China to control exports of 23 AI technologies: 14 proposed banned, rest require license
The draft, released October 16, 2026 by the commerce and science ministries, splits the technologies into 14 prohibited and 26 license-required — but where the performance thresholds for model weights and source code fall, and how the rules will hit China's dominant open-weight ecosystem, remain unresolved.
China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) published a draft revision of the country's catalog of technologies prohibited or restricted for export on October 16, 2026. The draft adds 40 technologies across artificial intelligence, biomedicine and photovoltaics, of which 23 are AI technologies — including large language models, agents, chip design and data clusters. That is according to Bloomberg, apparently the original source behind most subsequent coverage.
The split is twofold: 14 technologies are proposed for outright export prohibition, while 26 are "restricted" and require an export license. The catalog was first published in 2001 and previously revised in 2008, 2020, 2023 and 2025, according to Global Times, which is the most detailed available source on the draft's contents.
What is proposed for prohibition, and what requires a license
Global Times reports that the draft would prohibit export of the core technologies of large AI models trained on "national key sensitive data," as well as AI agents serving critical national information infrastructure, national defense and core government affairs. On the chip side, design technologies for 2.5-dimensional and three-dimensional AI chips, along with 3D flash memory, are proposed as "restricted" — that is, license-required, not banned.
Bloomberg adds another, longer-term layer of impact: model weights and source code for products above certain performance thresholds are also to be license-required. Weights are the numerical parameters a model develops during training; if they are transferred, the recipient in effect gets the finished product without paying for the training run itself, as Crypto Briefing explains in its review of the proposal. The thresholds — the point that determines whether the rules hit a handful of frontier systems or a far larger swath of the industry — are not specified anywhere in the available material.
The process: consultation until November 15, interim registration already in force
The public consultation is open until November 15, 2026, giving companies and other stakeholders a defined window to object or request clarification. In the meantime, an interim requirement of pre-export registration is already in effect for the affected technologies, according to Crypto Briefing — a point so far confirmed only through secondary sources.
The proposal follows consultations in July 2026 with major companies, including Alibaba and ByteDance, which according to Crypto Briefing concerned limits on data transfers. It is unclear how those talks shaped the draft. Note also that November 15 is a consultation deadline, not a date on which the rules take effect.
Nikkei Asia independently confirms that Chinese authorities on the Friday proposed adding AI, biopharmaceuticals and photovoltaic technology to the export-control list, against the backdrop of persistent trade and technology tensions with the West.
The open-source carve-out
Global Times reports that the draft contains an important exception: technologies that have entered the public domain through open source, published academic papers or other means are not covered by the controls.
The carve-out stands in a certain tension with one of the big open questions: how the rules will treat publicly released model weights going forward. The question is not academic. In the week the draft was released, four Chinese models — DeepSeek, Z.ai, Xiaomi and Tencent — held places among the five most-used on OpenRouter's ranking of large language models, according to Rest of World. China's open-weight ecosystem is thus globally dominant, and how the line between "already publicly released" and "newly controlled, but openly released" will be handled in future releases is unresolved in the available material.
The authorities' rationale
A spokesperson for the Ministry of Commerce stressed, according to Global Times, that the measures not only guard against the risk of technology escaping control and spreading in an uncontrolled manner, but also ensure normal international technology exchange and cooperation, and that they promote science in a "positive and benevolent direction" — safe, controllable and beneficial to humanity. This is the authorities' own characterization, not an independent assessment.
What remains unresolved
Several essential points remain open. The performance thresholds for model weights and source code have not been published, and they will in practice determine how deep into the industry the controls reach. None of the sources provides access to the draft itself or the complete list of the 23 AI technologies — all coverage is secondary, with Crypto Briefing explicitly building on Bloomberg. And it is uncertain what changes the consultation feedback may bring before the final catalog is set.
What is fixed is the timeline: the draft was released October 16, the registration requirement is already in effect, and stakeholders have until November 15, 2026 to respond.