Comment by ImaCake

In my experience, LLM usage follows an exponential distribution i.e. most people using LLMs are not using many resources, but a very small few are using a massive amount of resources. Most LLM usage is trivial and tiny, most of what I use LLMs for could be done by a local model with a decent web retrieval tool. What some people I know use LLMs for requires massive volumes of tokens. Its that small power user base which I would bet they are targetting.