[conradev]

They’re already optimizing GPU die area for LLM inference over other pursuits: the FP64 units in the latest Blackwell GPUs were greatly reduced and FP4 was added