China's first frontier AI export controls would restrict model weights, agents and chip design

What the announcement does

On October 16, 2026, China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology issued Announcement No. 48 of 2026, the country's first export-control proposal aimed at frontier AI. According to a Cloud Security Alliance (CSA) research note, the announcement carries two instruments with different legal force: a draft revision that would add 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export, and an interim pre-export registration requirement that is already binding.

The draft, open for comment until November 15, 2026, lists 14 prohibited and 26 restricted items: 23 in AI, five in biomedicine and 12 in photovoltaics. Under China's technology-trade framework, prohibited technologies cannot leave the country at all, while restricted ones require a license. ChinaTalk, cited by CSA, notes that licensing is administered by provincial authorities, although central authorities have intervened in prominent cases.

What gets controlled

The AI portion of the annex reaches model weights, key checkpoints, source code, architecture and training and inference optimization know-how. These would be restricted if a model 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. ChinaTalk observes that the 10^26 FLOP figure mirrors the threshold in the 2023 U.S. AI executive order and the since-rescinded 2025 AI Diffusion rule, and estimates that no released Chinese model currently clears it.

Data provenance also matters. Models trained on China's "core data," as defined by the Data Security Law, would be prohibited outright; models trained on "important data" would be restricted, with an exemption for trace use (up to 0.1 percent) of desensitized data that neither materially improves capability nor reappears in outputs. Technology already in the public domain through open-source release or academic publication is excluded.

Hardware is in scope too. The draft restricts AI and memory chip design at process nodes of 14 nanometers or below, along with compute-cluster technologies for deployments above 10,000 accelerators with interconnects of at least 800 GB/s.

Agents and offensive cyber tools

Agents are controlled independently of the models beneath them. Planning and task decomposition is restricted above benchmarks such as 20-step subtask chains with 80 percent decomposition accuracy and 90 percent subtask success; multi-agent collaboration at 60 or more concurrent agents; and tool calling at 200 or more integrated third-party tools with 90 percent accuracy thresholds. High-performance agents designed mainly for critical infrastructure, defense or core government use would be prohibited outright.

Cyber capability gets its own categories. AI malicious-code generation, adaptive evasion and variant generation would be prohibited where usable against high-sensitivity systems, while AI-assisted vulnerability discovery would be restricted under six tests, including scoring 75 percent on authoritative benchmarks or autonomously chaining two or more low- or medium-severity flaws into a working exploit.

The rule that already applies

The registration requirement has been in force since the announcement date and covers exports that have started but are not yet complete. Exporters must register through MOFCOM's unified business platform before shipping any listed technology. Four triggers apply: transfer of technology ownership abroad, licensing domestic technology for overseas use, a change of ownership or control that makes a foreign entity the owner or controller, and any other case MOFCOM designates. The form asks for recipient details, destination, technology description, end use and estimated value, and exporters must decide for themselves whether a technology falls within scope.

Where the line falls for enterprises

MOFCOM draws the boundary at custody rather than access. API, app, web, SaaS and lightweight on-device use do not count as exporting a large model, while deploying a restricted model locally outside China does — although automatic licensing may apply on a Chinese company's own overseas cloud. Analysts cited by CSA read the registration form as potentially capturing intra-group code sharing, so multinationals face exposure through licensing deals and ownership or control changes, not just product exports.

A MOFCOM 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's last brush with AI came in 2020, when a revision added recommendation-algorithm technology alongside the proposed TikTok sale to a U.S. buyer, and was reported as likely to complicate that deal.

Why it matters

Frontier AI supply chains now sit under formal outbound controls from both of the largest AI-producing jurisdictions. CSA's earlier research documented the first U.S. use of export-control authority to suspend a broadly deployed commercial frontier model; Announcement No. 48 adds the Chinese counterpart on the export side. For companies building on Chinese models, chips or engineering teams, the exposure is immediate — the registration obligation applies today, the thresholds are unusually concrete, and the open-source carve-out will decide how much of the Chinese AI ecosystem remains reachable from abroad.