China’s AI Power Profile: Advantages, Dependencies, and Limits

1) Computing Power: Advanced Chip Dependency and Choke Point Sanctions

Training and inference for AI systems rely heavily on graphics processing units (GPUs) and specialized AI accelerator chips. Chinese AI companies and research institutions, particularly LLM developers such as Baidu (Wenxin Yiyan), Alibaba (Tongyi Qianwen), ByteDance (Doubao), Reports indicate that training a frontier-level AI model typically requires tens of thousands of NVIDIA A100 or H100 chips. The U.S. export ban on high-performance GPUs to China (covering A100, H100, and their Chinese variants A800/H800) has been fully implemented, forcing reliance on stockpiled chips or lower-performing domestic alternatives such as Cambricon, Tiansuan, and Hygon. These domestic chips lag 1-3 generations behind NVIDIA in manufacturing process, power consumption, software compatibility, and ecosystem support.

China’s advanced chip manufacturing is similarly constrained. SMIC, its only high-end process manufacturer, is hindered by export restrictions on essential equipment, making mass-production of AI chips at 7nm or smaller nodes difficult. As a result, from design to deployment, China’s AI computing base remains externally dependent—a core variable limiting competitiveness.