A leading Chinese AI executive estimates training a frontier model requires approximately 50,000 NVIDIA GB300 systems or 200,000 Huawei 950 accelerators, highlighting the capital intensity of frontier AI. He notes that China's largest domestic models are in the tens of billions of active parameters, while frontier models operate around 800 billion active parameters.
Access to compute infrastructure, not human capital, is the binding constraint for China's frontier AI development, making export controls highly effective.