China's First Frontier AI Export Controls Target Model Weights

What Announcement No. 48 actually covers

The proposal adds 40 technologies to China's Catalogue of Technologies Prohibited or Restricted from Export: 23 in AI, 5 in biomedicine, and 12 in photovoltaics. Fourteen items are prohibited outright and 26 require a license.

The AI section is the one that matters for anyone building or buying models. It reaches frontier model weights, key checkpoints, source code, architecture, and training and inference optimization know-how. A model triggers restrictions if it meets any one of three tests: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 position on public evaluation rankings, global call volume, or comparable measures at the time of export.

That last test is the slippery one. A model below the compute and parameter thresholds can still be captured if its public rank or call volume puts it in the top 10 when the export happens. ChinaTalk, which analyzed the draft, calls it "by far the most up-for-interpretation" criterion. For compliance teams, that means the restriction is dynamic, not a fixed spec sheet.

Agents get their own treatment, independent of the underlying model. Planning, multi-agent coordination, tool use, and game-theoretic decision technologies are restricted above set benchmarks. High-performance agents built for critical infrastructure, defense, or core government use are prohibited outright.

The registration rule is the part already in force

The catalogue additions are a draft open for comment until November 15, 2026. The interim pre-export registration requirement applies now, including to exports that have started but are not yet complete.

Exporters must register through MOFCOM's unified business platform before exporting any listed technology. Registration covers four triggers: transfer of technology ownership abroad, licensing of domestic technology for overseas use, a change in ownership or control that makes a foreign entity the owner or controller, and other cases MOFCOM designates.

The form asks for recipient details, destination, technology description, end use, and estimated value. Exporters must decide for themselves whether a technology falls within scope—there is no pre-clearance step. Analysts read the form as potentially capturing intra-group code sharing, which could pull multinationals into the regime through licensing deals and ownership changes, not just product exports.

MOFCOM's spokesperson described the revision as routine and grounded in Article 20 of the Foreign Trade Law, stressing that it "does not target any specific country, region, or entity." The catalogue has been revised several times since 2001; the August 2020 revision added recommendation-algorithm technology and coincided with the proposed TikTok sale to a U.S. buyer.

Service access vs. custody: where the line falls

MOFCOM says API, app, web, SaaS, and lightweight on-device use do not count as exporting a large model. Deploying a restricted model locally outside China does.

That distinction determines most enterprise exposure. A U.S. company calling a Chinese model through an API is not exporting it. The same company downloading weights and running them on its own servers outside China is. Automatic licensing may apply on a Chinese company's own overseas cloud, which gives Chinese providers a path to serve foreign customers without a full license review.

The draft also excludes technology that has already entered the public domain through open-source release or academic publication. That carve-out matters for the open-weight ecosystem, though the frontier thresholds are set high enough that most released Chinese models currently fall below them. ChinaTalk estimates that no released Chinese model exceeds the 10^26 FLOP threshold today.

Why this matters beyond China

Frontier AI supply chains now face formal outbound controls from both of the largest AI-producing jurisdictions.

The 10^26 FLOP threshold mirrors the figure used in the 2023 U.S. executive order and the since-rescinded 2025 U.S. AI Diffusion rule. That symmetry is not accidental: both governments are converging on compute as the measurable proxy for frontier capability. The difference is direction. The U.S. controls what goes into China; China is now controlling what comes out. For companies already tracking U.S. export controls, the compliance burden doubles—especially as OpenAI moves to buy Google TPUs in a break from Nvidia, showing how quickly AI supply chains are being redrawn.

The registration requirement is live, the draft catalogue is not final, and the "global leading" test leaves room for interpretation at the moment of export. Legal teams will be reading this annex closely between now and November 15.

What it means for US, UK, and EU companies

If you use a Chinese AI model through an API, app, or SaaS product, Announcement No. 48 does not restrict you directly—those uses are explicitly excluded from the export definition. If you license Chinese technology, acquire a Chinese AI company, or deploy Chinese model weights on your own infrastructure outside China, you may fall under the registration requirement that is already in effect.

The draft catalogue is open for comment until November 15, 2026. Companies with Chinese AI suppliers or partners should review their licensing and deployment arrangements before that date, because the interim registration rule applies now and covers exports already underway.

Announcement No. 48 is China's first formal attempt to control frontier AI exports, and it arrives with a split personality: a draft catalogue that invites comment and a registration rule that does not wait. For anyone with Chinese AI partners, the next two weeks are the time to map exposure, not after the catalogue is finalized.