Edge AI inference faces two core challenges: mobile systems can transport tensors but lack an interface that conveys the inference context those tensors belong to. Prior work has tried to bridge this gap with dynamic partitioning, collaborative inference, and split pipelines, yet these approaches remain tied to specific applications, blind to the network, or context‑free.\ \ This paper introduces Inference on Demand (IoD), an abstraction that exposes split inference through a CAMARA‑style API and carries a split‑aware tensor context in‑band across the 5G user‑plane. IoD is evaluated on a programmable User Plane Function (UPF) testbed using a partitionable visual AI workload. Results show that runtime‑driven adaptive split selection can operate as a network‑facing service, sidestepping the rigid cost profile of static configurations.\ \ When service intent, runtime partitioning, and tensor transport are all visible to the mobile data plane, split inference becomes more than mere offloading—it turns into a network‑native AI service. Programmable telco infrastructure can therefore coordinate computation, context, and connectivity, delivering a unified compute‑and‑network substrate for user‑centric applications.\ \ Review: IoD tightly couples AI inference with the 5G network, offering a practical route to flexible, cost‑effective edge AI services and laying conceptual groundwork for future network‑native AI platforms.