Challenges in the Telecommunications Industry and AI-RAN’s Return Expansion Mechanism

1. Current Status and Challenges in the Telecommunications Industry

GSMA Intelligence reports that the average ARPU in Canada, the United States, Japan, South Korea, the United Kingdom, France, and Germany, as well as the global average, has remained mostly flat. Adjusted for inflation, this trend indicates a real decline in revenue. Meanwhile, telecom operators continue to increase capital expenditures.

As investment costs rise, telecom operators must find effective strategies to recover infrastructure costs and maximize return.

2. Expansion of the GPU Market Driven by AI Demand

Since the emergence of generative AI, AI has been rapidly transforming society, driving the expansion of the GPU market, which serves as its computational backbone.
The GPU as a Service (GPUaaS) market, which delivers GPU resources over networks, is expected to grow significantly. According to the latest Fortune Business Insights report (published in February 2025), the market size is projected to expand from 4.31 billion USD in 2024 to 49.84 billion USD by 2032, with a compound annual growth rate (CAGR) of 35.8%.

The Global GPUaaS Market Size

3. Aggressive Business Strategies: New Return Generation and Expansion

Transforming into AI-Carrying Infrastructure

SoftBank is advancing a next-generation social infrastructure initiative to distribute computing resources nationwide in preparation for the AI era. (For more details, please refer to the AI-RAN special feature.)

AI-RAN, which integrates RAN control functions and AI server functions on the same hardware platform, is one of the key components of a distributed AI data center. This enables telecom operators to provide both mobile communication services and GPU as a Service (GPUaaS) on a unified hardware platform, creating new revenue opportunities. With AI-RAN, telecom operators can enter the GPUaaS market and seize significant business opportunities.

Flexible Business Strategies

AI-RAN improves infrastructure utilization by optimally balancing two distinct functions: RAN and AI. This enables telecom operators to enhance equipment efficiency by dynamically reallocating server resources. During off- peak hours, server resources originally reserved for ensuring service quality during peak mobile traffic can be allocated to AI workloads, increasing return.

If further growth in demand for GPUaaS is expected, telecom operators can adopt a strategy to expand return by proactively increasing server capacity and allocating more resources to AI workloads.

In other words, they can make investment decisions that align with the timing and scale of growth in the GPUaaS market.

Capital Investment in Response to AI Demand

Initial Target for AI-RAN: Urban Areas

Urban areas, where population and industries are concentrated, have a high demand for AI and GPUaaS, making them a key initial target for AI-RAN.

Similarly, mobile traffic is also concentrated in densely populated urban areas. To improve communication quality in congested areas, advanced antenna technologies utilizing wideband frequencies and coordinated control between adjacent cells are highly effective. The application of AI in RAN enables throughput optimization at cell boundaries and overall capacity enhancement across the entire network.

AI-RAN Architecture Optimized for Urban Areas

To effectively leverage advanced antenna technologies and inter-cell coordination control for improving throughput and capacity in congested areas, it is suitable to adopt a C-RAN (Centralized-RAN) architecture, which enables centralized control of multiple cells on the same server.

A C-RAN architecture involves deploying only Radio Units (RUs) at cell sites, while the Distributed Unit (DU) and Central Unit (CU), which manage and control multiple RUs, are implemented on servers at aggregation sites.

C-RAN Architecture: Servers at Aggregation Sites Centrally Control Multiple Cell Site RUs

4. Return Evaluation

The previous chapter outlined the current state and challenges of the telecommunications industry, demonstrating how AI-RAN enables participation in the GPUaaS market, unlocking new business opportunities. This chapter examines the return structure of AI-RAN, presents a case study, and quantitatively evaluates its return potential.

AI-RAN Return Structure (C-RAN)

The return of AI-RAN consists of revenue from mobile communication services and GPUaaS. Providing these services requires capital expenditures and operational expenses at both cell sites and aggregation sites, as outlined below.

C-RAN Architecture: Servers at Aggregation Sites Centrally Control Multiple Cell Site RUs

GPUaaS Revenue Potential

NVIDIA launched the Aerial RAN Computer-1 as a deployment platform for AI-RAN, featuring the NVIDIA GB200-NVL2 server as its core component. This server dedicated 100% of its computing power to AI workloads and achieves approximately 25,000 tokens per second when running Llama-3-70B FP4, making it capable of handling mainstream large language models. When deployed as GPUaaS, the NVIDIA GB200-NVL2 server is projected to generate 20 USD per hour and 15,000 USD per month in return per server.

Return Analysis in Central Tokyo

We will conduct a case study assuming the deployment of AI-RAN in the Shibuya area of central Tokyo, one of the cities with the highest mobile traffic concentration in Japan.
We will evaluate the return that can be generated by providing GPUaaS while securing server resources for RAN cell accommodation. Furthermore, in this case study, we will prepare two scenarios with different allocation ratios of RAN and AI workloads processed by the servers, and evaluate the Return on Investment (ROI).
Increasing the RAN allocation ratio reduces the number of servers required to cover the area, leading to lower TCO and reduced GPUaaS revenue. Conversely, decreasing the RAN allocation ratio increases the number of required servers, resulting in higher TCO and greater GPUaaS revenue.

＜Case Study RAN Design Conditions＞

・600 cells (100 MHz bandwidth, 4T4R antenna configuration)
・C-RAN architecture, where 600-cell RUs are hosted on GB200-NVL2 servers

*This case study is an example based on a hypothetical scenario and does not represent actual design values ​​or plans.

＜Return Evaluation Assumptions＞

・Revenue: Only GPUaaS revenue is considered (mobile communication service revenue is excluded as it remains constant).
・TCO: Evaluation focuses on TCO at aggregation sites (cell site investments are excluded as they remain constant).
・Operational expenses (OPEX) such as rent and electricity costs are based on pricing in Japan.
・Capital expenditure (CAPEX) amortization period: 5 years
・Evaluation period: 5 years

＜Workload Allocation Scenarios＞

1. 67% RAN, 33% AI (RAN Heavy)
2. 33% RAN, 67% AI (AI Heavy)

Scope of Return Estimation (C-RAN)

In the RAN Heavy scenario (67% RAN, 33% AI), the five-year ROI reaches a maximum of 33%. The revenue-to-CAPEX ratio over five years, excluding OPEX based on Japan’s market conditions, is approximately 2x.

In contrast, the AI Heavy scenario (33% RAN, 67% AI) prioritizes aggressive revenue expansion by allocating more resources to AI workloads. This results in an increase in the required number of servers, leading to a higher TCO compared to the RAN Heavy scenario. However, the additional GPUaaS revenue surpasses the increase in TCO, driving the five-year ROI up to 219%. The revenue-to-CAPEX ratio, excluding Japan-based OPEX, is approximately 5x.

The results of these two scenarios do not indicate that one is the correct choice over the other. If the goal is to enter the GPUaaS market while maintaining a conservative AI workload allocation, the RAN Heavy approach provides a balanced path to new revenue generation. However, if there is a clear outlook for significant market growth, increasing CAPEX and AI workload allocation enables a more aggressive return expansion strategy.

The TCO comparison of AI-RAN against x86-based vRAN (virtualized RAN) and ASIC-based custom BBUs (Baseband Units) under the same conditions is shown below. Unlike x86-based vRAN and ASIC-based BBUs, which cannot generate GPUaaS revenue, AI-RAN demonstrates a clear advantage in maximizing return potential.

AI-RAN Return Potential for Covering 600 Cells in Urban Tokyo

・ROI = (5-year GPUaaS revenue − 5-year TCO at aggregation site) ÷ 5-year TCO at aggregation site
・Revenue-to-CAPEX Ratio=5-year GPUaaS revenue ÷ CAPEX at aggregation site

5. Towards Social Implementation of AI Infrastructure

AI-RAN addresses the challenges faced by the telecommunications industry while leveraging the growth of the AI market to create new business opportunities. At the same time, it provides high reliability, low latency, and robust security, offering unique added value that contributes significantly to future industrial development and economic growth.

AI-RAN technology is constantly evolving. Through continuous technological development and validation, its value and competitive advantage continue to be demonstrated. However, implementing technological innovations in society requires more than just technological progress; shaping them into sustainable business is essential. Research is being conducted on both technological innovation and business strategies for social implementation, aiming to provide society with a new infrastructure that carries AI. Stay tuned for further updates.