China carves out open-source AI from its frontier export controls
On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48, a proposal to add 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export, 23 of them in artificial intelligence, alongside an interim pre-export registration requirement already in effect.
The announcement, analyzed in a research note by the Cloud Security Alliance AI Safety Initiative posted on October 23, 2026, marks the first time China's technology-trade regime has reached frontier model weights, agent architectures, and AI-enabled offensive cyber capability with explicit technical thresholds.
China's catalogue has touched AI before. The August 2020 revision added personalized recommendation and AI interaction interface technology, coinciding with the proposed sale of TikTok to a U.S. buyer and reported as likely to complicate that deal. Unlike those earlier additions, Announcement No. 48 proposes specific tests for restriction: 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. Weights, key checkpoints, source code, architecture, and training and inference optimization know-how are restricted if any one test is met. Agents are controlled independently, with planning, multi-agent coordination, tool use, and game-theoretic decision technologies restricted above set benchmarks, and high-performance agents built for critical infrastructure, defense, or core government use prohibited outright.
The announcement contains two instruments with different legal status. The catalogue additions are a draft open for comment until November 15, 2026. The registration requirement applies now, including to exports that have started but are not yet complete. 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. Exporters must decide for themselves whether a technology falls within scope. The registration form asks for recipient details, destination, technology description, end use, and estimated value.
The line between service access and custody determines most enterprise exposure. 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 count, though automatic licensing may apply on a Chinese company's own overseas cloud. Multinationals also face exposure through licensing deals and changes of ownership or control, and analysts read the registration form as potentially capturing intra-group code sharing. Any transfer of ownership, licensing for overseas use, or shift of control to a foreign entity can trigger registration.
The official annex contains a carve-out: any technology that has already entered the public domain through open-source release or academic publication is excluded from management. A MOFCOM spokesperson, quoted in the announcement, stressed support for open source, open collaboration, and resource sharing, with convenient measures such as automatic licensing for technology exports involving open-source AI technologies, Chinese companies building and operating their own computing centers overseas, conducting chip fabrication overseas, utilizing overseas resources for joint research and development, and participating in the formulation of international standards. The spokesperson described the revision as routine and grounded in Article 20 of the Foreign Trade Law, stating it "does not target any specific country, region, or entity" and aims to prevent uncontrolled diffusion while protecting legitimate cooperation.
Hacker News commenters noted the tension. One observed that Chinese models have gained popularity precisely because they are open source, and that restricting powerful models would make them difficult to advocate for and hard for global users to adopt. Another asked what strategic interest China has in supporting open-weight publication, suggesting soft power or commoditizing competition as possible motives. A third noted the rules appear to match previous guidance: restrictions on top-10 models, exemption for open-source collaboration.
The system here is a dual-layer technology-trade regime. The first layer is the catalogue itself, which classifies items as prohibited, restricted, or uncontrolled. The second layer is the licensing infrastructure that implements those classifications, administered by provincial authorities with central intervention in prominent cases. Data moves through this system when an exporter submits registration details to MOFCOM's unified business platform, which then routes to commerce and science authorities for licensing decisions. The design decision that produced this outcome is the explicit exclusion of public-domain technology from the catalogue's scope, paired with automatic licensing for open-source-related exports.
This creates a bifurcated path. A Chinese company training a frontier model faces a choice: release weights openly and fall outside the catalogue, or keep them proprietary and face restriction if the model meets any of the three thresholds. The open-source path eliminates export-control friction entirely for the weights themselves, though related training infrastructure and optimization know-how may still require registration if transferred. The proprietary path preserves competitive exclusivity but subjects overseas deployment to provincial licensing, with the risk of central intervention.
The trade-off is between control and diffusion. Open release sacrifices the state's ability to track where weights travel and how they are used, but gains global adoption, derivative development, and standard-setting influence without bureaucratic overhead. Restriction preserves state oversight but adds friction that competitors without such controls do not face. The spokesperson's emphasis on "common prosperity and safeguarding common security" frames open source as aligned with state interest, but the mechanism rewards Chinese companies that open their weights with automatic regulatory clearance while penalizing closed alternatives with licensing uncertainty.
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. ChinaTalk, cited in the CSA note, estimates that no released Chinese model currently exceeds this threshold. This means the immediate practical effect of the draft catalogue is limited, but the registration requirement applies now to any technology the exporter self-determines to be listed. The self-assessment design places compliance risk on the exporter, with MOFCOM retaining discretion to designate additional cases.
A system that exempts public-domain releases from export controls while restricting proprietary equivalents will see frontier capabilities concentrated in open weights if the cost of openness is lower than the cost of licensing friction. Whether that concentration serves the strategic interests suggested by commenters—commoditizing competitors' proprietary advantages, building dependency on Chinese-trained infrastructure, or establishing de facto standards—depends on whether the exemption persists as models cross the threshold. The open question is what happens when a Chinese model both exceeds \(10^{26}\) FLOPs and is released openly: the weights are excluded from management, but the training know-how and optimization techniques may still require registration if transferred, and the automatic licensing for "open source AI technologies" may or may not cover the full stack needed to replicate or extend the model.