AI-Native O-RAN Edge Architecture

• AI-Native O-RAN Edge is a converged system combining cloud-native O-RAN and AI orchestration to manage real-time and batch workloads at the network edge.
• It employs unified workload orchestration, multi-timescale control, and partial federated learning to ensure low-latency, high-performance, and secure AI model management.
• The architecture guarantees efficient resource utilization, dynamic admission control, and robust security through open interfaces, enhancing service continuity and resilience.

AI-Native O-RAN Edge denotes the synthesis of advanced AI capabilities and cloud-native Open Radio Access Network (O-RAN) architectures to orchestrate, manage, and deliver both telecommunications and generic AI workloads at the network edge. This paradigm extends O-RAN's core principles—disaggregation, openness, modularity, and cloud-nativeness—beyond network optimization, enabling monetizable distributed AI services and autonomous, context-aware, low-latency edge intelligence (Polese et al., 9 Jul 2025). The defining characteristics of an AI-Native O-RAN Edge are unified workload orchestration, real-time and non-real-time AI-RAN loops, open and standardized interfaces extensible for multi-vendor ecosystems, and secure life-cycle management for massive fleets of ML models.

1. Converged Architecture and Key System Components

The AI-Native O-RAN Edge is anchored on a converged architectural stack that integrates O-RAN Service Management & Orchestration (SMO), an AI-RAN Orchestrator, and distributed AI-RAN Sites:

• AI-RAN Orchestrator: Embedded within the SMO, this module extends conventional service management by integrating AI resource abstraction, unified inventory (vCPU, GPU, FPGA, storage), workload classification (real-time RAN, real-time AI, batch AI, non-RT), admission control, and policy enforcement. It enables both real-time scheduling (10–100 ms loop) and batch orchestration (minutes scale), facilitating flexible deployment models for AI workloads with varying latency, throughput, and geo-targeting requirements (Polese et al., 9 Jul 2025).
• AI-RAN Sites: These are AI-ready O-Cloud edge clusters, leveraging commodity servers with multi-core CPUs, up to four NVIDIA GPUs, and NVMe SSDs. Sites terminate the AI-O2 interface, deploy Kubernetes (or OpenShift) clusters for container orchestration, implement resource isolation (CPU pinning, cgroup QoS, SR-IOV), and support both AI-for-RAN (CU/DU/xApp/rApp acceleration) and AI-on-RAN (generic inference/training) (Polese et al., 9 Jul 2025).
• Open Interfaces: Integration of the AI-O2 interface (protobuf-based extension of O2 for AI-specific policies and telemetry), alongside traditional O-RAN A1, E2, and O1, ensures backward compatibility for RAN functions while supporting full lifecycle and telemetric management of AI workloads.

The architecture supports real-time orchestration, batch job optimization, and geographic model placement to minimize inference latency (<5 ms edge, <20 ms cloud), with admission control that prioritizes ultra-reliable low-latency communication (URLLC) and preemption of low-priority AI jobs during RAN demand peaks (Polese et al., 9 Jul 2025).

2. AI-Oriented Orchestration, Scheduling, and Resource Management

Resource management unifies both telecommunications (RAN) and AI/ML job orchestration via mathematical optimization and feedback-driven control:

• Batch Optimization: Orchestration seeks to maximize system utility $\sum_i U_i(R_i(x,t))$ subject to capacity and deadline constraints, where $x_{i,r}\in\{0,1\}$ indicates job-to-resource assignment and $t_i$ the start time. The utility $U_i$ is a concave function of inference throughput. Constraints enforce resource capacities, $$t_i^{\mathrm{start}} + \tau_i(x_i) \leq T_i^{\mathrm{deadline}}$$, and deadlined, high-priority jobs preempt batch jobs as needed (Polese et al., 9 Jul 2025).
• Multi-Timescale Control: Real-time scheduling loops (≤100 ms) leverage KPI feedback (e.g., buffer occupancy, sudden inference latency spikes) from the E2 interface, while batch loops (minutes) globally optimize analytics and model training using mixed-integer programming or heuristic search (Polese et al., 9 Jul 2025).
• Partial Federated Learning: Distributed learning at the edge is facilitated through partial federated topologies—only the shared parameter layers of multi-task models are exchanged across nodes, while task-specific experts are retained locally. This reduces bandwidth, accelerates convergence, and allows contextual adaptation to local environmental features (Farooq et al., 2024).

A robust policy engine within the orchestrator enforces operator-defined intent hierarchies (e.g., “URLLC > AI inference > batch training”) and supports admission control policies that prevent RAN SLO violations in >99.9% of load peaks (Polese et al., 9 Jul 2025).

3. Model Lifecycle Management, Self-Learning, and Versioning

Continuous training, versioning, and deployment of thousands of heterogeneous ML models across the cell–edge–regional–cloud stack necessitate AI-native lifecycle automation:

• Pipeline: Cloud- or region-side pipelines continuously retrain models on fresh telemetry (RAN KPIs, enriched context), validate candidates, and push artifacts into a distributed version repository indexed by accuracy, stability, security, and resource footprint (Bensalem et al., 24 Jan 2026).
• RL-Driven Update Manager: Update decisions are governed by a reinforcement learning (MDP/Q-learning) policy that trades off inference latency, model accuracy, system stability, and security—issuing rolling update, rollback, or canary deployment commands to the edge via a container orchestrator (e.g., Kubernetes) (Bensalem et al., 24 Jan 2026).
• Resilience and QoS Guarantees: Canary rollouts, telemetry-based anomaly detectors, and namespace-level resource isolation collectively maintain bounded latency (≤10 ms for dApps), zero-SLA violations, and high system resilience even under aggressive model churn (Bensalem et al., 24 Jan 2026).

This framework supports massive horizontal scaling, with the version repository implemented via geo-distributed KV stores (e.g., Cassandra, etcd) and Update Manager policies extended to multi-domain and multi-model deployments (Bensalem et al., 24 Jan 2026).

4. Multi-Task Learning and Distributed Edge Inference

Edge hosting of multi-task deep learning models is a key enabler for scaling AI-nativeness in O-RAN:

• MTL Architectures: Customized gate-control mixture-of-experts (CGC) models, with shared and task-specific sub-networks, dynamically allocate compute for co-located downstream RAN applications (secondary carrier prediction, user positioning, link classification). Uncertainty-based loss weighting automatically tunes loss prioritization according to task noise profiles (Farooq et al., 2024).
• Federated Topologies: Partial federated learning (e.g., FedSim) synchronizes shared backbone weights while retaining local task-expert specialization, balancing global propagation learning with local context adaptation (Farooq et al., 2024).
• Deployment Insights: MTL provides significant gains for certain tasks (e.g., 8% test MAE improvement for positioning, 2–3% for secondary carrier prediction). In data-sparse settings, global aggregation supersedes MTL; task grouping must be optimized to avoid negative transfer, particularly for heterogeneous label/feature distributions (Farooq et al., 2024).

Service lifecycle is dramatically simplified: a single MTL model instance orchestrated from the Near-RT RIC or co-located with the DU replaces multiple per-task models, optimizing edge memory and compute footprints.

5. Security, Interoperability, and Performance Isolation

AI-native O-RAN edge must address vulnerabilities introduced by open interfaces, containerized execution, and distributed AI control:

• Security Framework: Zero-trust architecture with mutual authentication (TLS/mTLS), Trusted Execution Environments (TEEs) for confidential code and model execution, model attestation (signed containers), and continuous vulnerability scanning for all software artifacts in the lifecycle (Abdalla et al., 2021).
• Performance Isolation and QoS: Strict resource binding (CPU pinning, GPU affinity, cgroup-based quotas) and dynamic admission control ensure that AI jobs do not degrade RAN SLAs. Bandwidth for AI workloads is regulated (e.g., >1 Gbps for AI bursts), with adaptive queue management when RAN KPIs are at risk (Polese et al., 9 Jul 2025).
• Open APIs and Certification: All orchestrator modules expose gRPC/protobuf APIs for vendor-neutral automation. Model packaging follows OCI, MLflow, and ONNX standards; deployment is coupled with K8s PodSecurityPolicy templates and AI-O2 conformance tests for certification (Polese et al., 9 Jul 2025).

AI-edge platforms maintain strict separation between RAN real-time xApps/rApps and batch AI workloads, guaranteeing deterministic low-latency while affording flexible infrastructure sharing across vendors.

6. Evaluation, Use Cases, and Measured Gains

Empirical evaluations across prototype and simulation studies demonstrate the architectural efficacy and gains of AI-Native O-RAN Edge models:

• Resource Utilization and Latency: Co-scheduling AI and RAN workloads on GPU clusters increases overall GPU utilization by 30% compared to dedicated, siloed deployments, while preserving RAN latency within sub-100 μs jitter bounds under concurrent execution (Polese et al., 9 Jul 2025).
• AI Efficiency: Batch-aware geographic model placement reduces total edge job completion time by 25% (Polese et al., 9 Jul 2025).
• Resilience/Admission Control: Admission control blocks low-priority AI jobs to maintain RAN SLA in >99.9% of bursts (Polese et al., 9 Jul 2025).
• Service Continuity: RL-driven model update policies allow xApps and rApps to achieve ≥95% of maximum possible accuracy while reducing disruptive model churn by up to 80%, avoiding SLA violations and instability (Bensalem et al., 24 Jan 2026).
• Case Study: With 2 × NVIDIA GPUs simultaneously serving a 1.5 Gbps AI-for-RAN DU workload and a large-language-model chatbot instance, up to 200% aggregate GPU utilization is observed with full service continuity (Polese et al., 9 Jul 2025).

7. Open Challenges and Research Directions

Despite demonstrable progress, several research vectors remain critical:

• Standardization: No unified semantic or information-element definitions for AI workload descriptors, telemetry, or performance metrics are yet available in O-RAN (A1/O2/E2) (Feng et al., 4 Dec 2025).
• Security at Scale: Automated attestation, resilient update pipelines against adversarial code, and secure AI model versioning remain active challenges (Abdalla et al., 2021).
• Latency and Real-Time Constraints: Deploying “zApps” (<1 ms loop) for PHY-level AI remains unaddressed in commodity O-RAN stacks; future work targets sub-ms control for URLLC (Abdalla et al., 2021).
• Explainability and Robustness: Online monitoring for stability under distribution shifts, explainable AI for critical control loops, and proactive canary/rollback mechanisms require further research (Abdalla et al., 2021).
• Multi-Agent and Autonomous Control: Coordinating multiple AI agents (multi-timescale, multi-domain) with conflict resolution and trust guarantees is essential for 6G agentic intelligence (Feng et al., 4 Dec 2025).

Ongoing R&D is advancing digital twin integration, semantic-aware API extensions, continuous learning, and hierarchical orchestration—paving the way for operational, robust, and open AI-native O-RAN edge platforms.