"The 2.8x H20 performance figure is specifically for FP4 inference workloads. For training large models, which is what defines Frontier AI development, the picture is much less favorable for Huawei. Nvidia H100 and B100 still dominate training workloads by significant margins."