Beijing has put numbers on how far its AI export curbs would reach. On October 16, 2026, the Ministry of Commerce and the Ministry of Science and Technology published draft changes for public comment covering 40 technologies across artificial intelligence, biomedicine and solar manufacturing. Caixin reported that 14 of them would be banned from export outright and the other 26 would need export licenses. Twenty-three of the 40 are AI-related, which makes this the most detailed look yet at where China wants to draw a line around its own AI.
The story has moved fast since the draft appeared. The Economic Times framed it as export curbs that draw a line around China’s AI crown jewels, and the framing fits. The draft names advanced large models, AI agents, chip-design technologies and computing clusters. It also arrives while Chinese labs keep shipping open models, a tension we covered in our look at DeepSeek’s open-weight bet and the caution around it. These are draft proposals, not final rules, so everything below describes what Beijing floated and what it has not yet decided.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
China AI Export Controls at a Glance
The draft revises China’s catalogue of technologies that cannot be exported or can only leave the country with permission. Both ministries opened it for public comment on October 16. A ministry spokesperson, unnamed in the available report, said the measures aim to ensure that AI technology is “safe and controllable.”
Three features set this proposal apart from earlier tech-export notices. It names concrete AI categories instead of speaking in broad terms. It sets at least one numeric threshold, a model size of 2 trillion parameters. And it comes with a temporary registration requirement that took effect the same day the draft was published, so companies did not get a quiet grace period while the catalogue was debated.
The draft does not name a single Chinese AI company, model or product as covered. Nor does the available reporting identify any government official behind the text. Readers who see claims that the rules aim at DeepSeek, Alibaba, Baidu, Huawei or any other firm should treat them as unconfirmed.
The 40-Technology Breakdown: 14 Bans and 26 Licenses
The headline split is simple. Of 40 technologies, 14 face an outright export ban and 26 would require a license, according to Caixin. AI dominates the list with 23 entries. The remaining 17 sit in biomedicine and solar manufacturing, and the available reports do not break those down further.
| Item | Figure or detail | Status |
|---|---|---|
| Draft published | October 16, 2026 | Public comment stage |
| Publishing bodies | Ministry of Commerce and Ministry of Science and Technology | Confirmed |
| Total technologies covered | 40 across AI, biomedicine, solar manufacturing | Draft |
| Outright export bans | 14 technologies | Draft |
| License required | 26 technologies | Draft |
| AI-related entries | 23 of the 40 | Draft |
| Large-model size marker | 2 trillion or more parameters | Could be restricted |
| Cluster marker | Coordination of more than 10,000 AI chips | Included in draft |
The ban-versus-license divide matters for planning. A banned technology cannot leave China under any deal. A licensed one can, but only after a government review, which hands regulators a veto and a timeline lever over every overseas partnership.
The 2 Trillion Parameter Line for Large Models
The most quotable number in the draft is 2 trillion. Models at or above that parameter count could fall under export restrictions. Parameter counts are a blunt measure, since architecture, training data and efficiency tricks now matter as much as raw size. Still, a hard number gives lawyers something to work with, and it signals that Beijing views the biggest systems as strategic assets.
A second rule goes further. The proposal would prohibit exports of core technologies for AI large models trained using sensitive national data. That shifts the question from how big a model is to what it learned from. Any developer who trained on data Beijing regards as sensitive could find its model weights, training methods and related know-how off the table for foreign buyers.
For context on the pace of the global race, the frontier keeps moving. Our comparison of GPT-6.1 Sol, Sonnet 5.5 and Gemini 4 Argon shows how quickly Western labs are stacking releases. China’s draft suggests its regulators want a say in which of their own frontier systems cross the border.
AI Agents and Sensitive Environments
The draft treats AI agents as their own category. Agents built for highly sensitive environments, including national defense and critical information infrastructure, would be subject to export controls. The wording ties the restriction to the use case, not the product label. An agent framework sold to a retailer and a near-identical one built for a defense network would sit on opposite sides of the line.
That approach mirrors the debate in the West, where regulators and labs argue over capability tiers. A cyber-capability threshold is already part of how OpenAI handles its newest release, as we detailed in our piece on the GPT-6.1 Terra critical cyber tier. Beijing’s agent rule reaches for the same idea from the export side: the riskier the deployment, the tighter the gate.
Computing Clusters and Chip Design Technology
The proposed AI controls also cover chip-design technologies and computing clusters. One marker stands out: technology capable of coordinating more than 10,000 AI chips. Orchestration software at that scale is what turns a warehouse of accelerators into one training machine, and it is hard to reproduce without experience running such systems.
China cannot buy its way past every chip constraint, so the know-how to run large clusters efficiently has become a prized capability. Putting the coordination layer under export control protects that investment. It also tells foreign buyers that Chinese hardware and software are not a package they can pull apart and rebuild abroad.
Access Is Not Transfer: The Draft’s Central Distinction
The part of the draft that matters most to everyday users is also the easiest to miss. The proposal would broadly allow overseas users to reach Chinese AI through websites and cloud platforms. What it targets is something different. Some transfers of underlying technology, and some overseas deployments, would need government approval.
Using a service versus owning the capability
The proposal separates access to Chinese AI services from foreign acquisition of the models, data or expertise needed to operate or reproduce them independently. A developer in Berlin or Austin who calls a Chinese model through a cloud endpoint would stay on the permitted side. A foreign firm that wants the weights, the training data or the engineers to run its own copy would face review. Caixin said the proposal preserves exemptions and streamlined approvals for much international AI collaboration.
| Activity | Likely treatment under the draft | Basis |
|---|---|---|
| Overseas user chats with a Chinese model on a website | Broadly allowed | Access through websites and cloud platforms |
| Overseas company calls a Chinese model on a cloud platform | Broadly allowed | Access through websites and cloud platforms |
| Foreign buyer acquires the models, data or expertise to run a copy | Government approval likely | Underlying technology transfer |
| Chinese firm deploys a restricted model in an overseas data center | Government approval likely | Overseas deployment |
| Export of core tech for models trained on sensitive national data | Prohibited | Proposed ban |
| Export of agents for national defense or critical infrastructure | Export controls apply | Proposed control |
The ministry spokesperson addressed the effect on Chinese AI companies expanding abroad in similar terms. The outcome, the spokesperson said, depends on “whether it involves the transfer of restricted technologies from within China to locations outside its borders.” That sentence carries the whole logic of the draft. Selling a service is fine. Moving restricted know-how out is where the review starts.
The Temporary Registration Rule Already in Force
Public comment usually means waiting. Not here. A temporary registration requirement for certain transfers took effect on October 16, 2026, the day the draft appeared. It applies while the catalogue is under consultation.
The registration covers transfers of technology ownership, overseas licensing and transfers of ownership to a foreign entity. Incomplete exports are covered too, which means a deal that is mid-flight can still get pulled in. A company that signed a licensing term sheet last month and has not yet closed should assume the requirement touches it.
The design is practical. Regulators get visibility into who is moving what, and companies get no window to rush deals out before the final text lands. Any firm with a pending foreign transaction involving AI know-how now has a compliance step to add to its closing checklist.
What Is Confirmed and What Is Not
Early coverage of fast-moving policy stories tends to blur drafts and decisions. The table below separates what reporting supports from what remains open.
| Claim | Status | Note |
|---|---|---|
| Draft published October 16 by two ministries | Confirmed | Reported by Caixin and others |
| 40 technologies, 14 banned, 26 licensed | Confirmed as draft | Caixin figures |
| 23 of 40 are AI-related | Confirmed as draft | Multiple reports |
| Models of 2 trillion or more parameters could be restricted | Confirmed as draft | Not a final threshold |
| Temporary registration in force from October 16 | Confirmed | Applies to certain transfers |
| Rules are final | Not true | Eventual scope depends on finalization |
| Rules target a named company such as DeepSeek or Huawei | Unconfirmed | No company named in available reports |
| A specific named AI model is banned | Unconfirmed | No model named in available reports |
Why Beijing Is Drawing This Line Now
Export controls are a familiar tool for China. Caixin’s page for this story lists related headlines on stepped-up enforcement of critical mineral export controls and a ban on dual-use exports to 10 U.S. entities. The AI draft extends that toolkit from raw materials and hardware into software and know-how.
Chosun Ilbo’s English edition ran the story under the headline China Tightens AI Export Controls Amid Tech Rivalry. That rivalry is the backdrop. For several years, Washington has restricted chip sales to China, and Chinese companies have responded by building their own capability and pushing it abroad. Now that capability has value Beijing wants to protect.
The timing also tracks the pace of releases. The global AI market is cranking out new models almost weekly, and Chinese labs keep publishing open models. A policy that lets users touch Chinese AI but keeps the crown jewels at home fits a strategy of winning global usage without handing rivals the blueprint.
How the Draft Compares With US Chip Controls
The US approach has centered on hardware, limiting what advanced chips Chinese buyers can purchase. China’s draft works from the other end. It targets models, agents, cluster software and design know-how, the layers that sit on top of silicon. Bruegel’s analysis carries a headline that captures the contrast, saying China’s outbound AI strategy starts with models and governance and ends with chips.
Two differences stand out. First, the US controls limit what leaves American firms and reaches China, while China’s draft limits what leaves China and reaches the world. Second, Washington’s tools lean on hardware performance thresholds, whereas Beijing’s draft mixes size thresholds, data sensitivity and use-case sensitivity. Chip supply still shapes the whole picture, as our report on TSMC weighing a CoWoS price hike for Nvidia chips shows, but the Chinese draft puts the policy spotlight on software and models.
Market Impact for Chinese AI Developers
No reporting yet quantifies a financial hit, and no company has been named, so any stock-level forecast would be guesswork. The practical effects are easier to describe. Chinese AI firms that expand overseas will need to map each planned export against the restricted list and decide whether it is a service, a license or a transfer.
Where the friction lands first
Joint ventures and overseas data center plans carry the most exposure, because they usually involve moving weights, engineers or training know-how. Pure API businesses carry the least, since the draft broadly allows access through websites and cloud platforms. Companies selling model access abroad can keep operating, but anyone planning to host a restricted model on foreign soil should expect an approval step.
Pricing pressure is a second angle. Chinese labs have competed hard on cost, a point visible in our Kimi K3.1 Pro versus Opus 5.5 versus GPT-6 Astra price comparison. If the final rules slow self-hosted deployments abroad, API access becomes the main route for foreign customers, which keeps pricing and terms of service in the spotlight.
What It Means for Global Builders and Buyers
For most developers outside China, the draft changes little today. If you call a Chinese model through a hosted API, the draft’s access carve-out suggests that workflow stays intact. If your roadmap depends on downloading weights, licensing source technology or hiring a Chinese team to stand up a private deployment, the picture is less settled.
Procurement teams should do three things now. Inventory every vendor relationship that involves Chinese AI technology. Flag any contract that transfers ownership, grants an overseas license or places technology under foreign control, since those fall under the registration rule. And track the final text, because thresholds such as the 2 trillion parameter line could move before adoption.
Governments are also tightening their own stance on which AI services their agencies can touch, as the Swedish decision in our report on Sweden blocking ChatGPT, Claude and Copilot illustrates. AI policy is turning into a patchwork, and companies operating across borders will have to read each jurisdiction’s rules on their own terms. For more coverage of models and policy, see our AI and machine learning section.
Five Predictions for the Final Rules
These are analysis, not reported facts. They rest on the structure of the draft and the patterns of past rulemaking.
- Thresholds will get fuzzier, not sharper. A single parameter count ages badly as architectures change, so expect the final text to pair size with other tests or leave room for regulators to judge case by case.
- The access carve-out will survive. Beijing gains influence when foreign users depend on Chinese AI services, so cutting off cloud and website access would work against its own goals.
- Registration will become routine paperwork. The temporary rule gives regulators a data set, and that data set will likely shape how licenses are granted.
- Western regulators will cite the draft. Officials debating model-level controls, a topic touched on in our coverage of Anthropic’s stance on court blocks for risky models, now have a major government’s text to point to.
- Named-company clarity will arrive through guidance. Since the draft names no company, firms will learn their status through follow-up notices and licensing outcomes, not the catalogue itself.
Frequently Asked Questions
What did China announce on October 16, 2026?
The Ministry of Commerce and the Ministry of Science and Technology published draft changes to its export catalogue for public comment. The draft covers 40 technologies in AI, biomedicine and solar manufacturing.
How many technologies would be banned or licensed?
According to Caixin, 14 technologies would be banned from export and 26 would require export licenses. Twenty-three of the 40 are AI-related.
Are the AI export rules final?
No. The proposal is a draft released for public comment, and the eventual scope depends on how it is finalized. A temporary registration requirement for certain transfers did take effect on October 16.
Can overseas users still use Chinese AI models?
The draft would broadly allow overseas users to access Chinese AI through websites and cloud platforms. It targets transfers of the underlying technology and some overseas deployments, which would need government approval.
What is the 2 trillion parameter threshold?
Models with 2 trillion or more parameters could be subject to export restrictions under the draft. The figure is part of a proposal, so it could change before the rules are final.
Does the draft target DeepSeek, Alibaba, Baidu or Huawei?
That is unconfirmed. The available reports do not identify any specific Chinese AI company or model as directly covered.
What does the temporary registration cover?
It covers transfers of technology ownership, overseas licensing and transfers of ownership to a foreign entity. Incomplete exports are also covered.
