According to semiconductor research firm SemiAnalysis, NVIDIA AI accelerators with a combined rated power consumption exceeding 1.5 gigawatts have flowed into China despite U.S. export restrictions, equivalent to approximately 1.5 million B200 GPUs or 2.14 million H200 GPUs based on power specifications. The analysis was derived from tracking AI accelerator production and flow patterns, with findings reportedly validated by the U.S. Commerce Department.
The findings have drawn attention following recent U.S. criminal cases and policy developments. In March, U.S. prosecutors indicted Supermicro co-founders, alleging approximately $510 million in restricted servers were diverted to China via a Southeast Asian intermediary. Taiwan authorities also investigated illegal exports of 130 servers containing NVIDIA B300 GPUs, with 74 units eventually reaching China. On January 12, the U.S. House passed the Remote Access Security Act to extend export controls to overseas cloud computing services, a proposal NVIDIA has previously opposed.