AI-RAN Sites: AI-Enhanced RAN Clusters

Updated 3 July 2026

AI-RAN sites are clusters that fuse disaggregated RAN infrastructure with edge AI resources to enable real-time signal processing, AI inference, and service orchestration.
They employ a hierarchical architecture combining physical RAN components, edge AI modules, and management layers for low-latency control and robust multi-vendor interoperability.
Advanced AI techniques, including RL-based agents and hierarchical planning, are used to optimize resource allocation, maintain strict SLOs, and enhance energy efficiency in 6G deployments.

An AI-RAN site is a geographically co-located or distributed cluster that fuses disaggregated radio access network infrastructure—such as base stations, RAN Intelligent Controllers (RICs), and transport links—with edge AI compute resources and service orchestration layers. The resulting network entity is capable of both real-time RAN signal processing and on-site AI inference/training, supporting native, low-latency intelligence to accelerate network services, automate operations, and expose programmable APIs for emerging verticals in 6G and beyond (Srinivasan et al., 14 Feb 2026, Shah et al., 12 Jul 2025, Polese et al., 9 Jul 2025, Ananthanarayanan et al., 2024).

1. Site Definition and Hierarchical Architecture

An AI-RAN site, as defined for 6G, comprises three hierarchical strata: (a) physical/RAN elements—including gNodeBs, O-DU/O-CU, fronthaul/backhaul; (b) edge-AI modules, such as GPU/ASIC-accelerated servers hosting AI workloads and service pipelines; and (c) management layers involving event-driven RICs (Near-RT and Non-RT) with containerized xApps/rApps for optimization and automation. Human- and agentic-assistant interfaces span these layers for real-time co-management and troubleshooting (Srinivasan et al., 14 Feb 2026, Shah et al., 12 Jul 2025, Polese et al., 9 Jul 2025).

Typical architectural breakup:

Layer	Core Components	Functionality
Physical/RAN	RU, DU, CU, fronthaul/backhaul	PHY/MAC, scheduling, RRM, protocol stack
Edge AI	GPU/NPU servers, containers	AI inference/training, native AI services
Management/Orchestration	RIC (xApp/rApp), dashboards	Control, policy, monitoring, policy framework

Modern sites routinely deploy these within a single container orchestrator (K8s/OpenShift), integrate multi-domain telemetry, and adhere to open interface standards (O1/O2/E2/A1/Y1, and extensions like AI-O2 and O2+) to enable robust multi-vendor interoperability (Polese et al., 9 Jul 2025, Maxenti et al., 2023). The AI-RAN paradigm thus extends the O-RAN "O-Cloud" concept to host a fully orchestrable, mini edge data center (Polese et al., 9 Jul 2025).

2. Resource Orchestration and Control Loops

Resource orchestration within AI-RAN sites is fundamentally cross-domain, involving both classical RAN scheduling and dynamic allocation of AI compute. Typical frameworks include:

xApp/rApp pipelines: Real-time and non-real-time control logic hosted in the Near-RT and Non-RT RICs, interacting with O-DU/O-CU and the AI infrastructure through open interfaces (Shah et al., 12 Jul 2025).
End-to-end orchestrator: A higher-order entity (e.g., CAORA or HAF orchestrator) takes global site state—RAN KPIs, AI workload forecasts, infrastructure utilizations—and solves joint resource provisioning, placement, and scheduling at different timescales (Shah et al., 12 Jul 2025, Li et al., 8 May 2026).

For compute sharing, leading schemes leverage RL-based agents (e.g., SAC) (Shah et al., 12 Jul 2025, Shah et al., 10 Mar 2025) or hierarchical LLM-critic agentic planners (Li et al., 8 May 2026). These systems address compute partitioning (e.g., NVIDIA MIGs for RAN/AI isolation), service migration with interruption cost modeling (reload time, VRAM/DRAM fitting), and enforce SLOs across highly heterogeneous services.

Mathematical models typically maximize weighted utility functions

$U = \sum_{n} w_{\text{RAN}} f_{\text{RAN}}(n) + w_{\text{AI}} f_{\text{AI}}(n)$

subject to CPU, GPU, bandwidth, and strict latency constraints per domain (Polese et al., 9 Jul 2025, Maxenti et al., 2023).

3. Functional Services and Capabilities

AI-RAN sites enable a spectrum of natively converged services:

Real-time signal processing: Sub-ms beamforming, link adaptation, RRM, enabled by tightly-coupled RAN-AI co-execution (Polese et al., 9 Jul 2025, Li et al., 11 Jul 2025).
Edge AI services: Hosting of third-party DApps, on-site large model inference (LLM, CV, analytics), federated and split learning, all orchestrated to avoid impacting RAN control latency (Ananthanarayanan et al., 2024, Ding et al., 17 Jul 2025).
Telemetry aggregation and caching: Local preprocessing of high-frequency KPIs, model checkpoints, buffer states, and feature maps, for efficient backhaul utilization.
Mission-critical operations: For LAE/UAV applications, semantic-aware rApps and RL-powered xApps enable closed-loop, <50 ms control of fleets based on synergy between radio metrics and AI-processed scene semantics (Abdalla et al., 1 Jan 2026).

Dynamic sharing is enforced either by agentic RL (SAC) (Shah et al., 10 Mar 2025, Shah et al., 12 Jul 2025), optimization-based scaling (Maxenti et al., 2023), or deadline-driven hierarchical agentic frameworks (Li et al., 8 May 2026). Strict resource isolation and scheduling priorities ensure RAN SLA preservation even with bursty AI loads (Shah et al., 10 Mar 2025, Polese et al., 9 Jul 2025, Li et al., 8 May 2026).

4. Performance Metrics, Evaluation, and Guarantees

Evaluation of AI-RAN site orchestration is based on domain- and function-specific KPIs including:

Planning/Operation/Tuning Accuracy: Defined as the fraction of correct plans, tool uses, or tuning steps executed by a co-management AI assistant (Srinivasan et al., 14 Feb 2026).
SLO Fulfillment: Fraction of requests (RAN and/or AI) that meet preannounced delay or throughput deadlines (Li et al., 8 May 2026).
Resource Utilization: Typically GPU/CPU occupancy, measured as average fraction of time compute is not idle (Shah et al., 12 Jul 2025, Shah et al., 10 Mar 2025).
AI-RAN coexistence: RAN SLA compliance (e.g., 99%) even as AI workloads make maximal use of shared hardware (Shah et al., 10 Mar 2025, Shah et al., 12 Jul 2025, Li et al., 8 May 2026).
Energy and Scalability: Energy minimization, reduction in total draw (e.g., 15–34% energy savings via AI-driven cell sleep/on-off) (Catalan-Cid et al., 4 Jun 2026, Li et al., 11 Jul 2025).

Empirical validations employ trace-driven simulations (e.g., Barcelona 5G, Azure LLM traces), testbed deployments (e.g., EURECOM OAI/OpenShift), and field trials (e.g., 5000-site China 5G-A cluster) (Shah et al., 12 Jul 2025, Abdalla et al., 1 Jan 2026, Bouknana et al., 24 Nov 2025, Li et al., 11 Jul 2025).

5. Agentic and Conversational Co-Management Frameworks

Advanced sites increasingly incorporate agentic assistants for human-in-the-loop and automated co-management:

Turn-based conversational UI: Decomposes user queries into hierarchical intents, invokes tool/knowledge sources, and aggregates answers with ground-truth validation (Srinivasan et al., 14 Feb 2026).
Three-layer agentic architecture: Interface (dashboards, chat), intelligence (pipeline orchestration, response generation), knowledge (pattern-matched data routers) (Srinivasan et al., 14 Feb 2026).
Mitigation of hallucination: RAG grounding, deterministic numeric checks, multi-path validation, user feedback loops, and cached gold-standard templates (Srinivasan et al., 14 Feb 2026).

Measured accuracies: design/planning (78%), operation (89%), tuning (67%), average dialog response time (13 s). Hallucination rates remain non-negligible (43%), driving research in robust retrieval-augmented designs and alignment with real-time telemetry (Srinivasan et al., 14 Feb 2026).

6. Standardization, Open Interfaces, and Multi-domain Operation

AI-RAN sites rely on rigorous standardization for interoperability and scalability:

O-RAN interfaces: O1 (telemetry), O2 (infra management), AI-O2 (AI resource exposure), A1 (policy), E2 (near-RT RIC for xApps), Y1 (radio analytics for orchestrators) (Polese et al., 9 Jul 2025, Shah et al., 12 Jul 2025, Shah et al., 10 Mar 2025).
Multi-vendor orchestration: Open hardware and software stacks (e.g., OAI, FlexRIC, RedHat, Rancher) deployed as containers and microservices with Northbound APIs for developer integration (Bouknana et al., 24 Nov 2025, Polese et al., 9 Jul 2025).
Deployment models: Sites can be macro/small-cell co-located, standalone edge datacenters, or mobile (e.g., research UAVs), with orchestration from local to regional to central hierarchies (Polese et al., 9 Jul 2025, Ananthanarayanan et al., 2024).

APIs expose AI-as-a-Service (AIaaS) primitives, deterministic latency/gflops SLAs, and flexible life-cycle management across domains, supporting operator monetization schemes as well as hybrid energy-optimization strategies (Li et al., 11 Jul 2025, Maxenti et al., 2023).

7. Research Challenges and Outlook

Open research problems at AI-RAN sites include:

SLO violation mitigation: Deadline-aware resource allocation, handling service interruption during migration, and supporting mixed workloads with hard/soft latency constraints (Li et al., 8 May 2026).
Model lifecycle and federated operation: Continuous retraining, traceable datasets, synchronization across handovers, robust privacy mechanisms for distributed learning (Ananthanarayanan et al., 2024, Ding et al., 17 Jul 2025, Li et al., 11 Jul 2025).
Standardization: Extension of E2 service models for semantic features, introduction of SLA descriptors for AI/LAE, and universal KPI reporting (Abdalla et al., 1 Jan 2026, Polese et al., 9 Jul 2025, Li et al., 11 Jul 2025).

Empirical and trial evidence indicates that AI-RAN sites, when implemented with principled multi-layer orchestration and open interfaces, achieve marked gains in latency, energy, capacity, and operator agility, forming the foundation for intelligent, distributed 6G (Li et al., 11 Jul 2025, Polese et al., 9 Jul 2025, Abdalla et al., 1 Jan 2026, Srinivasan et al., 14 Feb 2026).