Beijing has spent three years watching the United States restrict what chips it can buy. On October 16, 2026, it moved to restrict what AI it can sell. China's Ministry of Commerce and Ministry of Science and Technology jointly published Announcement No. 48 of 2026, a draft proposal adding 40 technologies to the national catalogue of items prohibited or restricted from export. Twenty-three of those 40 are artificial intelligence technologies.

The draft is open for public comment until November 15. It is not yet law. But a temporary pre-export registration requirement took effect the same day the announcement was published — meaning any Chinese company that wants to transfer ownership, license rights, or control of a listed AI technology to a foreign entity must register on the commerce ministry's platform before doing so, right now, while the consultation is still running.

What the 23 AI Technologies Actually Cover

The 23 items span large-scale models, intelligent agents, computing clusters, AI chip design, voice recognition, high-performance translation, deepfake detection, and hacking-related AI systems. The ministry framed the list as targeting "information processing" AI — technologies capable of being repurposed for attacks on critical infrastructure, defence industries, or core government systems by non-state actors or malicious entities.

The most consequential category on the list is the core technology of large models: weights, key checkpoints, source code, architecture designs, and training and inference optimisation techniques. Beijing has set three separate thresholds, any one of which triggers the restricted classification. A model qualifies if it was trained using 10²⁶ floating point operations or more. It also qualifies if it has 2 trillion parameters or more. The third route is ranking in the top 10 globally on mainstream public benchmarks or major industry platforms at the time of export.

That 2-trillion-parameter bar is the number the whole industry will be calculating against. Microsoft's MAI-Base-2, which completed pre-training on October 15, carries 2.8 trillion total parameters — a figure that would clear the threshold if it were a Chinese model seeking to export weights. China's own labs are thinking at even larger scales: DeepSeek has stated its intention to train an 8-trillion-parameter model on domestic chips, and Alibaba has disclosed plans for a 10-trillion-parameter Qwen model. Neither company is named in the draft. Both would be caught by the 2T threshold if they sought to transfer weights abroad under the rules as written.

The Agent and Chip Controls

The list sets specific numerical thresholds for intelligent agents as well. A planning system falls under the restricted category if it can decompose a plain-language instruction into 20 or more steps with at least 80% accuracy and 90% success. Multi-agent setups running 60 or more agents in coordination are also covered, as are tool-use systems that work with 200 or more third-party tools and select the correct one at least 90% of the time.

On the hardware side, the draft covers AI chip designs requiring 14-nanometre or finer manufacturing processes, or chips using 2.5D and 3D packaging. Networking infrastructure that connects clusters of more than 10,000 AI accelerators is also listed. This mirrors, in structure if not in specifics, the kind of hardware controls the US Bureau of Industry and Security has used since 2022 — except here the exporting country is China, and the targets are not named.

The most stringent category is a full export ban, not a restriction. It applies to AI that writes and mutates malicious code for attacks on critical infrastructure, where the system beats multi-engine malware detectors at least 60% of the time or a single mainstream antivirus product at least 85% of the time. AI-assisted vulnerability discovery systems are banned if they can build complete working exploits for known flaws in mainstream operating systems at least half the time, or autonomously discover five or more undisclosed high-severity flaws in a major open-source project in a single run.

Models trained on what China's Data Security Law defines as "core data" — the most protected category — face an outright export ban. Models trained on the somewhat less sensitive "important data" category are restricted unless that data made up no more than 0.1% of the training mixture and was cleaned such that it cannot resurface in outputs.

What the Rules Leave Alone

Two exemptions matter a great deal for Chinese AI companies with global products. First, anything already in the public domain through open-source releases or published academic papers is excluded from the controls. This directly protects DeepSeek's open-weight models and anything else a Chinese lab has already published under a permissive licence.

Second, and crucially for companies like Alibaba Cloud, ByteDance, and Baidu: offering a restricted model through an API, a consumer app or website, a SaaS product, or a lightweight on-device version does not count as a technology export under this draft. That means Kimi, Qwen, and similar products could continue serving foreign users through their apps and interfaces even if the underlying model weights became restricted. Running a restricted model on the company's own overseas cloud infrastructure would get an automatic licence. Open-sourcing a covered technology also receives automatic clearance.

Deploying a restricted model on servers operated by a foreign entity does count as a technology export and would require a licence. That is the line the draft draws: it controls who holds the weights and the source code, not who can use the API.

The MOFCOM spokesperson described the controls as "precise and limited," adding that "the vast majority of AI technologies can still be exported freely" once the final rules take effect. That claim holds up under scrutiny for consumer-facing products. For frontier model weights and advanced chip designs, it does not.

"AI Must Remain Under Human Control"

The ministry's stated rationale echoes language Xi Jinping used when meeting Donald Trump in September 2026: the goal is to ensure "AI always remains under human control." The MOFCOM spokesperson also pointed to precedent, noting that "all major economies" already control these technologies to some degree. The US Commerce Department restricted exports of Anthropic's Mythos 5 and Fable 5 models globally in June 2026, one of the first instances of a Western government placing a specific AI model under export control.

The symmetry is striking. As UNCTAD's recent analysis showed, the economic value generated by AI infrastructure is concentrating rapidly at the infrastructure layer, and governments on both sides of the US-China competition are now treating frontier model weights as strategic assets requiring the same sovereign control as advanced chips. Beijing's controls, if they pass as drafted, would give MOFCOM the same kind of gate-keeping authority over Chinese LLM exports that the US Commerce Department already has over advanced semiconductors.

The ministry was consulting Alibaba, ByteDance, Zhipu, and other domestic AI and chip firms on the shape of these controls since at least the summer, according to AI Weekly's tracking. A Financial Times report in July had already flagged that Beijing was considering exactly this kind of control. The October 16 announcement was not a surprise inside China's AI industry; the specifics of the numerical thresholds were.

The Timeline and What Comes Next

Comment period closes November 15. The ministry has not stated a timeline for finalising the rules after that. The temporary registration requirement that took effect on October 16 applies not just to new transfers, but to transactions already under way and not yet completed — a provision that may create friction for deals currently in progress between Chinese AI companies and overseas partners.

The final threshold numbers — exactly where the 2-trillion-parameter line sits, whether the FLOPs bar will be adjusted, how "core data" is operationalised — will be shaped partly by the industry feedback Beijing receives over the next 30 days. Chinese labs that want to keep the option of licensing their biggest frontier models to foreign enterprises have a narrow window to push for thresholds that their near-future models will not clear. Based on the scale of what DeepSeek and Alibaba have publicly said they are building, that window may already be closing.