AI-RAN Technical and Economic Analysis: Cost Modeling and Revenue Opportunities of Open Programmable Architecture

This study focuses on the technical and economic analysis of AI-RAN. Addressing the 5G ROI dilemma and economic concerns about 6G deployment, it constructs a joint cost-benefit model for GPU-based RAN. Key conclusion: AI-on-RAN revenue can offset the additional expenditure from heavy GPU asset investment, achieving a maximum 8x ROI across multiple scenarios, which provides critical support for the economic feasibility of 6G deployment.

5G ROI Dilemma

Large-scale 5G deployment has not met expected ROI, raising concerns about the economic feasibility of 6G rollout. Operators need to explore new business models to revitalize existing assets.

AI Computing Power Supply-Demand Imbalance

The surge in AI inference/training demand has led to insufficient data center computing power and rising costs, presenting opportunities to convert idle computing capacity in communication infrastructure into AI services.

Technical Foundation for Resource Sharing

The AI-RAN architecture shares idle GPU-accelerated RAN capacity (during off-peak hours) with AI workloads, enabling flexible resource allocation based on the Open RAN universal GPU/software-defined platform (traditional ASICs have fixed functions and low utilization).

Lack of Economic Case

Previous AI-RAN economic cases lacked systematic demonstration. This study uses a quantitative model to answer: Can additional GPU investment be recovered? What is the ROI in different scenarios? What are the key influencing factors?

Benchmark Data for Heterogeneous Platforms

Collect processing benchmark data from x86 servers (with accelerators) to GPU platforms, covering computing efficiency differences across scenarios, providing a technical basis for cost modeling.

Traffic and Demand Profiles

RAN utilization fluctuates diurnally (dropping to below 20% during off-peak periods), while AI demand has characteristics like stable inference and bursty training. It is necessary to balance RAN service quality and AI throughput.

Joint Cost Model

CAPEX: GPU hardware procurement, infrastructure upgrades, software transformation; OPEX: increased energy consumption, maintenance costs, network interconnection fees. Model assumption: Prioritize RAN service quality, with AI only using idle resources.

AI Revenue Evaluation

Based on LLM inference market pricing, consider token depreciation, dynamic demand, and GPU service density to evaluate the revenue potential of AI-on-RAN, and analyze the impact of demand fluctuations on ROI.

Multi-Scenario ROI Results

AI-on-RAN revenue can offset additional GPU expenditure, achieving a maximum 8x ROI across multiple scenarios, far exceeding the return level of traditional telecom infrastructure.

Key Success Factors

1. GPU utilization: Higher utilization leads to lower unit cost; 2. Demand complementarity: Optimal when RAN off-peak periods overlap with AI peak demand; 3. Service pricing: Maintain premium through differentiated services (low latency, edge deployment).

Implications for 6G Deployment

• Enhanced economic feasibility: AI revenue subsidizes infrastructure investment, changing the 6G investment logic (communication-computing convergence);
• Importance of open programmability: Supports flexible resource scheduling, aligning with the Open RAN concept;
• Ecosystem building: Need to collaborate with AI service providers and establish standardized transaction mechanisms.

Risk Challenges

• Technical risks: Resource isolation, real-time scheduling, and fault recovery need verification;
• Regulatory uncertainty: Compliance costs related to data privacy and model security;
• Competitive landscape: Cloud service providers are deploying edge AI; operators need to develop differentiated advantages.