“Inference costs dominate training costs when serving a large language model to billions of users.”