Beijing wants a say over who can take its most advanced AI software out of the country. On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology proposed adding 23 AI technologies to the country's export-control list, including model weights and source code. It is the first time the AI sector has been named in such a list, and a temporary registration rule took effect the same day.
The proposal is still open for public comment, so parts of the regime are already in force while the final text is unsettled. For American readers, the move reverses the usual direction of AI export policy. Washington has spent years limiting chip shipments to China. Now Beijing is moving to control software assets that US developers and companies depend on.
What China's draft export list covers
The draft covers 40 technologies across three fields: artificial intelligence, biomedicine and photovoltaics. The AI list accounts for 23 of them, with 12 solar items and 5 biomedicine items making up the rest, according to CryptoBriefing's summary of the proposal. Of the 40 items, 14 are marked prohibited and 26 are restricted.
China has run an export-control catalogue covering dual-use technology for years, long before AI became a policy focus. Tech Insider notes that the new element is the explicit naming of large models, agents and model weights.
Bloomberg describes the AI list as covering large language models, agents, chip designs and computing clusters. Model weights and source code for products above certain performance thresholds are the most consequential piece.
Which AI technologies and thresholds are covered
The Next Web reports the thresholds and scope that the draft appears to set out. The performance thresholds are not published in full, and the draft names no specific model.
| Item | Reported scope |
|---|---|
| Large models | At least 2 trillion parameters, or a top-10 global ranking on performance or safety tests |
| Chip designs | Chips at 14 nanometers or below, and designs built by stacking chips in 2.5D or 3D |
| Model weights and source code | Products that exceed the performance thresholds |
| Open-source code and published research | Exempt |
| Running a covered model on overseas servers | Counts as an export |
Why model weights and source code change the export debate
For three years, AI export controls ran in one direction. Restrictions targeted hardware, and the policy argument was about which chips could move. This draft points the other way, at software. OODAloop describes the shift as new territory, because Beijing would regulate the overseas transfer of foundational software assets, not only physical semiconductors.
The practical difference is the cost of moving the asset. A chip is a physical object that customs officers can count and inspect. A set of model weights is a large file that can be copied over a network in an afternoon. Compared with chip controls, which depend on a small number of fabs and shipping routes, software transfers are harder to see. The reporting does not describe how Beijing plans to check them.
One detail matters for anyone running Chinese models in the cloud. Running a covered model on servers abroad would count as an export. At the same time, Chinese firms could keep serving overseas users through apps, websites and APIs. The reporting does not explain how those two points fit together, and the final text will decide how companies read them.
Which Chinese models could be affected
Bloomberg and AP reporting says Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max appear to be above the size threshold, according to The Next Web's summary. The draft itself names no model, so those readings come from reporters rather than from regulators.
The reach extends well beyond China. Chinese open-weight models such as Qwen, DeepSeek, GLM and Kimi serve as base models for much non-Chinese AI work, according to Carly's AI news roundup. A US startup that fine-tunes Qwen or Kimi builds on exactly the kind of asset the draft would regulate.
Who gains is less clear. Non-Chinese open-weight developers could win business if buyers steer away from Chinese base models, but that outcome depends on the final list and on how the open-source exemption is read. Chinese labs, meanwhile, could lose overseas distribution for covered models even while app and API access continues.
Implicator.ai reported before the draft that Beijing was weighing controls on model weights, training data and chip designs. That earlier reporting said the measures could reach Alibaba's downloadable models and TSMC's overseas fabs. The same reporting tied the proposals to Beijing ordering Meta's roughly $2 billion Manus deal unwound, which suggests acquisition controls are also in play.
Agents are on the list too. USA 24 recently reported on rogue Tencent-linked AI agents that tried to scrape Alibaba's Amap data, a reminder that agent behavior is already drawing scrutiny before any export rule applies.
Chip rules and the Qualcomm and TSMC question
The chip-design rule would reach designs for chips at 14 nanometers or below, along with designs built by stacking chips in 2.5D or 3D packaging. Earlier reporting says the proposed chip rules could affect production at Qualcomm and TSMC for designs based on Huawei, Alibaba and ByteDance chips. Those are possible knock-on effects, not confirmed outcomes.
For US chip designers and foundry customers, the risk is indirect. A rule that reaches Huawei, Alibaba or ByteDance-based designs could change orders, timing or contract terms for firms further down the supply chain. ByteDance is reportedly nearing a deal with MediaTek for custom AI data center chips, which shows how many custom designs now cross borders before they reach a server rack. USA 24 covered that ByteDance and MediaTek talks earlier.
The registration rule and the November 15 deadline
A temporary rule took effect October 16 and requires firms to register before exporting listed items, including incomplete deals, according to Daily.dev's summary of the proposal. Public comment runs until November 15, 2026. That is the only firm deadline in the reporting, and no final decision or timetable has been announced.
Coverage also disagrees on the start date. One outlet says the draft was published on Thursday, while most others date it to Friday, October 16. The gap is small, but it matters for anyone building a compliance record. It is also unclear whether the temporary registration rule applies to exactly the same items as the eventual list.
What US developers and businesses should do now
I did not find official guidance for US users in my searches, so the steps below are a practical starting point rather than a compliance checklist.
- Inventory every model in your stack, including its origin, its license and whether any component was fine-tuned from a Chinese open-weight base.
- Ask hosting providers where a Chinese model actually runs. Serving from servers outside China is the scenario the reporting flags as a possible export.
- Add contract language that lets you swap or retire a model if a restriction takes effect.
- Read the full draft text when it is published, especially the open-source exemption. Coverage does not say whether downloadable weights released under an open license count as open-source code.
- Track the November 15 comment window and any final list, and keep a dated record of the decisions you make.
What comes next
Washington's reaction is the largest unknown. I did not find a formal US government response in my searches. Before the comment window closes, three questions will shape how much US companies need to rethink their use of Chinese open weights: whether the final list publishes the performance thresholds, whether any named model is confirmed above the line, and whether the chip and acquisition proposals move on the same timeline as the software rules. Each answer would change the cost of building on a model that is one download away from a regulated export.
