“Microsoft's most recent text model, which debuted at rank 13 on the LM Arena, was trained on a cluster of 15,000 NVIDIA H100 GPUs.”