China’s Ministry of Commerce and Ministry of Science and Technology on October 16, 2026, proposed sweeping new export controls covering 40 technologies across artificial intelligence, biomedicine, and photovoltaics, according to the official draft catalogue and reporting from Bloomberg. Twenty-three of the listed items are AI-specific, spanning large language models, autonomous agents, chip design tools, computing clusters, and information-processing systems, a scope that puts Beijing’s rulemaking on a near-collision course with Washington’s own chip and model restrictions.
The timing matters as much as the substance. China has spent the past three years absorbing US export bans on advanced GPUs, EDA software, and chipmaking tools. Now it is building a mirror-image system of its own, one that could govern who outside China gets access to Chinese-made AI models, chips, and the code and weights behind them. The public comment window closes November 15, 2026, giving companies, trade groups, and foreign governments one month to respond before Beijing finalizes the rules.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What China’s Draft Export Control Catalogue Actually Covers
The draft revises China’s existing catalogue of technologies prohibited or restricted from export, adding 40 new entries split across three sectors. Within that total, 23 technologies fall under the AI umbrella, according to the MLex account of the filing. The listed categories include large AI models, intelligent agents, AI chip design methodologies, computing cluster architectures, and information processing technologies, a grouping broad enough to touch nearly every layer of the modern AI stack, from the silicon up through the software that runs on it.
What sets this proposal apart from a routine update is the explicit inclusion of model weights and source code as potentially controlled material, once a product crosses specified performance thresholds. That detail signals Beijing is treating the parameters inside a trained model the same way it has long treated physical hardware designs: as something that can be licensed, registered, or blocked from leaving the country. For any company training frontier models on Chinese soil, or using Chinese open-weight releases as a base for commercial products abroad, that is the single most consequential line in the draft.
Fourteen Prohibitions, Twenty-Six Restrictions
The 40 technologies split into two tiers of severity. Fourteen would be banned outright from export, a list that includes AI-driven malicious code generation and mutation technology, plus AI models trained on what the filing describes as Chinese core data. Twenty-six others would be restricted rather than banned, covering technologies the government links to hacking capabilities. Restricted items generally require a license before they can leave the country; prohibited items cannot be exported at all, regardless of licensing.
| Category | Count | Treatment |
|---|---|---|
| Total technologies in draft catalogue | 40 | Mixed: AI, biomedicine, photovoltaics |
| AI-related technologies | 23 | Large models, agents, chip design, compute clusters, info processing |
| Technologies proposed for outright prohibition | 14 | No export permitted, including malicious-code generation tech |
| Technologies proposed for restriction | 26 | License required, includes hacking-capable technologies |
| Public comment deadline | Nov. 15, 2026 | Open for industry and public submissions |
| Lead agencies | 2 | Ministry of Commerce, Ministry of Science and Technology |
Notice what isn’t in that 14-item prohibition list: generic large language models or chip designs aren’t banned outright. Beijing’s harshest language targets specific abuse cases, malicious code generation and models trained on sensitive domestic data, while routing broader AI model and chip-design technology into the restricted, licensable category. That is a meaningfully softer posture than an across-the-board export ban, and it leaves room for continued commercial AI exports once a license is granted.
The Interim Registration Rule Takes Effect Immediately
While the catalogue itself remains a draft subject to a month of public comment, Beijing has already put a temporary measure in place. Exporters must register covered technology transfers before export while the proposed catalogue changes are under review, according to the Geopolitechs translation of the official notice. That interim step means the registration requirement is live now, even though the underlying prohibition and restriction lists haven’t been finalized.
The registration scope is wide. It covers outright technology-ownership transfers, overseas licensing agreements, changes in ownership that would place China-based technology under foreign control, and exports that have already started but haven’t yet completed. In practice, that last category means deals already in motion, joint ventures, licensing agreements signed months ago, cross-border model-weight transfers mid-process, now need to be registered with Chinese authorities even though they predate the proposed rule.
What Falls Outside the Controls
Two carve-outs matter for how narrowly or broadly this will actually bite. First, technologies already in the public domain through published academic papers or open-source releases are excluded from the controls. Second, and more significant for the global AI industry, the proposal reportedly does not cover access through APIs, websites, apps, or software-as-a-service platforms, according to coverage cited in the fact pattern around the announcement. That means a foreign company calling a Chinese AI model through a hosted API, rather than downloading the weights and running them locally, would likely sit outside the scope of the new export registration and licensing regime as drafted.
That SaaS exclusion is a meaningful design choice. It suggests Beijing’s primary concern is the physical or digital transfer of the technology itself, the weights, the source code, the chip-design IP, rather than foreign usage of Chinese AI services delivered over the internet. Companies such as those behind DeepSeek’s hosted API or Zhipu’s GLM models could continue serving international customers through cloud endpoints even as weight downloads and source-code transfers face new friction.
Why Beijing Is Doing This Now
A spokesperson for China’s Ministry of Commerce framed the proposed measures as an effort to ensure that AI technology is “safe and controllable,” language that echoes Beijing’s long-running domestic push to keep frontier AI development within a regulatory perimeter it can monitor. That framing lines up with China’s parallel efforts over 2025 and 2026 to require security reviews for domestically deployed large models, and it extends that logic outward to anything crossing the border.
There’s also a clear geopolitical mirroring at work. Washington has restricted advanced AI chip exports to China through its Commerce Department’s Entity List and a series of expanding controls on GPU sales, EDA software, and chipmaking equipment since 2022. China’s new catalogue effectively builds the same kind of lever, but pointed the other direction: toward model weights, training data provenance, and AI chip-design methodology rather than physical silicon alone. Where the US controls what advanced hardware reaches China, China’s draft controls what AI software, weights, and design know-how can leave it.
How This Compares With Existing US and Allied AI Export Rules
| Dimension | China’s Oct. 2026 Draft Catalogue | Existing US Export Control Regime |
|---|---|---|
| Primary target | AI models, agents, chip design, model weights, source code | Advanced GPUs, EDA software, chipmaking tools |
| Legal mechanism | Export control catalogue (Commerce + Science and Technology ministries) | Entity List and Export Administration Regulations (Commerce Department) |
| Hosted-service treatment | APIs, apps, websites, SaaS reportedly excluded | Cloud and API access has faced separate, evolving compute-threshold rules |
| Open-source treatment | Published papers and open-source releases excluded | No blanket open-source exemption; case-by-case licensing |
| Current stage | Draft, open for comment until Nov. 15, 2026 | Multiple rounds already in force since 2022 |
The two regimes aren’t symmetrical, but the direction of travel is similar: both governments are narrowing the aperture through which advanced AI capability can move across borders, just from opposite ends of the supply chain. Washington has focused on hardware and the tools used to build hardware. Beijing’s draft reaches further up the stack, into the trained models and the code that makes them run, an area where US rules have moved more slowly and unevenly.
Historical Context: From Entity Lists to Model Weights
China has maintained an export control catalogue in some form for years, covering sensitive dual-use technologies well before AI became a flashpoint. What’s new in the October 2026 proposal is the direct naming of large AI models, intelligent agents, and model weights as controllable items in their own right, not just the chips or infrastructure beneath them. That’s a notable evolution from earlier rounds of Chinese tech export policy, which focused more narrowly on hardware, encryption, and select software categories.
The timing also follows a period in which Chinese open-weight models, from DeepSeek’s releases to Zhipu’s GLM family and Alibaba’s Qwen line, gained significant global adoption, partly because they were freely downloadable without the kind of licensing friction Western commercial models carry. If finalized, the new catalogue would formalize a path for Beijing to apply selective controls to future high-performance releases, even as it preserves the open-source exemption for technology already published. Readers tracking the broader shift in Chinese open-source AI infrastructure will recognize this as a continuation of a theme: open models as both strategic asset and strategic risk.
Market Impact: Chips, Clusters, and the Compute Supply Chain
AI chip and compute-cluster technology sits inside the 23-item AI list, which means the proposal potentially touches the design methodologies behind Chinese AI accelerators, not just the finished models running on them. That lands at a sensitive moment for the global chip supply chain, with memory shortages already squeezing AI hardware budgets and Nvidia’s position in China under pressure from domestic alternatives. Readers following the shift in Nvidia’s China market share as Huawei’s chips gain ground will want to watch whether Beijing’s new rules end up protecting that domestic chip-design advantage by restricting how Chinese design IP can be licensed abroad.
Large compute deals are also part of the backdrop. Cross-border AI infrastructure agreements, like the kind of GPU supply arrangements seen in the Tencent-Oracle chip pricing deal, depend on a predictable rulebook for what technology can move where. A new Chinese registration requirement covering technology-ownership transfers and overseas licensing adds a compliance step to any arrangement that involves Chinese-origin AI chip design or compute-cluster architecture, even if the deal itself is structured through non-Chinese intermediaries.
Memory supply is a related pressure point. The ongoing HBM4 memory shortage has already pushed AI hardware costs higher across the industry. Export-control friction on the chip-design side of China’s AI stack could compound that squeeze for companies trying to source or build AI accelerators that depend on both advanced memory and Chinese-origin design methodology.
Impact on AI Model Competition and Licensing
The model-weights provision is the part of the proposal most likely to ripple through the competitive landscape among large language models. Chinese labs releasing frontier-class models, in the vein of comparisons like GLM 5.3 Prime versus Qwen3.8 Max, have built global developer mindshare partly on permissive weight downloads. If a future model crosses the performance thresholds the draft references, exporting its weights or source code outside China could require registration and, potentially, a license, a meaningfully different experience for international developers than simply pulling a model from Hugging Face.
That creates a bifurcated path for Chinese AI companies: keep a model below the undefined performance threshold and distribute weights freely, or push past it and accept new export friction, unless the model is routed through the SaaS/API exclusion instead of a weights download. Expect Chinese AI labs to watch the threshold definition closely once the final catalogue is published, since it will shape decisions about open-weight release strategy for 2027 model generations.
Below is a simplified view of how the proposed registration and licensing flow is structured, based on the draft’s description of scope:
Proposed AI export control flow (draft, subject to Nov. 15 comment period):
1. Technology falls under one of 23 AI categories?
-> No: standard export rules apply
-> Yes: continue
2. Is the technology already public (published paper / open-source release)?
-> Yes: excluded from controls
-> No: continue
3. Is the transfer via API / app / website / SaaS access only?
-> Yes: reportedly outside scope of controls
-> No: continue
4. Does it match one of the 14 prohibited items
(e.g. malicious code generation/mutation tech, models trained on core data)?
-> Yes: export prohibited
-> No: continue
5. Does it match one of the 26 restricted items (e.g. hacking-capable tech)?
-> Yes: license required; interim registration mandatory now
-> No: standard registration/review appliesWhat Industry and Trade Groups Are Watching For
With the comment period open until November 15, the most consequential unknowns are the exact performance thresholds that trigger model-weight and source-code controls, and how “AI chip design” will be defined in the final text. Industry trackers covering the filing, including AI Weekly, have flagged the weights inclusion as the detail most likely to draw formal comments from Chinese AI firms and multinational partners during the review window.
Multinational semiconductor and cloud companies with joint ventures or technology-licensing arrangements inside China will also be parsing the registration requirement’s “ownership change” language closely, since it could capture restructurings, spinoffs, or stake sales that shift a China-based AI technology under majority foreign control, even where no technology physically crosses a border.
Timeline: From Draft to Final Rule
The clock on this proposal is short by the standards of major trade policy. The draft catalogue was published October 16, 2026, and the comment window closes exactly one month later, on November 15, 2026. That compressed schedule leaves little room for extended back-and-forth before Beijing moves toward a final version, especially compared with the multi-month or multi-year review cycles that have accompanied some prior Chinese regulatory overhauls in tech and data policy.
In the meantime, the interim registration requirement already applies, which means the practical compliance burden on exporters starts well before the catalogue itself is locked in. Companies with AI technology transfers in progress, whether that’s a licensing deal, a joint venture restructuring, or a mid-flight export, need to register now rather than wait for the final rule, since the registration scope explicitly includes exports “already started but not completed.”
How the Story Is Being Covered Across Outlets
Coverage of the proposal spread quickly across financial and trade-focused outlets within hours of the announcement. The Economic Times framed the move as part of a broader pattern of China tightening its grip on strategically sensitive technology exports, alongside biotech and solar. That framing echoes coverage from Bloomberg and MLex, both of which emphasized the AI-specific portion of the catalogue as the section most likely to draw international attention and formal objections during the comment period.
The breadth of coverage, spanning financial newswires, specialist AI policy trackers, and regional business outlets, within a single day of the announcement underscores how closely global markets are watching any move that touches the AI supply chain. Given how sensitive AI chip and infrastructure stocks have proven to export-control headlines out of both Washington and Beijing over the past several years, expect continued scrutiny of every procedural step between now and the November 15 deadline.
Predictions: Where This Goes From Here
- Expect a wave of formal comments from Chinese AI labs and multinational chip and cloud firms before the November 15 deadline, concentrated on the undefined “performance thresholds” for model weights and source code.
- The SaaS and API exclusion will likely survive into the final rule largely intact, since closing it would disrupt the hosted-API business model that Chinese AI labs are actively using to monetize overseas usage.
- Watch for a parallel or follow-on notice clarifying exactly what counts as “AI chip design methodology,” since that term currently covers a wide and commercially sensitive range of IP.
- Expect Washington and allied governments to cite the proposal as justification for maintaining or tightening their own existing AI and semiconductor export restrictions on China.
- Expect at least one high-profile cross-border AI licensing or compute deal to be delayed or restructured around the interim registration requirement before the catalogue is finalized.
What This Means for Global AI Developers and Companies
For developers and companies outside China, the near-term practical takeaway is narrow but important: using a Chinese AI model through a hosted API or app likely remains unaffected under the draft as described, while downloading weights, licensing source code, or acquiring a stake in a China-based AI technology provider now sits under a registration requirement that is already in effect, ahead of the catalogue’s final form. Companies that rely on Chinese open-weight models as the base layer for commercial products, rather than consuming them via API, have the clearest incentive to track the thresholds that will determine which future releases fall under restriction.
The broader signal is that AI export control is no longer a one-directional story about Western restrictions on Chinese hardware access. Beijing is now building its own regulatory lever over the software, weights, and design know-how that make frontier AI possible, and the two regimes, American and Chinese, are increasingly shaping the same global AI supply chain from opposite directions.
Frequently Asked Questions
What did China propose on October 16, 2026?
China’s Ministry of Commerce and Ministry of Science and Technology proposed adding 40 technologies across AI, biomedicine, and photovoltaics to the country’s catalogue of technologies prohibited or restricted from export, with 23 of those technologies specifically AI-related.
Which AI technologies are covered by China’s export control proposal?
The 23 AI technologies include large AI models, intelligent agents, AI chip design, computing clusters, and information processing technologies, along with model weights and source code for products exceeding specified performance thresholds.
What is the difference between prohibited and restricted technologies in the draft?
Fourteen technologies would be banned from export outright, including AI-driven malicious code generation and mutation technology and models trained on Chinese core data. Twenty-six other technologies, including those with hacking capabilities, would be restricted, meaning they could still be exported with a license.
When does the public comment period end?
The public comment deadline is November 15, 2026. The 40-technology catalogue remains a draft until the comment period closes and the ministries finalize the rules.
Does this affect companies that use Chinese AI models through an API?
According to reporting on the proposal, access through APIs, websites, apps, or software-as-a-service platforms is not covered by the controls. The focus is on direct transfers of technology such as model weights, source code, and chip design methodology.
Are open-source AI models exempt from the proposed controls?
Technologies already in the public domain through published academic papers or open-source releases are excluded from the proposed controls, according to the draft.
Is there already a rule in effect, or is everything still a draft?
The catalogue of prohibited and restricted technologies is still a draft open for comment. However, an interim measure is already in effect: exporters must register covered technology transfers before export while the proposed changes remain under review.
How does this compare to US export controls on AI technology?
US export controls, enforced largely through the Commerce Department’s Entity List and Export Administration Regulations since 2022, have focused primarily on restricting advanced AI chips, EDA software, and chipmaking equipment from reaching China. China’s draft catalogue runs in the opposite direction, targeting AI models, agents, chip design methodology, and model weights that could leave China, rather than hardware entering it.
