Really the only reason to have a local setup is for 24/7 on-demand high-volume inference that can't tolerate enormous cold starts.