Isidora avatarIsidora & Cos Explain

China Wants to Control AI Exports. Will You Lose Access to Chinese Chatbots?

Painterly view of a vast data center hall at dusk with glowing blue server racks, an ornate old customs gate and an open ledger on a wooden desk in the foreground, no people, logos or text
Summary

On 16 October 2026 China's commerce and science ministries proposed adding 40 technologies to the list of what may not leave the country freely, 23 of them in artificial intelligence, and began requiring registration of such exports at once. The draft targets who holds the weights of frontier models, not who uses them: using a Chinese model through an app, website or API stays outside the controls, and anything already open-sourced is exempt. This explainer covers what is in the draft, what is not, the 2020 TikTok and 2026 Manus precedents, three ways it could end, and my view. Comments close on 15 November.

On Friday, 16 October 2026, China’s Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48. It proposes adding 40 technologies to the catalogue of technologies that are prohibited or restricted from export: 23 in artificial intelligence, 5 in biomedicine and 12 in solar power. The public can comment until 15 November.

For anyone who uses Chinese AI models such as Kimi, Qwen or DeepSeek, the obvious question is whether the door is about to close. The short answer from the text is no, not for ordinary users. The longer answer is more interesting, because the draft shows what Beijing actually wants to keep at home.

What the draft covers

The 23 AI items cover large models, AI agents, computing clusters and information processing. According to the ministry spokesperson, 14 of the 40 technologies are proposed as prohibited exports and 26 as restricted. The AI examples given were “AI-driven generation and mutation of malicious code”, which would be banned outright, and “AI-assisted vulnerability discovery”, which would need a licence.

For large models, the restricted tier covers the core of the model: weights, checkpoints, source code, architecture, and training or inference methods. Coverage by The Decoder and China AI Dispatch lists the triggers as training compute of at least 10^26 operations, at least 2 trillion parameters, or a ranking among the world’s top 10 models. A separate case covers models trained on China’s legally defined “important data”. Early reports read the combination of these conditions slightly differently, so the final annex will be the authority. Bloomberg also reports that design technology for AI chips at 14 nanometres or better, and for advanced packaging, would be restricted.

One part is already in force. From the day of the announcement, any company exporting a listed technology must register first on the ministry’s platform. That covers transferring ownership abroad, licensing it abroad, and a foreign entity taking control of the Chinese owner. Exports already under way but not finished are included.

What it does not cover

This is the part that matters most to users. The spokesperson said that offering a restricted model abroad through API calls, an app or website, a SaaS product, or a lightweight on-device version is not a technology export under these rules. What does count is deploying the model’s weights on servers in another country. Even that can get an “automatic licence”, which the ministry describes as roughly “file and it goes out”, if the servers belong to the Chinese company’s own overseas cloud.

Anything already in the public domain through open-source release or published papers is outside the scope. New open-source releases also qualify for automatic licensing. China AI Dispatch notes that Kimi K3 (2.8 trillion parameters) and Qwen3.8 (2.4 trillion) are already above the parameter line, but both have open weights, so both are exempt.

So a person chatting with a Chinese model in a browser is untouched. As China AI Dispatch puts it, this is “a rule about who holds the weights”. The company that would feel it is a foreign firm that wants the next closed frontier model on its own servers, or a foreign buyer that wants to purchase the lab that built it.

Why now: TikTok and Manus

Beijing has used this catalogue before. In August 2020, it added recommendation algorithms and AI interface technologies to the restricted list, the first revision in 12 years, just as Washington was forcing a sale of TikTok’s US business. A Chinese trade expert told state media the change meant the sale could need Beijing’s approval. The translated 2020 list shows how broadly those items were written.

The more recent case is Manus, an AI-agent startup founded in China that moved to Singapore. After Meta announced it was buying the company, the commerce ministry said in January 2026 that it would examine the deal against export-control and technology-transfer rules. In April, China’s planning commission ordered the acquisition unwound, and in August Manus said it would resume as an independent company. The new rule’s “change of control” trigger turns that one-off intervention into a standing procedure.

There is also a mirror image. The 10^26 threshold is the same line Washington drew in its January 2025 AI Diffusion Rule, which controlled the weights of closed models above it and left open weights alone. The US rescinded that rule in May 2025. China is now adopting the logic the US dropped.

The official reasons, and the unofficial ones

The ministry’s stated reason is safety: AI tools that write malware or find vulnerabilities could be used against critical infrastructure, defence and government systems by “non-state actors or malicious entities”. It also promises easier exports for Global South countries and for members of its World AI Cooperation Organization.

Analysts see more than safety. Ryan Etzcorn of Ankura China Advisors told Bloomberg that targeting AI, solar and biotech, where China has a growing edge, signals that “steadier relations with the West do not mean it will stand still”. The list puts frontier AI next to polysilicon and gene therapy. That looks like industrial policy and bargaining power as much as security.

Three ways this could go

Scenario A: it passes roughly as drafted. Most AI stays freely exportable, as the ministry promises. Chinese labs keep open-sourcing, with a form attached. Users abroad notice nothing. Foreign companies that want to self-host future closed Chinese models, or buy Chinese AI startups, face licences and delays.

Scenario B: it gets tighter. Thresholds could be lowered, or more items added, since the ministry says the list will be “dynamically adjusted”. Beijing could also encourage labs to stop open-sourcing their best models, so that new releases fall under the rule. The API and app carve-out is the line to watch. If that narrows, ordinary users would be affected, and nothing in the current text suggests that.

Scenario C: it is loosened or traded. Comments from Chinese companies could soften the thresholds. Because the catalogue sits alongside trade tensions, parts of it could also become a bargaining chip in trade talks, eased or enforced depending on relations with Washington and Brussels.

Opinion: the weights you already have stay yours

I don’t see this as China cutting the world off from its AI. I see it as China copying a playbook: keep the crown jewels at home, sell the service abroad, and let open-source spread your standards. It is the same bargain the US tried in 2025, and frankly a more coherent version of it.

For users, the practical reality is reassuring. A browser or app session is explicitly outside the rules. Open weights already published cannot be recalled, and US providers already serve Chinese models from US data centres.

The losers are elsewhere. Europe’s plan for “sovereign” AI often means running the best open models on European servers. If the best open models come from the US and China, and both governments now claim a say over their future weights, that plan rests on someone else’s goodwill. The rule is a reminder that open-source today is not a guarantee of open-source tomorrow. Europe should treat it as a deadline, not a headline.

This is not legal or investment advice.

Sources

Comments