Yuyuan Tantian explains the rumors of China's AI technology export ban

Over the past two days, a claim has been circulating in the market and among overseas policy research circles: Is China about to close the door on AI technology exports? This question can be traced back to October 16. On that day, the Ministry of Commerce and the Ministry of Science and Technology publicly solicited opinions on the revision of the Catalogue of Technologies Prohibited or Restricted from Export (hereinafter referred to as the Catalogue). Among the items were 23 AI-related technologies, including large models, agents, computing clusters, and information processing. After the Catalogue was released, some people raised questions: China emphasizes technological openness, so why suddenly impose "export restrictions" on AI? Is this a signal of a policy shift? Tan Zhu has learned that since the policy was announced, two leading companies, DeepSeek and MiniMax, have smoothly completed registration procedures and finished open-sourcing and overseas authorized use. In the first half of this year, relevant departments began drafting the consultation version of this Catalogue and, beforehand, extensively organized research institutions, industry organizations, and leading companies in the sector to conduct in-depth assessments. Lawyers Liu Jing and Liu Hongmei of Beijing Dacheng Law Offices participated in these discussions. They shared a detail with Tan Zhu: compared with the initial draft from several months earlier, the control points for some AI technology items kept being adjusted in line with the latest industry developments. Many of the concerns and suggestions raised by companies at the meetings were well reflected in the policy released this time. What they care about more now is whether there will be clearer operational guidelines going forward. Another detail is that in the relevant Q&A, the Ministry of Commerce gave noticeably more and more specific explanations of "what is not regulated" than of "what is regulated." This also shows that the core of the Catalogue is to delineate boundaries, rather than simply restrict AI technology exports. Tan Zhu has also sorted out several key points to help everyone understand how this set of rules works. What exactly does this set of rules regulate? People who do not work in technology can easily confuse "technology exports" with "product exports." Some may ask: if a company uses a restricted large model to provide API calls, APP/webpage interaction, SaaS applications, and the like to external users, does that fall within the scope of large-model technology export management? The answer is clear: it does not. The focus of this regulation is mainly on three types of situations: First, data security. Large models that provide services after being trained on China's core data are proposed to be prohibited from export; large models that provide services after being trained on important data are proposed to be restricted from export. Simply put, as long as the model training process involves concerns about national data security, the model will be subject to export constraints regardless of whether it is large or small or how many parameters it has. Second, sensitive uses. For example, agent technology that reaches a certain level of performance is itself proposed to be restricted from export; but if this technology is specifically or mainly used in highly sensitive scenarios such as national critical information infrastructure, national defense and military industry, and core government affairs, its nature changes and it is directly upgraded to proposed prohibition of export. The logic conveyed here is very clear: where the same technology is used and for whom it serves determine how strict the regulation it faces will be. Third, the computing foundation. Technologies such as AI processing chip design technology and high-speed interconnect technology between chips or nodes used for large-scale intelligent computing clusters are all proposed to be included in export management. These belong to technologies at the level of AI's key physical infrastructure. After reading these three points, it is clear that this set of rules regulates the underlying logic of technology. When companies use restricted large models to provide API calls, APP/webpage interaction, SaaS applications, local on-device lightweight deployment, and other such services to external users, none of these fall within the scope of large-model technology export management. Simply put, AI companies can continue the commercialization they are supposed to do. After the relevant measures are formally implemented, the vast majority of AI technologies can still be exported freely. Considering the practical needs of international cooperation, the Catalogue also sets up two types of facilitation arrangements. One is open-source technology and specific international business. For technology exports in situations such as open-sourcing AI technologies, Chinese companies building and operating their own computing centers overseas, going abroad to conduct chip tape-out manufacturing, using overseas resources for joint research and development, and participating in international standard-setting, facilitation measures such as automatic licensing apply, basically achieving "report and export immediately"; for relevant technologies that have already entered the public domain through open-sourcing, publication of academic papers, and other means, they are proposed not to be included in the scope of export management. The other is international AI cooperation. For technology exports needed by members of the World AI Cooperation Organization to carry out international cooperation on capacity building, technology sharing, and joint rule-building, more facilitation and support will be provided. Liu Hongmei told Tan Zhu that after the new rules take effect, technologies in the prohibited export category must not be exported; for technologies in the restricted export category, companies can submit applications to the competent authorities and seek to obtain licenses, and this will not disrupt companies' original overseas layouts and business plans. This also confirms the domestic position of "encouraging openness." In the face of the new rules, companies can still remain composed. Why push this set of rules now? Looking globally, the U.S. approach stands in sharp contrast to China's. Over the past few years, the United States has formed two lines of thinking around AI exports. First, differential treatment by country. In January 2025, the U.S. Department of Commerce published the Framework for Artificial Intelligence Diffusion, which was designed to manage advanced computing chips and some advanced model weights based on the "sensitivity" of destination regions and other factors, strictly controlling the flow of advanced AI capabilities to specific entities. Second, dividing circles by camp. The CEO of Anthropic once proposed a "coalition strategy": uniting a group of "democratic countries" into a small circle, controlling the AI technology supply chain, drawing more countries in, and isolating rivals. The logic behind this is simple: whoever masters the most advanced AI has a greater advantage in rule-making. Therefore, while the United States guards the threshold of frontier technology, it also sets limits differently according to country and target, thereby stabilizing its own advantages. But China is not taking that path. This revision is a routine matter carried out in accordance with laws and regulations. Since 2001, China has adjusted and revised the Catalog multiple times. Tan Zhu has learned that the focus of regulation is the act of technology export, not the technology itself. This approach requires policy to break down control standards in sufficient detail, turning them into objective, quantifiable provisions. Why can China do this? Since the standards must be refined and objective, "risk assessment" must be made independently verifiable. This raises a deeper question: who has the capacity to assess risk and carry out governance? On this question, the premise the United States presupposes is that security can only be guaranteed if technological capability is held in the hands of a few. But from the perspective of technological development, leaders do indeed have an advantage when training frontier large models, yet the development of open source is shaking this premise at its root. First, the threshold for training large models is very high, but the threshold for conducting safety testing is not that high. According to industry estimates, training top-tier large models often requires investing tens of thousands of compute cards, but assessing their safety risks requires no heavy reinvestment in training from scratch. This breaks the cognitive limitation that "only technology leaders are qualified to conduct risk review." More research institutions and even government departments are now equipped to conduct safety testing on large models and reproduce safety issues. Second, relying on open source, governance capacity is rapidly becoming widespread. After weights are opened, models are no longer entirely "black boxes." Safety researchers and small and medium-sized technical teams around the world can all polish safety tools at relatively low cost. Since technical-level problems can be solved, rule design can pursue more precise implementation results. The temporary management measures for pre-export registration introduced this time are aimed at grasping relevant technology export demand in advance and making regulation more scientific. Moreover, to give enterprises a clear and unified set of judgment standards, the Catalog also sets explicit quantitative indicators: for example, computing power exceeding 10²⁶ floating-point operations, parameters exceeding 2 trillion, and capability reaching the global top ten level. In this way, enterprises no longer need to guess the direction of regulation, and the government can direct limited administrative resources to the links where risks truly exist. Why build a technology import and export management system? In the past, when facing external technology blockades, we were more on the defensive. But as China moves to the frontier in more and more technology fields, how to manage technology has become a new issue. In this role transition, how to avoid affecting normal commercial cooperation is a problem that had not been fully solved before. Therefore, the focus of this policy adjustment is not "how many technologies are controlled," but rather to build technology import and export management into a finely operable national capability. In the future, it will support our participation in global technological competition at two levels: at the micro level, achieving targeted interception. Undercurrents still exist in international technological competition, and the pressure to prevent core technologies from being illegally transferred and to curb sensitive technologies from flowing out disguised as commercial cooperation has always existed. This management mechanism enables regulators, within the intricate global commercial network, to precisely identify and cut off technology transfers that may endanger national security. At the macro level, forming the capacity to shape rules. An increasingly obvious trend is that China is no longer a passive acceptor of international rules. By establishing transparent, non-discriminatory management standards based on objective technical indicators, China's rules can provide a reference for international technology governance. In this process, we can use multilateral platforms to discuss the boundaries of AI governance with all parties. When this set of rules is gradually recognized or aligned by more multilateral cooperation mechanisms, China will have more institutional-level discourse power in global technological competition. At present, domestic AI technologies and digital services are accelerating their going-global. This technology catalog is only a small step in China's effort to fill the gaps in technology import and export management. But the direction it points to is clear: within the same institutional framework, the security bottom line and open cooperation can absolutely coexist.