China Wants a Licence for Its Best AI: Draft Export Controls on Model Weights and Agents, Taiwan’s Chip Crackdown and What It Means for Your AI Stack (AI Trends, 16 October 2026)

What’s trending in AI on 16 October 2026: For three years the AI export story ran in one direction: Washington restricting the chips that flow into China. Today the traffic got heavier both ways. China’s commerce and science ministries published a draft that would put 23 AI technologies on the country’s export control list, including large language models, AI agents, computing clusters and, above a performance threshold, model weights and source code. A temporary registration rule took effect the same day. Hours later, Bloomberg reported that Taiwan plans to make unlicensed shipments of AI chips and servers to China a criminal offence, as part of its trade commitments to the United States. Together they mark a shift: the models themselves, not only the hardware they run on, are now treated as strategic goods. If your business runs DeepSeek, Qwen, Kimi or GLM, buys AI servers, or licenses technology across the Pacific, this affects your plans. Below: what Beijing’s draft covers, what it leaves out, what Taiwan is changing, and a four-week plan to work through before the consultation closes on 15 November.

Key takeaways

China wants a say over who gets its best AI. The draft lists 40 technologies, 23 of them AI-related, from language models and agents to chip design know-how and model weights.
It is a draft, but registration has already started. Comments close on 15 November. An interim rule requiring companies to register before exporting listed items, including deals already under way, applies now.
Apps, APIs and open source are largely outside it. Reports say Chinese firms can keep serving foreign users online, and published open-source code and research are exempt. Hosting a covered model on servers abroad would count as an export.
Taiwan is closing its own gap. It plans to add AI servers and advanced chips to its controlled list and make unlicensed exports to China a crime, aligning parts of its regime with US rules.
The practical risk is supply, not access. Expect friction around future weight releases, cross-border licensing and AI hardware, so check where your AI stack depends on one country’s goodwill.

For the first time, the models themselves are on a control list, not just the chips they run on.

1. What China published today

On 16 October, China’s Ministry of Commerce and Ministry of Science and Technology released a draft update to the catalogue of technologies whose export is prohibited or restricted. According to The Next Web, which drew on Bloomberg and the Associated Press, the draft adds 40 items: 23 linked to artificial intelligence, 12 to solar power and five to biomedicine, such as gene therapy. Tech Insider and Crypto Briefing report that 14 of the 40 fall into the prohibited tier, where no export is allowed, and 26 into the restricted tier, which requires a licence. The reports do not say exactly how the AI items split between the two tiers.

The AI items named so far are broad. They include large language models, AI agents, AI chip designs, computing cluster architectures, and model weights and source code once a product passes a performance threshold. The Next Web adds design know-how for AI chips made at 14 nanometres or below, or built with 2.5D or 3D stacking, plus some memory chip designs. Tech Insider says the prohibited examples include AI-driven techniques for generating and mutating malicious code, and models trained on what China classes as core data.

A ministry spokesperson said the rules are meant to prevent misuse and abuse and to “ensure that AI remains under human control”. The ministry also said the plan does not target any particular country or company. Public comments close on 15 November, after which a final version will be published.

AI makes up more than half of the new list, and its items reach from silicon designs to model files.

2. Which models would be caught

The draft itself names no models, and Crypto Briefing notes the performance thresholds have not been published. Bloomberg’s reporting, as summarised by The Next Web, says the controls would apply to models with at least two trillion parameters, or to models that rank in the world’s top ten on performance and safety tests. Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8-Max are reported to be above the size threshold. A ranking test would move with the market, so a new release could cross the line on launch day.

That matters because Chinese open-weight models are now part of many Western AI stacks. As we noted when looking at open-weight models trailing the frontier by about 4.4 months, companies pick DeepSeek, Qwen, GLM and Kimi for price, self-hosting and data control. The draft does not ban those uses, but it gives Beijing a lever over the very top of each model family, which is exactly the tier that buyers chasing frontier performance want.

3. What the draft leaves out, and what counts as an export

The exemptions are as important as the list. The Next Web reports that open-source code and published research are excluded, and that Chinese companies can keep serving users abroad through apps, websites and APIs. In other words, a developer in Ohio calling a Chinese model’s API is not, on current reporting, an export problem. Tech Insider treats the API exclusion as less firmly confirmed, so it is worth watching in the final text.

The line is drawn at where the model physically runs. Running a covered model on servers outside China counts as an export, according to The Next Web, although a company running a covered model on its own cloud servers abroad could receive a licence automatically. Registration also reaches beyond shipping files. The National reports that it would cover changes in ownership or control of Chinese technology held by overseas entities, as well as licensing arrangements, and Tech Insider says it applies to exports that have started but are not complete.

Scenario	Status under the draft, as reported	What to watch
Using a Chinese model through its app, website or API	Reported as outside the controls	Whether the final text keeps the API exclusion
Downloading weights already published as open source	Exempt as published open-source material	Whether future frontier releases are still published openly
A Chinese firm hosting a covered model on servers abroad	Counts as an export; may get an automatic licence on its own cloud	Regional availability of top-tier models from Chinese clouds
Licensing Chinese chip design or model technology	Registration required; licence needed if restricted	Deal timelines and contract conditions
Acquiring or restructuring a company holding Chinese AI technology	Change of ownership or control may trigger registration	M&A and joint venture due diligence
Exports already under way	Interim registration applies now	Delays to deliveries in progress

Sources: The Next Web, The National, Tech Insider (16 Oct 2026). Based on reporting about a draft; the final rules may differ. Not legal advice.

Open weights are hard to control once they are public. As Crypto Briefing put it, weights posted for download are “notoriously hard to recall”. The likelier effect is on what happens next: whether Chinese labs keep publishing their strongest models openly, or hold back the top tier while releasing smaller versions.

Using a Chinese model is largely untouched. Moving one, or the know-how behind it, is what gets controlled.

4. Taiwan closes its chip loophole

The other half of today’s story comes from Taipei. Bloomberg, as reported by Bloomingbit, says Taiwan plans to add AI servers and advanced semiconductors to its export control list, so that shipping them to China without a licence becomes a criminal offence under its Foreign Trade Act. The economy minister told Bloomberg that some future controls will be aligned with US rules, reflecting commitments in Taiwan’s trade agreement with Washington. No timeline was given.

The change fills a real gap. Taiwan’s current law has no provision that treats unauthorised AI chip exports to China as a crime in themselves, so prosecutors in smuggling cases have had to rely on other laws. Startup Fortune reports that since mid-2025 Taiwanese firms have needed permits to sell to Huawei and SMIC specifically, and that the new approach would cover buyers across China rather than named companies. US authorities believe billions of dollars’ worth of Nvidia chips have reached China illegally in recent years, according to the Bloomberg report.

For buyers outside China, the effect is more paperwork rather than less supply. Taiwan builds a large share of the world’s AI servers, and tighter controls usually mean more end-user checks, longer lead times and stricter questions about where hardware ends up. That lands on top of the AI memory shortage that is already stretching delivery dates and budgets.

5. Why both moves landed now

Neither announcement came out of nowhere. The Next Web notes that the Financial Times reported in July that China was weighing export controls on AI models and chips, and Crypto Briefing says the draft builds on consultations with Alibaba and ByteDance that month about limits on data transfers. Knowledge about AI has already become an intelligence target, as we covered in AI know-how is the new spy target, so treating the models as controlled goods is a logical next step.

The timing is also diplomatic. According to The National, the US-China trade truce agreed last month cut tariffs on $30 billion of goods and opened an AI dialogue, and it has been extended by two months. Trade expert Wendy Cutler told the paper that the “fragility of our trade truce cannot be overstated” and expects a US response. Leonhard Xu of the consultancy Sinolytics called the announcement extremely broad and noted that an export control is usually not an export ban. Most listed technologies can still be exported, just with registration and, for some, a licence.

Four years after the first US chip rules, the controls now cover model files and design know-how.

6. What this means for your AI stack

For most businesses nothing breaks today. The models you already downloaded keep working, and API access is reported to be outside the rules. The risk is about the future: the next model you plan to adopt, the server order you plan to place, the licensing deal you plan to sign. Exposure depends on how you use AI, and the table below sorts the common cases.

How you use AI	Exposure	Why	First move
US or EU frontier models via API only	Low	Not covered by China’s draft	Note it in your vendor register and move on
Chinese models via API or chat app	Low to medium	API access reported as excluded, but final text could change; data handling questions remain	Confirm a fallback model and test it
Self-hosted Chinese open-weight models	Medium	Current weights stay usable; future top-tier releases may not be published openly	Archive the exact weights and licences you rely on
Buying AI servers or GPUs through resellers	Medium	Taiwan’s tighter controls mean more end-use checks and paperwork	Ask suppliers about end-user certificates and lead times
Licensing, JVs or M&A involving Chinese AI or chip technology	High	Registration applies now, including to deals in progress	Get trade counsel to review open deals this month

Delana analysis based on reporting from The Next Web, The National and Tech Insider (16 Oct 2026). Not legal advice.

Two points sit underneath the table. First, model choice is now a supply chain decision, not only a performance and price decision. When we looked at what NIST and Anthropic found in GLM-5.3, the concern was capability; now it is also continuity. Second, a single-vendor stack is fragile whatever the vendor’s nationality. Our guide to stress-testing your AI vendors and the Character.AI bankruptcy both point to the same answer: keep a tested alternative and know how you would switch.

The closer you are to moving technology or ownership across borders, the more the new rules matter.

7. A four-week plan before the 15 November deadline

The consultation window gives you about a month before the final text is due. Use it to learn where you are exposed, so that you are deciding with facts when the final rules arrive. Each week below takes a few hours from IT, procurement and legal.

Week 1 (by 23 October): Find every model

List every AI model in use, including ones embedded in SaaS tools, coding assistants and developer experiments. Note the provider, country of origin and whether you call an API or host the weights.
Flag anything from a Chinese lab, and record the exact version, licence and where the weights are stored.
Check your vendors’ models. Ask key SaaS suppliers which foundation models sit behind their AI features.

Week 2 (by 30 October): Protect continuity

Archive what you depend on. Keep verified copies of open weights, tokenisers and licence texts in storage you control, with checksums.
Name a fallback for each critical model, and run your own evaluation set against it so you know the quality and cost difference before you need it.
Remove hard-coded model names from applications so that switching is a configuration change, not a rebuild.

Week 3 (by 6 November): Check hardware and deals

Ask AI server and GPU suppliers how Taiwan’s planned rules affect lead times, end-user certificates and warranty support.
Review open deals that license Chinese AI or chip technology, or change ownership of a company that holds it. Registration applies now, including to transactions already under way.
Add an export-control clause to new AI contracts covering what happens if a model or component becomes restricted.

Week 4 (by 15 November): Decide and record

Brief leadership in one page: which models and suppliers are exposed, what the fallback is and what it would cost.
Consider a consultation response through a trade association or counsel if your business has a direct stake in the final rules.
Set a review date for when the final text is published, and assign one person to read it against your inventory.

8. The bottom line

Today’s announcements confirm that AI models are now strategic goods, handled by trade ministries as well as technology teams. China wants a say over where its strongest models and chip know-how go. Taiwan is putting criminal penalties behind its chip controls. The United States is expected to respond. None of this stops a business from using AI tomorrow morning, but it does change what a sound AI plan looks like. Know where every model in your stack comes from, keep a copy of what you depend on, and make sure switching is possible. Companies that do that now will not have to rush when the final rules arrive.

Frequently asked questions

What AI technologies does China’s draft export control list cover?

The draft published on 16 October 2026 adds 40 technologies, 23 of them AI-related. Reported items include large language models, AI agents, AI chip designs, computing cluster architectures, model weights and source code above a performance threshold, and design know-how for AI chips at 14 nanometres or below.

Can I still use DeepSeek, Qwen or Kimi?

On current reporting, yes. Access through apps, websites and APIs is reported to be outside the controls, and already published open-source weights are exempt. The open questions are whether future top-tier models will be released openly and whether the final rules keep these exclusions.

Which models would be covered?

The draft names no models and its thresholds have not been published. Bloomberg reports that the controls would apply to models with at least two trillion parameters or that rank in the world’s top ten on performance and safety tests, and that Kimi K3 and Qwen3.8-Max are above the size threshold.

Are the rules in force yet?

The catalogue is a draft open for public comment until 15 November 2026. An interim requirement to register before exporting listed technologies, including exports already under way, took effect on 16 October.

What is Taiwan changing on AI chip exports?

Taiwan plans to add AI servers and advanced semiconductors to its export control list and make unlicensed exports of them to China a criminal offence under its Foreign Trade Act. Its economy minister said some controls will be aligned with US rules under the two governments’ trade agreement. No timeline has been given.

What should businesses do now?

Inventory every AI model you use and where it comes from, archive the open weights and licences you depend on, test a fallback model, ask hardware suppliers about lead times, and have counsel review any open deals involving Chinese AI or chip technology before 15 November.