China’s Draft AI Export Controls: 23 Technologies, Thresholds and a Registration Rule

In short: On October 16, 2026, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) issued Announcement No. 48 of 2026, proposing revisions to the Catalogue of Technologies Prohibited or Restricted from Export. The draft adds 40 technologies, 23 of them in artificial intelligence. Comments are open until November 15, 2026, and a temporary registration requirement for certain transfers applies immediately. The catalogue itself has not yet been amended.

What the announcement says

The proposal covers three sectors: artificial intelligence (23 items), biomedicine (5), and photovoltaics (12). Of the 40 technologies, 14 are proposed as prohibited from export and 26 as restricted, meaning an export license would be required. Reuters and others also reported a separate focus on 12 solar-manufacturing technologies, which is the photovoltaic part of the same package.

The public summaries describe the 23 AI items only by broad categories: large models, intelligent agents, computing clusters, and related information processing. Readers who need the exact list should consult the official text, since secondary summaries do not reproduce every item.

The model thresholds

For large-model “core technology,” the draft lists model weights, key checkpoints, underlying source code, architecture design, training optimization, and inference acceleration. A model reaches the restricted threshold if any one of three conditions is met:

1. Training compute of at least 10²⁶ floating-point operations.
2. At least 2 trillion parameters.
3. “Globally leading” capability, for example a place in the global top ten of public evaluations at the time of export.

The “any one of” structure is important. A smaller model that ranks among the top ten globally would be caught even if it falls below the compute and parameter figures. The third test is also the vaguest. Public evaluations change monthly, and the draft’s wording leaves unanswered who determines the ranking and on which benchmarks. Analysts will watch the final text for a definition.

Examples of prohibited items include AI-driven malicious-code generation and models trained on China’s “core data,” which the accompanying FAQ describes as core national data.

Open-weight models and API access

The most-discussed question is how the draft treats open-weight and open-source models. According to the summary we reviewed, open-source releases and technologies already public through academic papers are proposed to stay outside the controls. At the same time, model weights are listed as restricted core technology. The reconciliation is not obvious from the summaries, which is why the South China Morning Post asked what the proposal means for open-weight technologies, and why Business Standard and ET Datacenters ran similar questions. A Chinese-language financial outlet (Futu) framed the open-source exemption as a benefit to leading internet platforms. The Decoder, by contrast, headlined that China plans to tighten controls on open-weight models and 23 key AI technologies. Different outlets are reading the same draft differently, and that disagreement is itself a signal that the text is ambiguous.

Access through an API, website, app, or SaaS is generally not treated as an export of the underlying technology. Tokenpost summarized the draft as separating model transfers from API access. For service providers that serve foreign customers from Chinese infrastructure, that distinction may be decisive. For companies that deploy a model on their own servers abroad, the answer may be different.

The immediate registration rule

Unlike the catalogue, which is only proposed, the temporary registration requirement is already in force. Parties must register on MOFCOM’s Technology Trade Management Information platform for:

• transfers of ownership,
• licensing for use abroad, and
• changes of owner or controller to a foreign entity.

Pending deals are covered, including transactions under way but not completed. The practical effect is visibility: authorities will see export demand before the final catalogue takes effect. Law360, Geopolitechs, and FindArticles all flagged the interim registration as the item to act on first.

Official rationale and outside readings

A MOFCOM spokesperson described the revision as routine catalogue work under the Foreign Trade Law and said it does not target any country, region, or entity. Stated AI goals include preventing misuse and keeping AI under human control. Global Times and China Daily echoed this, emphasizing national security, safety, and encouragement of open-source development and sharing.

Outside commentary reads the proposal against a backdrop of rivalry with the United States:

• Reuters and the Financial Times describe it as Beijing seeking a tighter grip on AI exports amid U.S. rivalry.
• Digitimes argues the controls reach beyond chips to design IP and computing infrastructure.
• Bruegel asks what lessons Europe should take; ThePrint calls it a warning signal for India.
• Bloomberg and Barron’s pair the story with Taiwan’s moves to curb AI chip exports to China, part of a U.S. trade deal, including reported prison terms for illicit AI server exports.
• The National links it to a U.S.-China trade truce that is fraying.

The picture is of a two-way tightening: the United States and its partners restricting chips going to China, and China now moving to restrict models and related know-how going out.

What search data shows

Google’s US daily trending feed on October 20 was led by sports, entertainment, and election news, and this topic was not in it. That is normal for trade-law news. The audience is lawyers, trade-compliance teams, AI developers, and investors. News volume was nonetheless high: a Google News search returned well over a hundred items across wire services, financial media, trade publications, and outlets in Asia and Europe, with coverage starting October 16 and continuing through at least October 20 with fresh analysis pieces. We did not obtain a keyword time-series, so we do not quantify any search spike.

Who should pay attention

• Chinese AI developers and their foreign partners. Joint ventures, licensing deals, and acquisitions involving model assets could trigger registration.
• Multinationals with Chinese R&D. If model work is done in China for a group that deploys elsewhere, the transfer of weights or code out of China could fall under the rules.
• Customers of Chinese models. Companies that run Chinese open-weight models locally should watch how the final catalogue treats them.
• Investors. Cross-border deals involving Chinese AI assets face more process, and potentially more uncertainty about approval.
• Policy teams. Governments weighing their own AI export rules will study the thresholds.

A practical checklist

1. Inventory planned or in-progress transfers

List any transfers of weights, source code, architecture, or related know-how that could meet the draft thresholds. Include transactions that are not yet closed.

2. Register covered transactions

Where a deal involves a transfer of ownership, licensing for use abroad, or a change of owner or controller to a foreign entity, assess whether registration on MOFCOM’s platform is required, and do so.

3. Review contracts

Check joint-venture, collaboration, and outsourcing agreements for change-of-control triggers and for provisions allowing overseas deployment of restricted models.

4. Document the technical boundary

Write down the difference between local deployment of restricted technology and remote API or SaaS access. If your product architecture relies on that distinction, record it now.

5. Consider commenting

The public comment window closes on November 15, 2026. Industry groups and affected companies have a chance to ask for definitions of “globally leading,” clarity on the open-source carve-out, and facilitation measures.

6. Track the final text

Compare the final catalogue with the draft thresholds and prohibited items when it appears.

Unresolved questions

• The precise meaning of the compute and parameter thresholds, and how they will be measured.
• The definition of “globally leading” capability and who sets it.
• How the open-source exemption interacts with weights listed as restricted core technology.
• The effective date of the final catalogue.
• Facilitation measures for open-source releases, overseas self-operated compute centers, and international standards work.
• How parallel U.S. and allied controls on AI and compute will interact with the Chinese regime.

Connecting the dots

The same week, the U.S. Department of Homeland Security ordered DeepSeek products removed from civilian federal systems, and the General Services Administration’s new clause on safeguarding government data in large-language-model systems took effect. Together with Beijing’s draft, these actions show models becoming regulated cross-border goods. Buyers need provenance, sellers need export analysis, and lawyers need to follow both sides of the border. A company that treats AI models as ordinary software may find that contracts, customs analysis, and procurement rules treat them differently.

How to read a draft like this

Draft export catalogues often change between proposal and final text. Thresholds get clarified, categories are merged or dropped, and transition periods appear. Planning should therefore assume the direction of travel, not the exact wording. The direction here is clear: model weights, agents, and the infrastructure used to train and run them are being placed on a managed-export list. A prudent company will build a record now of where its model assets sit, who controls them, and how they reach customers abroad. That record serves compliance with this proposal and any similar regime elsewhere, and it makes a future licence application faster if one is needed.

Bottom line

Treat the Chinese draft as a proposal with an immediate registration component. The catalogue may change after the comment period, but the interim registration rule is operating now. Start with an inventory of transfers, confirm whether any deals need registration, and keep an eye on the three definitional questions that matter most: compute, parameters, and “globally leading.” Everything else depends on the final text.

Sources reviewed: the Kurums law note of October 20, 2026 (a secondary summary of MOFCOM and MOST Announcement No. 48, which should be checked against the official text); Google News coverage from Reuters, the Financial Times, SCMP, Bloomberg, Global Times, Digitimes, Bruegel and others; and Google’s US daily trending-searches feed. This is analysis, not legal advice.