How about spending $50 billion in capex to do $100 billion worth of inference?
Everyone is hung up on the cost of training R1 but it’s 685 billion parameters. We still need all of those GPUs to actually use the model.
Everyone is hung up on the cost of training R1 but it’s 685 billion parameters. We still need all of those GPUs to actually use the model.
That makes both NVDA stock and big AI infrastructure spending less compelling, as those needs are scaled down via software efficiency and chip alternatives.