Not necessarily. There are lots of use cases for on device AI inference. I run YOLO on an Nvidia Jetson powered Lenovo Think Edge, which processes incoming video at full frame rates on four channels with recognition and classification for a bespoke premises security system. No clouds involved other than the Nix package manager etc. Make sure your argument May carry more weight when you're talking about ultra low power devices like an Arduino running AI inference locally that seems like more of a stretch.