China Wants Export Controls on 23 AI Technologies, Including Model Weights Above 2 Trillion Parameters

For years the export-control story in AI ran one direction: Washington restricting what China could buy. Today the arrow turned around. Beijing published a draft that would put its own models, agents and chip designs on a list of technologies that cannot leave the country without permission.

The headline number is 23. That is how many AI technologies China's commerce and science ministries want to add to the catalogue of restricted or prohibited exports. If you build on Chinese open-weight models, rent inference from a Chinese lab, or sell into China, the details matter more than the headline.

This post sticks to what the sources in hand report, flags where they differ, and separates what is proposed from what is already in force.

What Actually Happened

On Friday, October 16, China's Ministry of Commerce and Ministry of Science and Technology jointly published a draft update to the country's catalogue of technologies subject to export bans or restrictions. Per the Associated Press, the draft covers 40 technologies spanning AI, chips and solar power. Per The Next Web, the split is 23 AI items, 12 solar items and five in biomedicine, such as gene therapy.

Reuters described the AI portion as covering models, agents and computing clusters. Bloomberg's summary said the proposal could restrict overseas transfers of AI models and chip designs. The Next Web adds that the plan reaches model weights, source code and "other core parts" of large models.

The AI items fall into three groups:

• Large models. Per The Next Web citing Bloomberg, the curbs would apply to models with at least 2 trillion parameters, or models that rank in the world's top 10 on tests such as performance and safety. The AP adds that exports of core technologies for models trained on sensitive national data would be prohibited outright.
• Agents. Per the AP, AI agents designed for highly sensitive environments, such as national defense and critical information infrastructure, would also be banned from export.
• Chip design know-how. Per The Next Web, the draft would limit design knowledge for AI chips made at 14 nanometres or below, or built by stacking chips in 2.5D or 3D, plus some memory chip designs. The AP also reports several AI processing and memory chip design technologies on the list.

The ministry's stated rationale, quoted by both the AP and The Next Web, is to "prevent misuse and abuse" and ensure that "AI remains under human control." A ministry spokesperson called the move "necessary," cited national security, and said the plan does not target any country or firm.

One detail on timing. Per The Next Web, public comment runs until November 15, but a temporary rule took effect on Friday requiring firms to register before they export any of the listed items, including deals that are not yet complete. So the list is a draft, while the registration requirement is already live.

A note on discrepancies: a South China Morning Post summary indexed the same day described the plan as affecting 14 technologies, while Reuters, the AP, Bloomberg and The Next Web all point to 40 items with 23 in AI. We use the 40 and 23 figures because they come from four independent outlets and match the ministries' own draft as described by the AP.

The Carve-Outs Are Where the Business Is

Read the exemptions before reacting to the ban language.

Per The Next Web, open-source code and published research are exempt. Chinese firms could also keep serving users abroad through apps, websites and APIs, according to the ministry. In other words, the draft as reported does not cut off a developer in Austin who calls a Chinese model through an API.

What it does restrict is moving the model itself. Per The Next Web, running a covered model on servers abroad would count as an export. Firms that run it on their own cloud servers abroad could get a licence automatically.

That creates a clear split in how a Chinese model can reach you:

• Hosted API or app: continues, per the ministry statement as reported.
• Weights deployed on third-party infrastructure outside China: treated as an export of a covered technology if the model meets the size or ranking test.
• Weights on the lab's own overseas cloud: eligible for an automatic licence, per The Next Web.

The tension is with the open-source exemption. The sources do not explain how a model that is both above 2 trillion parameters and released as open weights would be treated. The AP reports that Kimi K3 and Qwen 3.8-Max are among models above the 2 trillion threshold, but also that the draft names no specific model. Until the final text clarifies it, nobody outside the ministries can say with confidence which released weights fall inside or outside the line. We covered Alibaba's open-model push earlier in our Qwen 3.8-Max post; the draft raises the question of whether that posture extends to the very top of the stack.

Why This Lands Now

The timing is the second story. Per the AP, the proposal arrived while China and the US were making progress in dialogues over AI safety and governance, including a notification mechanism for AI incidents that could threaten national security. We covered that channel in this earlier post. Cooperation on incident reporting and tightening on exports are running in parallel.

The AP also quotes two analysts. Omdia's Lian Jye Su said the move "can be seen as a response to Western countries posting sovereignty requirements on AI, robotics, telecommunications, and other key technologies," with the aim of better regulating access to Chinese technologies. Natixis economist Gary Ng said "the world is getting more fragmented in trade and investment, but also in access to technologies."

Washington has used this tool on its own champions. The AP notes that the US Commerce Department in June asked Anthropic to bar foreign nationals from using some of its most powerful models, Fable 5 and Mythos 5, citing national security, and that the restrictions were later lifted. Our write-up of that episode is here. The pattern is that frontier model access is now treated as a strategic export in both capitals, and that the rules can change quickly in either direction.

What the Sources Do Not Tell Us

Several things are unsettled, and it is better to say so than to guess:

• Final scope. This is a draft with a comment period. The final list could narrow or widen.
• How the ranking test works. The "top 10 on tests such as performance and safety" language is relayed by The Next Web from Bloomberg. The sources do not say which benchmarks or who decides.
• Licence mechanics. The automatic-licence path for firms using their own overseas servers is described in a single sentence. Terms, timing and fees are not reported.
• Market impact. None of the sources in hand give figures on revenue exposure for Chinese labs or Western customers. We are not going to invent one.
• Enforcement against third parties. Whether a Western cloud hosting already-released weights faces any consequence under Chinese rules is not addressed in the reporting.

What Builders Should Take From It

• Inventory your Chinese-model dependencies by access path. Separate hosted-API use from self-hosted weights. The draft, as reported, protects the first and puts the second in a grey zone for the largest models.
• Pin the weights you already depend on. If you run open-weight Chinese models in production, record exactly which versions and licences you hold today. Newer, larger releases may arrive with different terms.
• Keep a fallback that is not on the list. If a workflow depends on a Chinese frontier model, test a second provider on the same evals now, while it is a planning exercise and not an outage.
• Watch the November 15 comment deadline. That is the next date at which the scope could change, and the registration rule is already live for deals in flight.
• If you sell into China or partner with Chinese labs, ask about registration. Per The Next Web, the temporary rule covers incomplete deals. Contract language about model delivery and hosting location should be checked with counsel.
• Do not read this as a ban on API access. The sources are explicit that apps, websites and APIs continue. Plan for a narrower weights pipeline, not a closed door.

The Builder Take: The practical risk is not that Chinese models disappear from your stack tomorrow. It is that the cheapest way to use them, downloading weights and running them on your own cloud, is the path the draft puts under export rules. Know which side of that line each of your dependencies sits on before the final text lands.

Developer312 covers the AI business signals builders actually need to act on. Get the weekday briefing at developer312.com.