◆ THE EXCHANGE · Policy

China drafts export curbs on frontier AI model weights

The News

On October 16, 2026, China's commerce and science ministries proposed adding 40 technologies, 23 in AI, to the country's technology export-control list. Core technology of large models above set thresholds, including weights and source code, would be restricted, as would some AI chip design and computing-cluster technology. Comments close November 15; registration of listed technologies' exports began the same day.

Chandni Melwani

Chandni Melwani

Founder & Editor

Oct 20, 2026 · 1 MIN READ

The Ministry of Commerce building in central Beijing, a long stone-clad block behind trees and a yellow roadside railing, with a green bus passing under a blue sky with scattered cloud China Photo: N509FZ / Wikimedia Commons (CC BY-SA 4.0)

Know More

  • — Model thresholds (any one): at least 10²⁶ floating-point operations, 2 trillion parameters, or a global top-10 ranking at export.
  • — Prohibited (6 of 23): large models trained on China's 'core data' and high-performance agents for highly sensitive settings, among others.
  • — Clusters: interconnect, fault-tolerance, scheduling and mixed-vendor technology above set thresholds, for clusters over 10,000 AI accelerator cards.
  • — Must register: ownership transfers abroad, licensing for use abroad, a foreign entity becoming owner or controller, and other specified cases.

Offering a restricted Chinese large model abroad through an API or as SaaS would fall outside large-model export controls, a Ministry of Commerce spokesperson said in answers published on October 16. The core question, the spokesperson said, is whether restricted technology is transferred out of China, so localised deployment overseas would count.

On June 12, US export controls required Anthropic to restrict access to two of its models for foreign nationals, inside the US or out, until the controls were lifted on June 30.

The announcement’s annex proposes restricting exports of AI processing-chip design technology that meets any of six tests, including a 14-nanometre or smaller process, 2.5D or 3D advanced packaging, or simply having floating-point or integer computing ability. Memory-chip design for 3D flash and main memory (RAM) would be restricted too, under separate tests.

The ministry calls this revision routine, one of many since 2001, and says it “does not target any specific country, region or entity.”

The Read

The ministry says the controls guard against the risks of technology losing control and improper proliferation while safeguarding normal international technology exchange and cooperation. Chinese labs will probably serve restricted models abroad through APIs or self-built, self-operated overseas compute centres, since the first would fall outside large-model export controls and the second would get facilitation such as automatic licensing.

Room for Disagreement

Open-sourcing, which the annex also facilitates, is a third route: once a lab publishes a model's weights, foreign companies can run it on their own servers. The annex extends the same facilitation to exports for joint research and development using overseas resources.

Related

#Export Controls#China#AI Policy#Ministry of Commerce#Large Models#Model Weights#AI Chips#AI Agents#Open Source

Frequently Asked Questions

What does the pre-export registration require?

From October 16, 2026 until the revised list takes effect, exporters must register on the Ministry of Commerce's online platform before exporting a listed technology, giving details including the recipient, destination, intended use and expected value. Exporters should judge for themselves whether a technology is covered, and exports already begun but not finished are included.

What counts as a world-leading AI model under the proposal?

One ranked in the global top 10 at the time of export, in public assessments of overall performance, specialised capability, global call volume, safety, training efficiency or similar measures by major Chinese and foreign evaluators or large industry platforms, according to the annex to the announcement.

Which AI agent technologies would be covered?

Agent technology above set capability thresholds would be restricted, such as autonomous planning that breaks a natural-language instruction into a chain of at least 20 executable subtasks, architecture supporting 60 or more concurrent collaborating agents, or tool-calling that integrates at least 200 third-party tools, each with further performance conditions. Agents at that level used specially or mainly in highly sensitive settings, such as national critical information infrastructure, the defence and military industry, and core government affairs, would be prohibited.

Sources

Chandni Melwani

Chandni Melwani

Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.

Follow ↗

The Exchange Briefing

Get the five-beat briefing in your inbox.

The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.

Free · No spam · Unsubscribe in one click

More from The Exchange

Infrastructure

UK-founded Yamna signs MoU amid Oman AI data centre plan

UK-founded Yamna and an Omani ministry signed a memorandum of understanding on strategic cooperation in digital infrastructure on October 21, Oman News Agency reported. The UK government said Yamna was set to sign an agreement to build a large-scale AI data centre in Salalah, saying its first phase is expected to exceed £1bn and will create over 100 highly skilled jobs across the UK.

The breakwater of Salalah's port in Oman, with ships and cranes at the quay, seen from a rocky clifftop above the Arabian Sea

Chandni Melwani · Oman · Oct 23, 2026

Capital

Japan to fund a GlobalFoundries-run chip fab in New York

GlobalFoundries will run two new chip fabs at its Malta, New York, site under a US-Japan initiative providing up to $21 billion, with Japan providing the investment. GF is majority-owned by Abu Dhabi's Mubadala, and in a separate agreement the US Commerce Department will receive about 4% of its equity.

Aerial view of GlobalFoundries' Fab 1 chip plant in Dresden, Germany: large white and teal factory buildings surrounded by car parks and fields

Chandni Melwani · US · Oct 21, 2026