Abstract
The radio access network (RAN) landscape is undergoing a transformative shiftfrom traditional, communication-centric infrastructures towards convergedcompute-communication platforms. This article introduces AI-RAN whichintegrates both RAN and artificial intelligence (AI) workloads on the sameinfrastructure. By doing so, AI-RAN not only meets the performance demands offuture networks but also improves asset utilization. We begin by examining howRANs have evolved beyond mobile broadband towards AI-RAN and articulatingmanifestations of AI-RAN into three forms: AI-for-RAN, AI-on-RAN, andAI-and-RAN. Next, we identify the key requirements and enablers for theconvergence of communication and computing in AI-RAN. We then provide areference architecture for advancing AI-RAN from concept to practice. Toillustrate the practical potential of AI-RAN, we present a proof-of-conceptthat concurrently processes RAN and AI workloads utilizing NVIDIA Grace-HopperGH200 servers. Finally, we conclude the article by outlining future workdirections to guide further developments of AI-RAN.