Cowry Tech News
Policy

China’s draft AI export controls: what the proposed catalogue would cover, and what it leaves out

China has published a draft that would add 23 AI technologies to its export ban and restriction list, from frontier-scale model weights to AI chip design and 10,000-accelerator cluster software. What the draft covers, what it leaves out, and the pre-export registration already in force. Comments close 15 November.

By Kwesi Addo · 18 Oct 2026 · 14 min

Rows of shipping containers and red gantry cranes at Yangshan Deepwater Port in Shanghai.
Photo: Reb42 / Wikimedia Commons, CC BY 3.0

On Friday, China’s Ministry of Commerce and Ministry of Science and Technology published a draft revision of the Catalogue of Technologies Prohibited or Restricted from Export. That catalogue decides which Chinese technologies can leave the country freely, which need a licence, and which cannot be exported at all. The draft would add 40 technologies in three fields: artificial intelligence, biomedicine and solar power. Twenty-three of them are AI technologies.

Nothing in the catalogue changes yet (CTN’s Brief item has the short version). The ministries are taking public comments on the draft until 15 November. One part of the notice did take effect on Friday: anyone planning to export a technology on the draft list now has to register it with the commerce ministry before the export happens.

This piece explains the instrument, what the AI section of the draft list says, what the ministries have excluded, and what the temporary registration means for companies that build, license or deploy AI from China. It draws on three primary documents, all read in Chinese: Announcement No. 48 of 2026 from the Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology; Attachment 1, the draft technology list with its technical thresholds; and a MOFCOM spokesperson’s Q&A published the same day. Translations are CTN’s own. Reuters reported the draft independently on Friday, and its account matches the documents on the points below where it is cited. CTN has not asked the ministries for comment.

What the instrument is

China controls technology exports through two separate channels.

The first is the Export Control Law, which governs controlled items, including dual-use goods, and the lists attached to it. The second, used here, is the technology catalogue issued under the Foreign Trade Law and the Regulations on the Administration of Technology Import and Export. This catalogue covers technology itself: know-how, designs, source code, process recipes and the rights to use them.

Technologies in the catalogue fall into two classes. Prohibited technologies cannot be exported. Restricted technologies can be exported only with a licence from the commerce ministry. Anything outside the catalogue remains freely exportable.

According to the Q&A, Article 20 of the Foreign Trade Law gives the commerce ministry, working with other State Council departments, the job of drawing up and adjusting the catalogue. The spokesperson said the two ministries have revised it many times since 2001, adding and removing entries, and called this round routine work that “does not target any specific country, region or entity.”

The spokesperson also said products made with a listed technology fall outside the catalogue. It controls the technology, so chips, hosted model services and solar panels built with that technology are not the target.

The headline numbers

According to the announcement and the Q&A, the draft would:

  • add 40 technologies in total, with 23 in AI, 5 in biomedicine and 12 in solar power;
  • prohibit 14 of the 40 and restrict 26; and
  • group the AI entries under large models, agents, computing clusters and information processing.

The ministry gave AI-driven malicious code generation as one prohibited example and AI-assisted vulnerability discovery as one restricted example. Most entries carry numeric thresholds. The ministry describes the scope as “clear, precise and limited” and says “the vast majority of AI technologies can still be exported freely” once the measures take effect. Reuters reported the same 23-technology count and the ministry’s statement that most AI technology would stay freely exportable.

What the draft list would prohibit

The prohibited part of Attachment 1 has four AI-related entries:

  1. Deepfake detection built mainly to protect China’s cross-border data flows, with detection accuracy and recall of at least 95% on synthetic video, audio, images or multimodal content.
  2. AI-driven malicious code: technology that autonomously writes or mutates malware, evades sandboxes or mass-produces variants that slip past detection, where it could be used against critical information infrastructure, defence or “core government” systems. The entry sets success-rate thresholds, such as at least an 85% attack success rate against a single mainstream antivirus product.
  3. Large models trained on China’s “core data” and then used to provide services. Core data is the highest tier in China’s data classification under the Data Security Law.
  4. High-performance agents built specifically or mainly for highly sensitive uses, again naming critical infrastructure, defence and core government work. The restricted agent thresholds below define “high-performance.”

What the draft list would restrict

Builders should look hardest at the restricted part. Entries that would need an export licence include the following.

Large-model core technology

Most coverage has focused on this entry. It covers “the core technology of large models, including weight parameters, key checkpoints, underlying source code and architecture design, as well as training optimisation and inference acceleration techniques,” where the model meets any one of three tests:

  • training compute of at least 1026 floating-point operations;
  • at least 2 trillion parameters (written 20,000亿 in the Chinese text); or
  • capability at a “globally leading level.” The draft defines this as ranking in the global top 10, at the time of export, in public evaluations by mainstream Chinese or foreign evaluators and large industry platforms. Those evaluations can cover overall performance, specialist capability, global call volume, safety and training efficiency.

Models trained on important data

Important data is the tier below core data. A model trained on 0.1% or less of important data is out of scope if that data cannot materially improve the model’s capability in that field, has been desensitised, and will not reappear in outputs. Data that can be lawfully obtained in public is also excluded.

Agents

The draft sets thresholds for four kinds of agent technology:

  • planning that breaks a natural-language instruction into a chain of 20 or more executable sub-tasks, with at least 80% decomposition accuracy and 90% average sub-task success;
  • multi-agent systems that run 60 or more agents at once;
  • tool use with 200 or more integrated third-party tools and at least 90% tool-selection accuracy; and
  • game-style decision systems that beat 90% of humans on standard benchmarks.

AI chip and memory chip design

This entry covers the design of AI processing chips, including architecture, in-house instruction sets, intelligent-computing IP cores and the supporting software stack. It applies where a design has any one of six listed features. They include needing a manufacturing process of 14 nanometres or below, 2.5D or 3D advanced packaging, high-speed interconnect, or a non-von Neumann design such as compute-in-memory. Another listed feature is “having floating-point or integer computing capability.” Memory chip design is listed too, again with any one feature enough. For 3D NAND flash, the triggers are storage density of 8 gigabits per square millimetre or more, I/O speed of 1.5 gigabits per second or more, or the use of bonding technology. For DRAM, they are designs that need an 18-nanometre process or smaller, a vertical-channel transistor architecture, bonding technology, or multi-layer stacking.

Computing-cluster technology

The draft defines a “large-scale intelligent computing cluster” as one with more than 10,000 AI accelerator cards. For those clusters it lists chip-to-chip or node-to-node interconnect at 800 GB/s bidirectional or faster; fault tolerance for distributed training, such as recovering from a checkpoint within 120 seconds; communication scheduling, such as 98% linear scaling; and software that schedules and runs inference across AI chips from two or more vendors.

Other information-processing entries

The restricted list also includes AI-assisted vulnerability discovery, with thresholds such as autonomously chaining two or more low- or medium-severity bugs into a working attack with proof-of-concept code. It adds deepfake detection at 98% accuracy and above, image and video steganography, voiceprint recognition, and translation technology for Chinese and minority languages.

What the draft leaves out

The ministries were as specific about exclusions as about the list. For most companies the exclusions will matter more.

Selling access to a model does not count as exporting it. The Q&A says that using a restricted large model to provide API calls, app or web interaction, SaaS, or “lightweight local on-device deployment” to users abroad “does not fall within the scope of large-model technology export management.” Reuters reported the same exclusion. Under the draft, a Chinese lab could keep selling access to a restricted model worldwide without an export licence.

Running the model abroad does count. Overseas “localised deployment” of a restricted AI technology is a technology export under the draft. The ministry adds that where the deployment sits on the company’s own overseas cloud, which it calls a situation where “risk is relatively controllable,” automatic licensing can apply.

Technology already in the public domain is out of scope. AI technology that has entered the public domain through open-source release or academic publication would not be covered, so weights and code already published openly sit outside the list.

Some outbound activity gets a fast track. The Q&A lists cases where automatic licensing applies, which it says makes approval close to immediate: open-sourcing AI technology; Chinese companies building and running their own compute centres abroad; sending chip designs abroad for tape-out and manufacturing; joint research using overseas resources; and taking part in international standards work. Reuters reported the simplified licensing for open-source exports and overseas computing centres. The Q&A also promises “more facilitation and support” for technology exports among members of the World Artificial Intelligence Cooperation Organization (WAICO), the Shanghai-based intergovernmental body 29 countries agreed to set up in July, and support for AI capacity building in Global South countries.

Open-sourcing and public-domain status are handled differently. Technology that is already public is excluded. A new open release of something on the draft list is described as getting an automatic licence, so it still passes through the licensing system, even if quickly.

What exporters must do now

The temporary registration is the only part of the notice already in force. It applies to the draft list, before any final rules exist.

Section 3 of the announcement says that from the date of publication until the revised catalogue takes effect, technology exporters must register on MOFCOM’s technology trade management platform before exporting any technology on the draft list. Exports that have started but are not finished must also register. The Q&A says exporters have to judge for themselves whether a technology falls within scope. Reuters reported that registration of certain overseas transfers applies from Friday.

Four kinds of transaction need registering:

  1. transferring ownership of a technology abroad, such as assigning intellectual property;
  2. licensing a domestic technology for use abroad;
  3. a change of owner or actual controller of a domestic technology to a foreign entity; and
  4. other situations the commerce ministry specifies by notice.

The registration form asks for the exporter’s and recipient’s details, the destination country or region, the technology and its intended use, and the expected value of the export. Registration is not a licence, and the notice does not say registration can be refused. The Q&A says its purpose is to understand export demand and “guard against potential export risks” while the draft catalogue is being finalised.

The ministry’s stated reasons

The Q&A gives two lines of argument for the AI entries.

The first is safety. The spokesperson said technologies such as autonomous malicious code generation “can be used to attack critical information infrastructure, defence and core government systems” and could cause harm “if obtained by non-state actors or malicious entities.” The stated goal is to ensure that AI “always remains under human control.”

The second is openness. The ministry presents the open-source exclusion and the automatic-licence fast track as support for an “open-source innovation ecosystem.” It says “major economies” already control exports of AI, biomedical and solar technology to varying degrees, and that China is following “common international practice.”

On consultation, the spokesperson said the ministries assessed the scope in advance with research institutes, industry bodies and “major enterprises,” and will drop technologies from the list as they become outdated.

Analysis: what to watch

This section is CTN’s reading of the documents, not a statement by the ministries.

The top-10 test. Compute and parameter counts are fixed numbers. A top-10 ranking changes whenever a new model ships, so under the draft a model could move into or out of scope without any change to the model itself. The draft names the kinds of evaluators and measures that count but does not name specific leaderboards, or say how a company would show where its model ranked at the time of export.

A familiar number. The 1026 compute figure is not new to AI policy. The US executive order on AI signed in October 2023 (Executive Order 14110, 88 FR 75191) used the same number, more than 1026 “integer or floating-point operations”, as the training threshold that triggered reporting duties for model developers. That order was rescinded in January 2025 (Executive Order 14148, 90 FR 8237). The match is context only. The Chinese documents do not refer to the US order, and the two instruments do different jobs: one was a reporting trigger, the other would be an export-licensing line.

Chip-design breadth. As written, “having floating-point or integer computing capability” is one of six features, and any one of them triggers the entry. So is a design that “can be applied to specific scenarios such as autonomous driving, computer vision and smart voice.” Read literally, either one would reach a very wide range of AI chip designs. Whether it is meant as a standalone trigger or will be narrowed is the kind of question the comment period is there to settle.

Acquisitions. The third registration category, a change of owner or actual controller to a foreign entity, appears to bring a foreign takeover of a Chinese AI company that holds a listed technology inside the registration requirement, even if no code moves across a border.

Effect outside China. For companies and governments that use Chinese AI, the immediate effect is narrow. Access through an API, a chat app or a SaaS product is explicitly outside the draft controls. That holds for any business, in West Africa or elsewhere, that builds on a Chinese model through hosted access. The arrangements that would face registration now, and possibly licensing later, are the ones that move the technology itself: licensing frontier-scale weights to run on premises, hiring a Chinese vendor to build a large training cluster, or buying a Chinese AI company outright.

What is still open

The documents leave several questions unanswered:

  • The final text. This is a consultation draft. Scope, thresholds and facilitation measures can change after 15 November, and the documents give no date for a final catalogue.
  • Registration outcomes. The notice does not say what happens to a registered export the ministry views as risky before the catalogue is final.
  • Licensing timelines. Nothing says how long licences for restricted AI technologies will take once the revised catalogue applies, outside the automatic-licence cases.

Comments on the draft can go to the commerce ministry’s Security and Control Bureau by email, fax or letter until 15 November. Submissions made on behalf of an organisation must carry its official seal.

CTN will report on the final catalogue when it is published.

Sources