On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48, a draft revision to the catalogue of technologies prohibited or restricted from export. The revision adds 40 technologies across artificial intelligence, biotech, and solar. Twenty-three of the forty are AI-related, covering large models, agents, computing clusters, information processing, and chip design.
The announcement was widely reported as China's first frontier AI export control regime. That description is accurate but incomplete. The more consequential detail is in the implementation mechanics: the temporary registration requirement took effect on the day the announcement was published, before the rules themselves were finalized. Every covered technology transfer now has to be disclosed to the government — the technology, the parties, the destination, the end use, and the value — regardless of how the final catalogue reads.
That sequencing tells you what kind of instrument this is. It functions as a registration system with a gate that can be tightened later, not as an immediate ban.
What It Covers
The draft divides covered AI technologies into prohibited and restricted categories.
Prohibited, no export: models trained on China's "core data"; high-performance agents built for critical infrastructure, defense, or core government systems; AI systems with offensive cyber capability; and high-accuracy detection systems for machine-generated content with precision and recall above 98 percent.
Restricted, license required: foundation models that meet any one of three thresholds — training compute of 10²⁶ FLOPs or more, 2 trillion or more parameters, or a top-10 global ranking on a qualifying public evaluation. Models trained on "important data" also fall into this category. Alongside the models, the restricted list covers agents meeting quantified capability thresholds (sub-task chains of 20 or more steps, 60 or more concurrent agents, integration with 200 or more tools), AI chip designs at 14 nanometers or below, and computing clusters exceeding 10,000 accelerators.
The three-threshold structure is the part worth pausing on. A model qualifies for restriction if it crosses any one of them. Training compute, parameter count, and leaderboard position are three different ways of measuring whether a model is at the frontier. The draft's authors are not defining a specific capability. They are making sure that any model plausibly described as frontier-class is captured by at least one of the definitions.
Whether specific models — Kimi K3 at 2.8 trillion parameters, Qwen3.8-Max at 2.4 trillion — are formally listed has not been announced. Both exceed the parameter threshold and are expected to fall within the restricted category. The actual designation depends on a second test: what data they were trained on.
The Data Test
The draft's most distinctive feature is that the prohibited-versus-restricted distinction turns on training data, not on model capability.
Models trained on core data — the strictest tier under China's Data Security Law, reserved for data tied to national security and the "lifelines" of the economy — are prohibited from export outright. Models trained on important data — the middle tier, covering data whose leak or misuse could harm national security, the economy, public health, or public order — are restricted.
This is a different logic from the American model. U.S. export controls on AI have been organized around hardware: which chips can be sold, to whom, under what conditions. The Chinese draft is organized around the information that went into the model. The question is not how capable the model is, but where its training data came from.
There is a practical exemption: if a model uses no more than 0.1 percent important data, does not materially improve its capability from that data, has had the data de-identified and cleaned, and does not reproduce it in outputs, the model is not covered. The threshold is low enough that most models trained on public data plus a small amount of proprietary data will not be captured.
The data test is the part of the regime that has no American equivalent. The U.S. restricts exports based on what a model can do. China restricts exports based on what a model was built from.
What It Doesn't Cover
The draft's coverage is narrower than the headline count of 23 AI technologies suggests.
API access is exempt. Foreign users calling Chinese models through application programming interfaces, web interfaces, mobile apps, or software-as-a-service platforms do not trigger export licensing. A foreign business can purchase API tokens from DeepSeek, Qwen, or MiniMax without filing anything with the Chinese government. The Ministry of Commerce's spokesperson stated this explicitly: services provided through official APIs and mobile applications are not covered by the new export restrictions.
Lightweight on-device deployment is exempt. A quantized model running on a laptop or a phone is not treated as an export of the underlying model technology.
Technology already in the public domain is exempt. Models that have been open-sourced, and technical methods disclosed in published papers, fall outside the prohibited category. The existing weights and code of already-open models are not affected.
Overseas deployment is not exempt. This is the line that matters. Deploying a restricted model on infrastructure outside China — a foreign data center, a foreign cloud region — does constitute a technology export and requires a license. If the deployment is on the Chinese company's own overseas cloud, under conditions the government considers risk-controllable, automatic licensing applies.
The distinction the draft draws is between transferring the technology and selling access to a service. Weights, checkpoints, source code, architecture, training methods, and chip designs are treated as technology. An API call is a service. The first is regulated. The second is not.
For companies deploying models on their own overseas infrastructure, a separate category of automatic licensing applies: open-source releases of restricted models, overseas self-built computing centers, overseas chip fabrication, joint research using foreign resources, and participation in international standards setting. Automatic licensing still requires a filing, but approval is described as procedural and near-immediate.
The Design Logic
The Takshashila Institution's analysis of the announcement described the restrictions as designed "more as a switch than a wall." Almost nothing changes for exporters today. What China now has is the means to monitor, slow, or stop the flow of AI technology abroad.
That characterization is supported by the structure. The registration requirement is live. The catalogue is not final. The public comment period runs until November 15, 2026. The exemptions are broad enough that most current export activity — API calls, SaaS, open-source downloads of already-released models — continues unaffected. But every covered transfer is now on record, and the automatic licensing route can be narrowed or withdrawn.
The structural symmetry with American model-weight controls has been noted by analysts. Both governments have concluded that frontier model weights are a strategic asset requiring some form of control. The substantive difference is in what each regime targets. The U.S. controls chips and closed model weights. China controls the physical transfer of model technology while leaving service access open. One regime restricts the hardware that makes models possible; the other restricts the files that constitute the model, while allowing the model's output to flow freely.
What the First Filings Show
The most telling data point came a week after the announcement. CCTV's Yuyuantantian reported that DeepSeek and MiniMax had completed their registration filings and received clearance for open-source releases and overseas authorization of "some recently released and soon-to-be-released models."
The signal is in the default. The two most prominent Chinese open-weight model developers went through the process and received permission to release. The regime's default state is not prohibition. It is registration followed by clearance.
That outcome was not guaranteed. The draft could have been interpreted to require prior approval for every open-weight release, which would have frozen the open-source pipeline that made Chinese models the default choice for a large share of global developers. The first filings suggest the government intends the automatic licensing route to function as described: filing, then release.
Whether that holds for models above the thresholds is the open question. Both Kimi K3 and Qwen3.8-Max exceed the parameter threshold. Whether their open-weight releases proceed under automatic licensing, and how quickly, will determine whether the regime functions as a registration system or as a brake.
The Practical Impact
For global developers, the immediate impact is limited. API access to Chinese models is unchanged. Downloaded open-weight models are unchanged. SaaS and app-based access is unchanged.
The friction appears at three points. First, new model releases above the thresholds require registration before open-weight distribution. The filing is procedural, but it is a step that did not exist before October 16. Second, overseas deployment on non-Chinese infrastructure requires a license. A company that wants to run a restricted Chinese model in its own data center outside China is now inside the export control regime. Third, technology transfer transactions — licensing deals, IP transfers, changes in foreign ownership — are covered, including deals already underway but not completed.
For Chinese AI companies, the regime creates a compliance function that did not previously exist. A legal team that previously focused on data privacy and terms of service now has to file technology export registrations, track which models cross which thresholds, and determine whether a given deployment counts as a technology transfer or a service delivery.
For the market, the most consequential effect may be on the economics of open-weight releases. Chinese labs built their global adoption on giving away weights. The draft does not stop that. It makes it a licensed activity with a filing requirement. The cost of compliance is small relative to the benefit of global adoption. The risk is that the automatic licensing route becomes slower or narrower over time, and that the filing requirement becomes a de facto approval requirement for models the government does not want released.
What the Regime Reveals
The Chinese draft is best understood as an extension of a logic China has been building for years: data sovereignty applied to the model layer.
The Data Security Law established three tiers of data. The export control catalogue now applies those tiers to the models trained on that data. A model is not just a piece of software. It is a repository of the information that shaped it. If that information is sensitive, the model inherits the sensitivity.
The U.S. controls the hardware that makes frontier models possible, on the theory that without advanced chips, rivals cannot build them. China controls the models themselves, on the theory that the models embody data that should not leave.
Both theories have gaps. The U.S. theory has a cloud-access gap: compute can be rented abroad. The Chinese theory has a service gap: a model's capabilities can be accessed through an API without the weights ever crossing a border.
The draft does not close either gap. It creates a registration system that makes the flow of model technology visible, and a licensing mechanism that can be tightened. Whether it is used as a monitoring tool or as a brake depends on decisions not yet made.
The comment period ends November 15. The first filings have been approved. The switch exists. Whether it gets flipped is a separate question.
Sources: Ministry of Commerce and Ministry of Science and Technology Announcement No. 48 (October 16, 2026); Ministry of Commerce Spokesperson Q&A via Xinhua Finance (October 16, 2026); King & Wood Mallesons analysis (October 22, 2026); JTN Law Firm analysis (October 22, 2026); Takshashila Institution "China's AI Export Order" (October 18, 2026); The Decoder (October 16, 2026); South China Morning Post (October 16, 2026); East Asia Brief (October 20 and 24, 2026); Cloud Security Alliance Labs (October 23, 2026); CCTV Yuyuantantian via Lianhe Zaobao (October 24, 2026) and Hong Kong Economic Times (October 23, 2026); 21st Century Business Herald (October 16, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on the published draft, official statements, and law-firm commentary available as of the article date.
The catalogue is a draft open for public comment until November 15, 2026; final scope and thresholds may change.
No official list of covered models has been published, so which specific models fall inside the regime remains undetermined.
Parameter counts for named models come from third-party reporting, not from government filings.
Reporting standards differ across jurisdictions; figures may not be directly comparable.
Market size and contract value figures come from third-party research and are estimates.
This sector is changing quickly; claims can become outdated soon.
Sources
- Ministry of Commerce and Ministry of Science and Technology Announcement No. 48 (October 16, 2026)
- Ministry of Commerce Spokesperson Q&A via Xinhua Finance (October 16, 2026)
- King & Wood Mallesons analysis (October 22, 2026)
- JTN Law Firm analysis (October 22, 2026)
- Takshashila Institution "China's AI Export Order" (October 18, 2026)
- The Decoder (October 16, 2026)
- South China Morning Post (October 16, 2026)
- East Asia Brief (October 20 and 24, 2026)
- Cloud Security Alliance Labs (October 23, 2026)
- CCTV Yuyuantantian via Lianhe Zaobao (October 24, 2026) and Hong Kong Economic Times (October 23, 2026)
- 21st Century Business Herald (October 16, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on the published draft, official statements, and law-firm commentary available as of the article date.; The catalogue is a draft open for public comment until November 15, 2026; final scope and thresholds may change.; No official list of covered models has been published, so which specific models fall inside the regime remains undetermined.; Parameter counts for named models come from third-party reporting, not from government filings.; Reporting standards differ across jurisdictions; figures may not be directly comparable.; Market size and contract value figures come from third-party research and are estimates.; This sector is changing quickly; claims can become outdated soon.