FLOPs (6 · Nactive · D): 8.4 × 1023
H100 GPU-hours: 583K
Cost (@ $2/GPU-hr): $1.2M

MFU = Model FLOP Utilization: fraction of the GPU’s peak FLOP/s actually realized. 40% is typical for well-tuned dense training; MoE routing and small batches push it lower.