“GPUs typically achieve a Model Flops Utilization (MFU) of 20% to 50% on real workloads.”