"An H100 can serve more tokens per GPU of GPT-5.4 than if you had run GPT-4 on it. So it is producing more tokens of a model that is of higher quality."