> The main reason for the large energy costs of inference is that we are serving hundreds of millions of people with the same model.

its because thats how LLMs work, not because theyre so popular