True, but local setups can run LLM requests in parallel too. In this case efficiency gap is much narrower.