Every company building an AI assistant is also building hardware that listens and watches 24/7. Local on-device inference is the only way out.