AI-RAN Factory: AI-Driven Radio Networks

• AI-RAN Factory is an integrated platform that combines AI service provisioning with RAN control using semantic intent abstraction and digital twin validation.
• It employs a modular, layered architecture with LLM-driven agents and closed-loop optimization to manage both real-time and long-term network operations.
• Demonstrations show up to 19.6% energy savings and enhanced throughput while maintaining strict SLA compliance across heterogeneous networks.

An AI-RAN Factory refers to an agentic, automated platform integrating AI service provisioning and Radio Access Network (RAN) control within a unified, programmable infrastructure. This concept operationalizes the convergence of “AI-for-RAN,” “AI-on-RAN,” and “AI-and-RAN” paradigms, supporting both energy-efficient wireless communication and the orchestration of heterogeneous AI workloads for 6G network deployments. Architecturally, the AI-RAN Factory leverages semantic intent interpretation, LLM-driven orchestration, and closed-loop optimization to realize sustainable, adaptive, and high-performance AI-native RANs (Aroua et al., 20 Jun 2026, Li et al., 8 May 2026).

1. Architecture and Layered Composition

The AI-RAN Factory is structured as a modular, multi-agent stack that abstracts human operator/service manager intents into executable control and deployment actions. The architecture comprises distinct yet tightly integrated layers (Aroua et al., 20 Jun 2026):

• Operator/Service Manager: Accepts high-level semantic intents (e.g., “minimize RAN energy by ≥ 15% while keeping throughput ≥ 95%”) and SLA definitions.
• Semantic Intent Abstraction Layer: Parses natural language intents to structured intent objects, detects conflicts, and resolves priorities (notably between AI-for-RAN and AI-on-RAN objectives).
• SMO (Service Management & Orchestration): Hosts the LLM-based Semantic Coordination (SC) Agent, connected to O1 for enrichment DB, A1 for RIC, and a Telemetry Bus for KPI streams.
• SC Agent: Selects candidate rApps/xApps and their meta-parameters, routes configurations for validation.
• Digital Twin (DT) Agent: Simulates candidate configurations over historical/forecasted workload traces, validating constraints (throughput, latency, energy).
• CDM Agent: Handles deployment and monitoring, including feedback for drift or SLA violations.
• Non-RT RIC/ Near-RT RIC: Deploy rApps (long-term, ≥1 s control) and xApps (sub-second, real-time control) respectively.
• Data Plane + Edge AI Compute Farm: Executes both radio and AI workloads on shared hardware.

This layered separation enables semantic-to-physical intent closures, modular service upgrades, and tight orchestration between communication and AI services (Aroua et al., 20 Jun 2026, Li et al., 8 May 2026, Chatzistefanidis et al., 8 Aug 2025).

2. System Model, Joint Optimization, and Scheduling

AI-RAN Factory operation is governed by a multi-objective constrained optimization. The model considers carrier activation ($x \in \{0,1\}^N$), PRB allocation ($f_i$), and AI compute allocation ($a_k$). Objective:

$$\min_{x,f,a} \; \lambda_1 \sum_{i=1}^N x_i\,P_i(f_i) + \lambda_2 \sum_{k=1}^K Q_k(a_k)$$

subject to:

$$\begin{aligned} & \sum_{i=1}^N x_i\,C_i(f_i)\ge T_{\min} \ & L_0 + \Delta L(x,f) \le L_{\mathrm{comm}}^{\max} \ & g(a) \le L_{\mathrm{ai}}^{\max} \ & \sum_{k=1}^K a_k \le 1 \;;\; x_i\in\{0,1\},\; f_i\ge0 \end{aligned}$$

$\lambda_1,\lambda_2$ are operator-tunable weights, e.g., for Green-SLA (energy efficient) or AI-priority profiles. The optimization targets minimum aggregate radio and compute energy, subject to throughput and latency SLAs (Aroua et al., 20 Jun 2026).

Resource allocation is further extended by agentic and hierarchical controllers: slow-timescale placement is handled via LLM-based agent proposals, followed by event-driven convex GPU/CPU allocation minimizing deadline-weighted surrogate cost. This duality enables robust real-time enforcement of RAN hard-deadlines alongside opportunistic AI workload fulfillment (Li et al., 8 May 2026).

3. LLM-driven Coordination, Semantic Parsing, and Safe Enactment

Central to the AI-RAN Factory is the application of LLM agents for semantic abstraction and orchestration:

procedure HANDLE_SEMANTIC_INTENT(intent_text): intent_obj ← SemanticLayer.parse(intent_text) intent_set ← IntentResolver.prioritize(intent_obj) rapp_list ← SC_Agent.select_rApps(intent_set) for each r in rapp_list: params[r] ← SC_Agent.configure_params(r, intent_set) for each (r, p) in (rapp_list, params): sim_KPIs ← DT_Agent.simulate(r, p) if DT_Agent.check_constraints(sim_KPIs, intent_set): valid_list.add((r,p)) for each (r,p) in valid_list: CDM_Agent.deploy_rApp(r, p) while true: live_KPIs ← Monitor.fetch_metrics() if SC_Agent.detect_violation(live_KPIs, intent_set): HANDLE_SEMANTIC_INTENT(intent_text) break end procedure

Conflict resolution between goals (e.g., “critical AI service” vs. “energy saving”) occurs at this semantic intent processing layer. Deployment is mediated by Digital Twin emulation and validated by the CDM agent before activation in RICs (Aroua et al., 20 Jun 2026). This approach ensures that only candidates meeting all constraints (including safety/coverage/latency) are deployed, preserving system integrity (Li et al., 8 May 2026).

4. Quantitative Performance and Use Case Results

Operational effectiveness of the AI-RAN Factory is validated over multi-week sector-level studies (Aroua et al., 20 Jun 2026):

Use Case	Profile	Energy Gain	Coverage %	Throughput/UX %
AI-for-RAN	Steering	19.6 %	100 %	95.8 %
AI-on-RAN	UX-95	15 %	99 %	96.5 %
AI-on-RAN	UX-90	18 %	99 %	98.8 %

Notably, prediction-based sector steering achieved 19.6 % energy reduction while satisfying all coverage constraints and achieving 95.8 % throughput satisfaction. UX-aware DRL-based AI-on-RAN use cases delivered 15–18 % energy savings with ≥99 % coverage and high session satisfaction, underscoring the benefit of coordinated, agentic orchestration. The energy gain is quantified as:

$$\text{Energy Gain} = \frac{E_{\mathrm{base}} - E_{\mathrm{opt}}}{E_{\mathrm{base}}} \times 100\%$$

5. Design Principles: Modularity, Trust, and Scalability

Key architectural and operational guidelines are distilled as follows (Aroua et al., 20 Jun 2026):

• Modular Stack: Stateless LLM SC agents, with all actual actions validated by digital twin or monitoring layer before deployment.
• Open Interfaces: Use of O1 (enrichment DB), A1 (Non-RT RIC), E2 (Near-RT RIC), and telemetry streams for operational data collection and control flows.
• Scalability: Partition SC agents per sector cluster, employ hierarchical LLM or SLM for low-latency local coordination, and cache parameter templates to accelerate validation.
• Digital Twin Sizing: Use light simulators for bulk pruning, high-fidelity twins for final validation.
• Trust and Safety: Strict adherence to hard constraints (coverage, latency ceilings), operator override, audit logging of LLM rationale.
• Continuous Learning: Live metric feedback is used for model retraining and policy review, adapting weight parameters ($\lambda_1$, $\lambda_2$) and LLM prompts.

6. Comparative Results and Competitive Advantage

Extensive benchmarking against other multi-agent orchestration frameworks demonstrates superior SLA fulfillment and energy efficiency (Li et al., 8 May 2026). The hierarchical agentic approach (layered LLM agent with predictive critic and fast allocation) reached 90.0 % overall SLO fulfillment (+20.5 % over the previous strongest baseline) with AI-service request fulfillment increasing from 51 % to 85.3 %, maintaining RAN hard deadline fulfillment and efficient migration handling via critic gating. Empirical results confirm the scalability, adaptability, and robustness of the Factory blueprint across dynamic and heterogeneous network conditions.

7. Outlook and Deployment Trajectory

The AI-RAN Factory paradigm jointly optimizes communication and AI workloads by integrating semantic LLM mediation, deterministic optimization (digital twin, rApps), and validated closed-loop actuation (CDM). The approach enables holistic, energy-aware orchestration for 6G-ready RAN deployments, supporting modular upgrades, multi-vendor compatibility, and continuous self-adaptation (Aroua et al., 20 Jun 2026, Li et al., 8 May 2026). Operational recommendations emphasize stateless intent translation, incremental validation, and policy feedback to assure both performance and safety at scale. As demonstrated, this enables operators to realize sustainable, AI-native radio networks capable of efficient multi-objective optimization and real-time adaptability.