A Chinese lab is reportedly training an AI model of 3 trillion parameters, comparable to frontier models from American labs, despite having significantly less compute power due to export restrictions on advanced NVIDIA chips. This achievement suggests potential breakthroughs in training efficiency, domestic chip advancements, or circumvention of sanctions. US labs have had access to B200/B300s for approximately 1.5 years, which offer 66-92% more FLOPs per watt than the H100, H20, and A800 chips available for export to China.
China's reported 3T model challenges the perceived US lead in AI compute scaling, implying either significant efficiency gains or sanctions evasion.