it eats into all NVDA consumer-facing clients no? I can see why openai and etc are looking for alternative hardware solution to train their next model.