministry of Commerce and Ministry of Science and Technology propose adding 40 technologies (23 in AI, 5 in biomedicine, 12 in photovoltaics) to the catalogue of technologies prohibited (14) or restricted (26) from export. Consultation until 15 November 2026; prior declaration mandatory from 16 October.
According to the spokesperson, this is “normal” work in managing technology trade, carried out since 2001, which “does not target any country, region or entity”. The scope is said to be “clear, precise and limited”, manufactured products are not affected, and “the vast majority” of AI technologies will remain freely exportable. Access via API, app, web or SaaS is excluded from the controls.
Claims checked
Needs context “Normal work in managing technology trade […] that does not target any particular country, region or entity” (MOFCOM spokesperson, 16/10/2026)
The text indeed names no country, and according to the ministry the catalogue has been revised since 2001.
But the spokesperson himself justifies the measure by noting that “major economies” already control these technologies, and the threshold chosen for models (10²⁶ operations) is identical to the one in the January 2025 US rule. Reuters reads the draft in the light of the rivalry with Washington.[3][4][5]
Needs context “A clear, precise and limited scope” (MOFCOM spokesperson)
For models, the thresholds are quantified (10²⁶ operations, 2 trillion parameters, global top 10). The same goes for compute clusters (more than 10,000 cards, interconnects ≥ 800 GB/s).
For AI chip design, one criterion out of six is enough, including “floating-point or integer computing capability” or a process node of 14 nm or below. The scope is limited to AI processing chips, but this criterion appears broad on a plain reading. The “global top 10” criterion depends on public rankings that are not named.[2][3]
Partly verified “The vast majority of AI technologies can still be exported freely” (MOFCOM spokesperson)
APIs, apps, the web, SaaS and lightweight on-device deployments are excluded, as are open source and publications; several cases qualify for an automatic licence.
But the prior-declaration requirement applies from 16 October, including to exports already under way, and the final text, expected after 15 November, is not yet known.[1][3]
Incomplete China plans controls on “chips” (FT headline; AP refers to technologies “covering AI, chips and solar”)
The annex covers the design (architecture, instruction sets, intellectual-property blocks, related software) of AI and memory chips, not the chips themselves. The spokesperson specifies that products manufactured using the listed technologies are not affected.
The articles may clarify this in the body text; the FT is behind a paywall and we have only read its headline and standfirst.[2][3][6][7]
Accurate The 10²⁶-operation threshold has already been used by the United States
The “Framework for Artificial Intelligence Diffusion” rule of 15 January 2025 controlled the weights of closed models trained with more than 10²⁶ operations. On 13 May 2025, the Department of Commerce announced its rescission and asked that it not be enforced, ahead of the 15 May compliance date.[5][8]
Sourced counter-arguments
The “floating-point or integer computing” criterion only applies to designs of “AI processing chips”: the scope is therefore already narrowed by the category itself.[2]
Excluding APIs, SaaS and open source, and providing automatic licences, shows that the text seeks to protect assets without slowing the commercial expansion of Chinese companies: this is the ministry's argument, which Reuters also notes.[3][4]
Controlling the spread of AI tools capable of generating malware or finding vulnerabilities addresses a concern shared well beyond China.[3][4]
Questions that remain open
- How many of the 23 AI technologies will remain in the final text?
- How will the “global top 10” of models be determined?
- Will the European Union and the United States respond with their own rules on model weights?
In short
The figures in the announcement are accurate and can be checked in the annex. The official line on a text that is “precise, limited” and “targets no one” needs context: chip design is covered broadly, and the model threshold mirrors the one Washington abandoned. Headlines about “chips” oversimplify: it is the designs, not the products.
Sources
- MOFCOM–MOST Announcement No. 48 (2026): revision of the catalogue of technologies prohibited or restricted from export, consultation and prior declaration (in Chinese) — Ministry of Commerce of the PRC · accessed 18 October 2026
- Annex 1: proposed AI, biomedical and photovoltaic technologies (draft for consultation, .wps file attached to Announcement No. 48, in Chinese) — Ministry of Commerce of the PRC · accessed 18 October 2026
- Ministry of Commerce spokesperson answers questions on the catalogue revision (16/10/2026, in Chinese) — Ministry of Commerce of the PRC · accessed 18 October 2026
- Beijing seeks tighter grip on AI exports amid US rivalry — Eduardo Baptista, Reuters, 16/10/2026 (syndicated) — Reuters / WTVB · accessed 18 October 2026
- Framework for Artificial Intelligence Diffusion, interim final rule of 15/01/2025 (90 FR 4544) — Federal Register / Bureau of Industry and Security · accessed 18 October 2026
- China moves to curb tech exports including for AI and solar, 16/10/2026 — Associated Press · accessed 18 October 2026
- China unveils draft export controls on AI technologies and chips, 16/10/2026 (paywalled; headline and standfirst only) — Financial Times · accessed 18 October 2026
- Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule, 13/05/2025 — Bureau of Industry and Security · accessed 18 October 2026
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act), Article 51(2) — EUR-Lex / European Union · accessed 18 October 2026
AI-assisted content: this article was written with the help of artificial intelligence from public sources, then reviewed and approved by a human editor before publication. Spotted an error, want to respond? Exercise your right of reply · our method.