ConcludedRan 6 daysin the edition since 19 Oct 20268 sourcesupdated 24 Oct 2026
China's AI export control catalogue
The gist
Latest: China Proposes Export Controls on 23 AI Technologies, Firms Begin Registering
China's proposal establishes a licensing regime over its most capable AI models and creates an immediate registration requirement for any international transfer of covered technology, including open-source releases. The US Treasury's parallel proposal to ban American investment in Chinese AI companies adds a second front to the policy tension between the two countries.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
China's MOFCOM and MOST jointly issued Announcement No. 48 on October 16, 2026, proposing to add 40 technologies to the country's restricted and prohibited export catalogue, with AI as the largest category at 23 entries covering large language models, intelligent agents, computing clusters, and information-processing systems. Frontier models above 2 trillion parameters or trained with more than 10^26 FLOPs require a government license before weights or source code can leave the country. A pre-export registration requirement took effect immediately on October 16, independent of how the final catalogue resolves. The public comment window runs through November 15, and the final catalogue may differ from the draft.
Several carve-outs apply: access through APIs, websites, apps, or SaaS platforms is not covered, and already-public open-source technologies are exempt, with open-source technologies set up for automatic approvals. Since October 16, DeepSeek and MiniMax have each registered with the government before releasing model weights or code internationally, and both have continued to publish open-source models after filing. Open-sourcing weights for global download is classified as an export under the rules, so even open-source releases require a registration filing.
Nvidia CEO Jensen Huang, during a Taiwan visit with Lex Fridman, argued that US export controls make sense for physical goods but not for computing technology, with his core argument that broad sales generate wealth that funds R&D. The US Treasury separately proposed rules that would ban American investment in Chinese AI companies.
DeepSeek and MiniMax have registered with China's government and continued releasing open-source models since October 16, confirming the pre-export registration requirement in Announcement No. 48 is being followed in practice.
Open-sourcing model weights for global download counts as an export under the framework, so even open-source releases require a registration filing before weights leave the country. Separately, the US Treasury proposed rules that would ban American investment in Chinese AI companies.
Samsung is converting its Xian plant, which makes roughly 40% of its global NAND output, from 200-layer to 286-layer chips under a project called Monami, targeting 40,000 wafers per month by December 2027.
SK Hynix is upgrading its Dalian plant from 192-layer to 240-layer NAND. Both companies say the work adds no new capacity and complies with US export rules, but the US Commerce Department stated in 2025 it would license only continued operation, not technology upgrades, leaving the scope of any new approvals unclear.
OpenAI reportedly reversed its position on AI chip export controls, apparently to appease Nvidia, a move Miles Brundage called bad for US AI dominance and national security.
Reporting tied the shift to Nvidia's role as a $30 billion equity investor in OpenAI and guarantor of up to $105 billion for its Ohio data center. Anthropic backed chip restriction bills and left a trade association in September 2025 over lobbying against them. ByteDance, Alibaba, and Tencent ordered over 3 million Nvidia RTX PRO 5500 GPUs worth more than $35 billion, with Chinese orders reportedly exceeding Nvidia's projected 2027 global production.
China's MOFCOM and MOST jointly issued Announcement No. 48 on October 16, 2026, proposing that 23 of 40 new restricted export entries cover AI, with models above 2 trillion parameters or trained with more than 10^26 FLOPs requiring a government license before weights or source code can leave the country.
A pre-export registration requirement took effect immediately on October 16; the comment window closes November 15, 2026. APIs, SaaS, and app access are not covered, and already-public open-source models are exempt.
China's October 16 controls cover AI processor designs at 14nm or below, DRAM designs at 18nm or below, and interconnects for AI clusters with more than 10,000 accelerators, adding a hardware layer to restrictions previously described mainly in software terms.
At the International AI Summit in Dublin on October 14, Anthropic's international chief Guillaume Princen said open-weight AI models carry more safety risk by definition, citing developers' inability to control or monitor deployed copies. 'by definition, there's more risk with them,' Princen said.
An Institute for Progress report published October 19 estimates the US AI capability lead over China at approximately 6 months, finding adversarial distillation cuts what would be a 13-month advantage to about 6.5 months.
The report argues existing BIS, DOJ, and FTC authorities could widen the lead to over a year by restricting China's remote access to overseas training clusters. China's export controls, which took effect October 16, also cover advanced 3D NAND designs.
First in the edition · 19 October 2026
18 October 2026
Jensen Huang raised the possibility that chipmakers could refuse supply to customers developing unsafe AI.
16 October 2026added 18 Oct
China's interim AI export control catalogue took effect October 16, banning 14 categories outright, including systems that autonomously generate or mutate malware, and requiring government approval for 26 more, including large models, AI agents, and computing clusters.
APIs, apps, and already-public open-source technologies are exempt, and public comment closes November 15. On the same day, Nvidia CEO Jensen Huang argued against restricting computing exports, calling broad sales 'optimizing for a flywheel' and describing it as 'a more sophisticated long-term strategy'.
Jensen Huang defined sovereign AI as owning data, culture, and knowledge with guaranteed access, not as running on domestic hardware.
14 October 2026
Financial Times reported Chinese tech firms bought more than 1 million Nvidia gaming and workstation GPUs over the past year to supplement AI compute under export restrictions.