China has opened a consultation that could reshape how specified AI technology moves across its borders, while immediately requiring registration for covered transfers. The practical change is not a shutdown of Chinese models for overseas users. It is a new line between transferring particular technology out of China and letting a customer interact with a model through ordinary online services.
MOFCOM and the Ministry of Science and Technology, the agencies behind Announcement No. 48, propose revising China’s Catalog of Technologies Prohibited or Restricted from Export. Their consultation covers 40 proposed technologies across AI, biomedicine, and photovoltaics, including 23 AI-related technologies. Reuters independently reported the same core distinction: the proposal targets defined technology transfers, not routine access through an API, website, app, or SaaS product. That difference turns the notice into an immediate planning issue for developers, cloud operators, chip designers, investors, and parties arranging cross-border intellectual-property deals.
Temporary Registration Starts Before the AI Catalog Is Final
The consultation itself is not a final revision to the export-technology catalog. Public comments run through November 15 and can address proposed technology scope, technical indicators, control policies, and facilitation measures. MOFCOM and MOST have left the final shape of the catalog open to that process.
Yet the interim mechanism is already active. From publication of the announcement until a revised catalog is formally implemented, an operator planning to export technology listed in the notice’s annexes must register through MOFCOM’s Technology Trade Management Information Application before the export. The notice also reaches transactions that have begun but are not complete. For businesses, that timing is the part that matters today: a final classification may be pending, while the registration step for annex-listed technology transfers is not.
The notice gives the requirement a wider reach than a simple sale of software or a physical shipment of servers. It expressly covers transferring technology ownership overseas, including intellectual-property transfers; licensing technology held in China for overseas use; and a change that makes the owner or actual controller of China-held technology an overseas entity. MOFCOM can specify additional cases by notice. Registration information includes the exporter and recipient, destination, technology and intended use, and estimated export value.
That means the immediate question for a China-based company or a deal counterparty is whether a planned transaction fits the annex, not whether an overseas customer can continue clicking into a public-facing model service. An IP transfer, license for overseas use, or change that makes the owner or actual controller of technology held in China an overseas entity can raise a more immediate question under the notice than the familiar act of making a model query. The policy vocabulary will be familiar from EastFrontier coverage of Chinese AI-model and chip export controls and the wider export-control debate.
MOFCOM Draws the Line Between Transfers and Model Access
MOFCOM’s spokesperson frames the central test as whether restricted technology moves from within China to outside China. Under that explanation, localizing or deploying a restricted large model overseas falls within technology-export management. Deploying one on a company-owned overseas cloud may qualify for an automatic-license facilitation, according to the Q&A accompanying the proposal.
The same Q&A makes an important distinction for overseas users and for businesses whose products sit on top of Chinese models. API calls, interaction through websites or apps, SaaS applications, and on-device lightweight deployment using a restricted large model do not fall within large-model technology-export management, MOFCOM says. The statement is narrow but meaningful. It keeps the story centered on transfer of the technology itself rather than on every mode through which a user might reach a model.
There are boundaries on the other side as well. The primary notice says technology that has entered the public domain through open-source release or academic publication would not be managed. In its Q&A, MOFCOM also described facilitation measures for certain exports, including open-source AI, overseas computing centers built and operated by Chinese companies, offshore chip tape-out and manufacturing, overseas joint research and development, and work on international standards. Those are official policy positions laid out alongside a proposal, not proof that any individual project is automatically outside management.
A rule focused on weights, architecture, clusters, and overseas deployment is different from one aimed at ordinary model consumption. The draft is concerned with defined technical capabilities and their movement, not with a general prohibition on public access to Chinese AI tools.
Draft Thresholds Favor Technical Criteria Over Brand Names
The proposed annex is detailed enough to show that the 23 AI-related technologies are not 23 models or products. It points to technologies connected with large models, AI agents, computing clusters, and information processing. The official Q&A divides all 40 proposed additions into 14 proposed prohibited-export technologies and 26 proposed restricted-export technologies. That split is also a reminder not to describe every AI item as prohibited, or every one as merely restricted.
Selective examples show the approach. Proposed prohibited entries include a large model trained on China’s core data and certain high-performance AI agents aimed principally at sensitive infrastructure, defense-industry, or core-government settings. Proposed restricted entries include AI-assisted vulnerability discovery. The draft annex also covers large-model core technology such as weights, key checkpoints, source code, architecture, and training or inference techniques when listed thresholds are met.
Those thresholds are technical rather than company-specific. One proposed restricted category, for example, lists alternatives that include at least 10^26 floating-point operations, at least 2 trillion parameters, or globally leading capability as defined through contemporaneous public top-10 assessments. A proposed large intelligent-computing-cluster category covers technologies such as high-speed interconnects, distributed-training fault tolerance, communication scheduling, and heterogeneous-cluster inference or scheduling. The draft defines a large cluster as more than 10,000 AI accelerator cards.
MOFCOM says the proposed revision is routine technology-trade management meant to reduce loss of control or inappropriate diffusion while preserving normal international exchange and cooperation. It also says the measure is not directed at a particular country, region, or entity. Those are the government’s stated reasons, not an independent finding about the proposal’s eventual effects.
The next real test arrives when the comment window closes on November 15. The final revision will show whether the AI categories, technical indicators, and facilitation measures survive intact, change, or disappear. Until then, the clearest takeaway is unusually concrete: a consultation remains open, but temporary registration has begun for covered annex technology exports, and MOFCOM’s own distinction leaves ordinary API, web, app, SaaS, and lightweight on-device model access outside large-model technology-export management.