The draft rules, open for public feedback until November 15, distinguish between overseas users interacting with Chinese AI services and foreign entities seeking to independently operate or reproduce the underlying models and data.
Drawing the Line on AI Exports
Under the proposed framework, overseas users would generally be allowed to access Chinese artificial intelligence through cloud platforms and websites. However, the transfer of underlying technology or specific overseas deployments would require government approval.
The draft introduces 23 AI-related technologies to a list of banned or restricted exports. This includes advanced models, autonomous agents capable of performing tasks with minimal human intervention, and specialized technology used for chip design and large-scale computing facilities. A temporary registration requirement for certain technology transfers took effect immediately upon the announcement.
Alfredo Montufar-Helu, a Beijing-based managing director at Ankura, noted that the structure favored by Chinese leadership—keeping research and development onshore while selling operations and access abroad—mirrors where the market was already heading. Analysts suggest the move serves both commercial interests and national security concerns, such as safeguarding sensitive data and cyber capabilities.
Technical Thresholds and Industry Concerns
Leading Chinese AI firms like DeepSeek, Moonshot, and Z.AI currently offer models to international customers via web interfaces or cloud services. Some developers also utilize overseas computing facilities to bypass US export restrictions that block China from acquiring advanced Nvidia processors.
To determine which models fall under the new restrictions, the draft proposes three distinct tests:
- Training Workload: Requiring 10 to the power of 26 calculations.
- Model Size: Measuring at least 2 trillion parameters.
- Performance Ranking: Placing among the top 10 globally in specific efficiency, use, or performance assessments.
Legal experts and tech developers have sought clarification on these criteria. Law firms like Han Kun pointed out uncertainties regarding how regulators will measure models that activate only components per task, as well as questions surrounding the exact rankings that will be used.
Corporate Deals and Foreign Ownership
The regulations extend beyond software distribution to corporate transactions and international partnerships. Registration requirements now cover any transfers that place Chinese AI technology under foreign ownership or control, impacting ongoing deals as well.
These rules build on earlier government interventions, such as Beijing's directive ordering Meta to unwind a multi-billion-dollar acquisition of Chinese-founded startup Manus. While the draft provides simplified approval pathways for Chinese companies building their own computing centers abroad, transferring equivalent technology to foreign customers remains tightly controlled.
Legal analysts view the emerging framework as a direct counterpart to Western restrictions. Yang Jingru, a partner at Jincheng Tongda & Neal, remarked that while the US manages the inflow of hardware, capital, and computing power, China is beginning to systematically oversee the outflow of technological assets and models.
