ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsStructural4 min readDisconfirmAI Editorial

ChinaTalk Analysis of China's MOFCOM AI Export Control Draft: Unprecedented Interim Registration Mechanism and 40-Item List Equally Worthy of Attention

Quinn Ennis, author at ChinaTalk, provides a detailed reading of the draft Announcement No. 48 issued by China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) on October 16, 2026 (with a comment period until November 15). He points out that its unprecedented interim registration arrangement is as important as the 40 new technology items: it forces companies to self-assess whether open-source releases constitute an 'export' before rules are finalized, thereby building the dataset MOFCOM needs for final decisions. The article also analyzes how the 'automatic license' for open-source exemptions preserves China's global soft power for frontier models while setting up revocable checkpoints for their release.

2026-10-21 21:43 GMT+8

On October 16, 2026, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) jointly released the draft of Announcement No. 48, proposing to add 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export, with 23 related to AI. Of these, 14 are listed as prohibited exports (J codes) and 26 as restricted exports requiring licenses (X codes). The public comment period runs until November 15; it is currently not formal law.

After closely reviewing the draft, Quinn Ennis of ChinaTalk proposed a judgment: equally important to the list of 40 items themselves is the interim export registration system established simultaneously in the announcement. This system requires that prior to the official implementation of the revised catalogue, any export involving the above 40 technologies—including those already begun but not yet completed—must be registered in advance. Ennis notes that there is no precedent in the history of catalogue revisions for establishing an interim registration regime synchronously with draft additions.

The logic behind this design is that AI models are not physical items like gallium ingots, making the definition of what constitutes an 'export' technically unclear. By requiring companies to self-assess and register, MOFCOM is effectively collecting a dataset on 'which behaviors companies consider exports' before finalizing the rules, to inform its final licensing decisions. Triggers for registration include: transferring ownership of technology to overseas entities, licensing domestic technology abroad, changing the owner or actual controller of domestic technology to a foreign entity, and other circumstances specified by MOFCOM later.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Regarding frontier models, the draft sets three parallel thresholds; exceeding any one triggers regulation: training compute ≥10^26 floating-point operations (FLOPs), parameter scale ≥2 trillion, or 'capability reaching a globally leading level.' Citing Epoch AI data, Ennis points out that as of October 20, 2026, no released Chinese model has exceeded the FLOPs threshold; only four models exceed the 2-trillion-parameter mark: Kimi K3, ERNIE 5.0, Qwen 3.8 Max, and Qwen3.8-2.4T-A95B. However, the 'globally leading' standard does not specify a ranking body and covers multiple dimensions including 'overall performance, specialized operational capabilities, global call volume, safety performance, and training efficiency,' making the execution boundary highly ambiguous.

Regarding open source, the draft’s Q&A explicitly states that technologies already made public via open weights or literature are not regulated. But 'open weights' does not equal 'open source'—training data, preprocessing pipelines, and training code remain controlled intellectual property. For unreleased new models, the act of open-sourcing itself constitutes a controlled export requiring an 'automatic license.' Ennis argues that the term 'automatic' in the Q&A is qualified as 'basically achieving immediate submission and issuance,' leaving room for delay; during the interim registration phase, any company intending to open-source listed technologies will likely need to submit registration to MOFCOM first, effectively granting regulators the power to view technical specifications before a model is made public.

Another notable arrangement in the draft is codifying the transaction where Meta acquired Manus (a Chinese agent startup that had relocated to Singapore) but was forced to reverse the deal following a security review by China’s National Development and Reform Commission (NDRC). Ennis points out that this means similar transactions will henceforth be governed under export controls rather than investment review, meaning corporate redomiciliation cannot escape Chinese jurisdiction.

In terms of enforcement architecture, export licenses are reviewed by provincial commerce authorities in the exporter’s locality, in conjunction with provincial science and technology departments, rather than being directly approved by MOFCOM. However, citing precedents such as ByteDance’s 2020 application to sell TikTok and CATL’s 2023 application to export battery technology to Ford, Ennis explains that central authorities retain final decision-making power in cases involving the highest national security interests. The comments for Announcement No. 48 have been directed to the Bureau of Industrial Security and Import and Export Controls, rather than the Department of Trade in Services and Commercial Services typically responsible for previous catalogue revisions, suggesting that security agencies have gained greater authority in this round of rulemaking.

Ennis’s overall judgment is that this draft does not mean China intends to close the door on open-source models, but rather that MOFCOM is bringing AI hardware and software within its import-export regulatory jurisdiction. In the context of AI becoming an increasingly complex commodity, establishing professional capacity and enforcement mechanisms early will yield compounding benefits. However, the author also warns that this approach may backfire: if licenses can be revoked arbitrarily, adopters of non-Chinese models in non-WAICO member countries may switch to non-Chinese alternatives.

Boundaries that must be preserved: All the above analysis is based on the draft text and the Q&A accompanying MOFCOM’s release; the rules have not yet taken effect and content may be modified after the comment period. During the 2022–2023 catalogue revision, public consultation and expert deliberation led to the deletion or narrowing of several proposed additions. How the 'globally leading level' standard is defined in the final text, the actual execution speed of 'automatic licenses,' and whether the definition of open source covers releases like Kimi K3 which uses restricted-use licenses, all remain unknown.

Sources

  • chinatalk.media↗
Continue researching →