China Proposes Export Controls on 23 AI Technologies Spanning Foundation Models and Intelligent Agents

China's Ministry of Commerce and Ministry of Science and Technology proposed adding 40 technologies to the national export control list on October 16, 2026. The draft catalogue places 23 artificial intelligence technologies under strict oversight, including large language models, intelligent agents, and computing clusters.[2][3]

The proposal marks the first time Beijing has explicitly targeted trained neural networks and model weights for export restriction, shifting the regulatory focus from hardware to software. A temporary pre-export registration requirement for the affected technologies took effect immediately on Friday.[4][5]

Across the full 40-item catalogue, 14 technologies are marked as prohibited from export, while 26 face restricted licensing requirements. The public consultation period for the draft rules remains open until November 15, 2026.[1][7]

Defining the technical thresholds

The draft regulations target the core architecture of advanced artificial intelligence systems, requiring government approval to export model weights, checkpoints, and source code. These restrictions apply to models containing at least two trillion parameters, or those ranking among the world's top 10 in performance and safety metrics.[3][6]

Chinese technology giants like Alibaba Group Holding Ltd. and Moonshot AI currently operate models that cross this two-trillion-parameter threshold, such as Qwen3.8-Max and Kimi K3. Exporting the underlying architecture of these frontier systems will now require explicit state authorization.[3]

However, the Ministry of Commerce carved out a significant exemption for technologies already available in the public domain. Open-source code and published academic research are excluded from the new controls, protecting existing open-weight models that global developers currently use.[4][5]

The rules also distinguish between exporting underlying technology and providing commercial access to it. Chinese firms can continue serving users abroad through application programming interfaces, websites, and software-as-a-service platforms without triggering the export ban.[5]

"Most AI technologies will continue to be freely exportable," a spokesperson for the Commerce Ministry stated on Friday. The spokesperson added that China remains "committed to building an open-source AI innovation ecosystem" while managing national security risks.[4]

Securing core data and infrastructure

Beijing framed the export controls as a necessary defensive measure against malicious actors targeting critical digital infrastructure. The draft explicitly bans the export of artificial intelligence models trained on China's "core data," which the Data Security Law defines as information concerning national security and the economic lifeline.[3][5]

The prohibited list includes autonomous generation and mutation technology designed for malicious code. Regulators argue these specific capabilities could be weaponized to attack defense systems or core government networks if acquired by non-state actors.[4][7]

"The Chinese government's necessary management of these technologies aims to solidify the security baseline, prevent misuse and abuse, ensure that AI remains under human control, and promote international governance of AI," the Commerce Ministry spokesperson said in a statement.[3][6]

The move partially addresses regulatory gaps exposed by recent international acquisitions of Chinese startups. In early 2026, Chinese authorities ordered the unwinding of Meta's $2 billion acquisition of the application Manus, highlighting Beijing's growing discomfort with offshore technology transfers.[5]

Chip design and solar manufacturing

While artificial intelligence dominates the updated catalogue, the restrictions also extend into advanced semiconductor design and renewable energy manufacturing. The draft limits the export of design know-how for chips fabricated at 14 nanometers or below, as well as 2.5D and 3D stacked chip architectures.[5][6]

The inclusion of chip design methodologies represents a reciprocal countermeasure to ongoing United States export controls. While Washington restricts the flow of advanced physical graphics processing units into China, Beijing is now restricting the flow of domestic chip design intellectual property out of the country.[5]

Beyond the technology sector, the proposal targets 12 photovoltaic technologies and five biomedicine categories. The draft prohibits the export of manufacturing technologies for perovskite and tandem solar-cell modules, aiming to protect China's dominance in the global renewable energy supply chain.[4][7]

In the biomedicine sector, the controls cover sensitive platforms related to cell and gene therapy, pathogenic microorganisms, and small nucleic acid drugs. These additions reflect a broader strategy to maintain domestic leverage across multiple strategic industries.[4]

Navigating the new compliance landscape

The immediate implementation of the temporary registration rule forces Chinese technology companies to pause and evaluate their international partnerships. Firms must now register any pending deals involving the 40 listed technologies before finalizing overseas transfers.[5][6]

For the global open-source community, the November 15 checkpoint will determine how future Chinese models are released. While existing open-weight models are protected, developers remain uncertain whether future iterations of systems like DeepSeek or Qwen will face delays or outright bans.[1][7]

The application programming interface exemption provides a crucial commercial lifeline for developers seeking international revenue. By allowing cloud-based access to restricted models, Beijing ensures its domestic champions can still compete for global enterprise contracts without surrendering their core intellectual property.[5][7]

The draft catalogue formalizes a new symmetry in the global technology conflict. As the United States and its allies tighten the hardware supply chain, China is establishing the legal framework to hoard the software and data architectures built on top of it.[2][5]