Toward AI-on-RAN: Enabling on-RAN Edge Intelligence for AI-Native 6G Networks

The growing complexity and latency sensitivity of connected devices are driving the demand for real-time services to be delivered closer to end users. While current edge computing approaches (e.g., MEC, dUPFs) bring computation closer to users, they remain external to the RAN protocol stack, requiring explicit user-traffic routing, session anchoring, or client-side adaptation mechanisms. This breaks service transparency and ties applications to fixed deployment sites, which is unsuitable for highly mobile devices. To address this limitation, we propose the architectural vision and framework for the emerging AI-on-RAN paradigm, which integrates and serves AI workloads directly from the RAN infrastructure. By leveraging GPU-accelerated RAN platforms, our modular framework with new user-plane aware RAN Apps enables on-premises AI inference fully transparently to end users at the radio edge. We validate our design through a proof-of-concept implementation and OTA experiments with 5G RAN stack. Through two interactive AI case studies, we demonstrate significant latency reductions and efficient use of spare resources without degrading core radio functions. This work presents an early blueprint and prototype for AI-on-RAN, highlighting its technical feasibility and motivating further exploration toward AI-native 6G networks.