NVIDIA Blackwell B200 ships at scale to hyperscalers. 20 PFLOPS of FP4 compute per GPU. Training runs of 100 trillion parameters become feasible. PrevMain BlogNext