DeepSeek's Liang Wenfeng stated that China's largest bottleneck is compute resources, not talent, making it impossible to train models as large as US counterparts even with 50,000 GB300s or 200,000 Huawei 950s. DeepSeek currently operates with roughly 20,000 H-equivalent GPUs, with most having arrived recently, while Huawei's total capacity is limited to around 16,000 cards for DeepSeek and a hundred-something thousand for internet giants.
US export controls are effectively capping China's frontier AI model development by limiting access to high-end compute, forcing Chinese labs to pursue efficiency over scale.