The LWN comments say that you are correct for local AIs (but not LLM services), modulo some caveats about compiler flags and hardware used.