upvote
GPUs are much more efficient at parallelizing requests for LLMs so it's going to much more efficient to centrally host. Maybe big companies it would make sense to get their own though.
reply