China’s emerging AI export control regime is beginning to take shape, with two of the country’s most prominent artificial intelligence developers completing the registration process required to send listed technologies abroad. DeepSeek and MiniMax have secured open-source and overseas licensing rights, according to a post from an account affiliated with state media, marking an early test of how Beijing intends to balance national security concerns with global commercial ambitions.
The post, published late on Thursday, outlined how the export controls proposed on 16 October will operate. Its central message was that most AI technology will remain free to export. The controls are not a blanket ban on Chinese AI companies selling abroad. Instead, they create a registration and licensing system for a narrow set of advanced models, agents and computing technologies that the government considers sensitive. Public comments on the draft catalogue close on 15 November.
DeepSeek and MiniMax did not immediately reply to requests for comment. Reports said DeepSeek raised $15 billion from Tencent and CATL this week, a claim that has not been independently verified. The timing is notable: as Chinese AI companies seek capital and overseas customers, the government is also building a legal architecture to govern how their most advanced technology moves across borders.
A targeted list, not a broad embargo
China’s commerce and science ministries want to add 23 AI technologies to the national list of technologies that are banned or restricted from export. The list includes large models, agents and computing clusters. Companies must register before exporting any listed technology. The approach mirrors other Chinese export control regimes for dual-use technologies, where the state retains visibility over transfers while allowing most commercial activity to continue.
The catalogue is significant because it moves AI from a largely unregulated commercial field into a formal export control framework. China is home to a large open-source AI ecosystem, and its developers have become increasingly competitive internationally. The rules appear designed to prevent the most sensitive models and computing systems from being transferred to foreign entities without government knowledge, while avoiding a broad crackdown that would damage the industry’s global prospects.
The draft rules also reflect a wider global trend. Governments are grappling with how to control access to frontier AI without stifling innovation or trade. The United States has focused on restricting advanced chips and chipmaking equipment, while China is now building its own system for model weights, algorithms and computing infrastructure. The result is a fragmented regulatory landscape that companies must navigate country by country.
Data, sensitive uses and computing power
The post said the rules target three areas. The first is data. The rules would ban exports of large models trained on China’s “core data”. They would restrict models trained on “important data”, whatever their size. The distinction matters. In China’s data security framework, “core data” generally refers to data that could harm national security or major public interests if compromised, while “important data” covers a broader category of sensitive information. Models trained on such data can embed patterns, associations or memorized details, making model weights a potential channel for data transfer.
The second area is sensitive use. Agents above a set performance level face restrictions. Agents built mainly for critical information infrastructure, defence or core government affairs face a ban. This is a use-based control rather than a purely technical one. It means the same underlying model could be treated differently depending on what it is designed to do and who is expected to use it.
The third area is computing power. AI chip design and high-speed links for large computing clusters are on the list. These are the building blocks of frontier training. By controlling chip design and interconnect technology, the rules reach beyond models into the hardware layer. The catalogue also sets numerical thresholds for large models: training compute above 1026 floating-point operations, more than 2 trillion parameters, or capability in the global top ten. These thresholds are high, suggesting that only the most advanced systems will face restrictions.
What stays outside the rules
The post offered an important clarification for commercial AI companies. Offering a restricted model through API calls, apps, web pages, SaaS or lightweight local deployment is not considered a technology export. “Simply put, AI companies can continue with their commercialization efforts as usual,” the post said. That means overseas users could still access Chinese models through cloud services, mobile applications and software platforms, even if the underlying model is on the restricted list.
Open-source technology receives automatic licensing, which the post described as “immediate reporting and export”. Chinese companies’ own computing centres abroad, chip tape-out overseas, joint research and work on international standards are also covered by the automatic route. Technology already public through open source or academic papers falls outside the rules. Exports tied to cooperation among members of the World AI Cooperation Organization would receive more support.
This distinction between model access and technology transfer is central to the policy. API calls and hosted services are treated as commercial services, not exports of the model itself. Lightweight local deployment is also excluded, though heavier local deployment may still require review. Open-source releases are encouraged, but they do not override the data restrictions. A model trained on core data would still be banned from export even if its weights are published, according to the post’s description of the rules.
Legal consultation and industry feedback
Liu Jing and Liu Hongmei of Beijing Dacheng Law Offices took part in the consultations, according to the post. They told the account that control points for some AI items changed after the first draft. Many of the concerns companies raised show up in the published version. Liu Hongmei said restricted technologies can still go abroad with a licence, and that this would not disrupt companies’ overseas plans.
The involvement of legal experts and the revision of control points suggest that the government is trying to make the regime workable for industry. The registration process for DeepSeek and MiniMax is an early signal that companies can obtain clearances. It also gives the government a list of approved exporters and a mechanism to monitor future transfers. For companies, registration may become a competitive advantage, because it provides legal certainty for overseas partnerships and licensing deals.
The registration requirement could also create a two-tier market. Larger companies with legal teams and compliance staff may find it easier to navigate the process, while smaller developers may need to rely on open-source releases, API access or partnerships with registered firms. That could accelerate consolidation in China’s AI industry, even as the rules preserve room for commercial experimentation and global distribution.
A contrast with Washington
The post set the rules against the US approach. It cited the Commerce Department’s AI diffusion framework of January 2025. It also cited an “alliance strategy” it attributed to Anthropic’s CEO, Dario Amodei. “But China did not take this path,” the post said. The US framework sought to manage global access to advanced AI chips through a tiered system of countries and export licences. The Chinese approach, by contrast, focuses on registration and licensing for a narrow list of technologies while keeping most AI exports free.
The contrast is partly rhetorical, but it also reflects different policy tools. Washington controls access to advanced semiconductors and chipmaking equipment, where the US has leverage over global supply chains. Beijing is focusing on model weights, algorithms and computing clusters, where Chinese companies have growing capabilities. Both governments are trying to prevent their most advanced technology from strengthening strategic rivals while preserving commercial opportunities for their own firms.
In a separate post on Friday, the account said overseas internet users had urged other Chinese open-source models to follow suit. That detail points to a broader competition over open-source AI leadership. Chinese developers have gained a following among global developers, and the new rules appear designed to keep that ecosystem open while adding a layer of government oversight for the most sensitive systems.
Source: TNW | China News