Regulatory bans, export controls, and reliability failures are forcing AI developers to move inference to the edge.