The "humans as tools" model only works if the AI layer is centralized and owned by platforms. If inference runs on hardware you control, you're not callable - you're the one calling.
Been thinking about this a lot: https://www.localghost.ai/reckoning