An insight into the future of AI-RAN

Download

This article is based on a discussion with Dr. Alex Jinsung Choi, Chair of the AI-RAN Alliance on Tuesday 18th March 2025. The interview explores the overarching ambition of the AI-RAN Alliance, the key drivers motivating this evolution of RAN infrastructure, the use cases this will enable and ultimately how these advances help to address the wider issues facing the industry. 

What is the AI-RAN Alliance?

Announced originally at MWC 2024, the AI-RAN Alliance is a collaborative initiative aimed at advancing the integration of AI into the radio access network (RAN) in order to transform its capabilities. Since its announcement, membership has grown significantly – from 11 founding members to 89 (as of April 10, 2025) – with an additional 65+ new applications pending. Its membership encapsulates academic institutions, network vendors and telecom operators – of which there are now nine: T-Mobile US, SK Telecom, KT, LG Uplus, SoftBank, Boost Mobile, Indosat Ooredoo Hutchison, Turkcell and Globe Telecom.

Source: STL Partners

The AI-RAN Alliance is comprised of three working groups:

• AI-for-RAN: The integration of AI in the RAN to enhance performance and efficiency, particularly in relation to KPIs such as energy consumption and processing capacity.
• AI-and-RAN: The development of a common converged infrastructure platform at the RAN that can orchestrate heterogenous workloads (namely network and enterprise AI workloads). This is the target architecture the alliance is working towards.
• AI-on-RAN: Building on the tools developed by the two aforementioned working groups, this relates to defining the commercial mechanisms that can use these tools to achieve greater monetisation of the RAN, as it pertains to new enterprise services.

The three pillars of the AI-RAN Alliance

Source: STL Partners

What are the trigger points motivating this evolution of the RAN?

It is no secret that the industry is facing a range of existential issues – market saturation and revenue decline to name two – however, in spite of this turmoil there are strong reasons to be hopeful. While on the business side, fragmentation is causing harm to operators, on the technology side there isn’t this same fragmentation. Operators have the largest common infrastructure platform in the world. The AI-RAN Alliance therefore is looking to better harness the potential of this infrastructure footprint, and break the status quo. If the industry is to reverse its decline, it needs to embrace a more bold and ambitious approach – and look to new approaches, new partnerships, new services. The AI-RAN Alliance is about contributing a technological and business platform in pursuit of this.

To the question of what will drive operators to deploy GPUs specifically at the RAN, there are again industry-wide trends shaping this outcome.

• Particularly, in relation to the sphere of generative AI, there is a consensus that GPU-equipped infrastructure is a necessary prerequisite to support AI applications. While advancements are being made to run models strictly on a CPU, this is by no means eliminating the needed for accelerated hardware. As such, to leverage the RAN as a platform to support demanding enterprise AI workloads it must be equipped with GPU – or in other words “G-RAN” must be deployed (you heard it here first).
• In the RAN domain, telco capex decisions are based on traffic forecasts, and initial predictions point to “AI traffic” accounting for a substantial share of total mobile traffic by the end of the decade. Therefore, if “AI traffic” is to become a dominant form of mobile traffic – then operators must develop a platform that can support it sufficiently. In this case it can be best served by a distributed footprint of GPU infrastructure.
• In addition, there are structural forces shaping this need for GPU infrastructure. GPU-equipped hardware will replace CPU-only hardware across many product lines in the coming years. This will drive a wider shift in procurement decisions across the industry, prompting a shift to deploy more GPU resources, in order to capture a better cost of performance.

This migration to “G-RAN” will likely be an incremental investment, however, unlike the step change investments required to rollout a new generation of mobile infrastructure. This evolution will be more targeted.

What type of enterprise workloads do you envision AI-RAN supporting?

Although, the alliance is still in its early days, there are broadly three buckets of enterprise workload that is envisioned to be deployed on AI-RAN infrastructure.

Source: STL Partners

1. Edge AI inferencing: AI-RAN infrastructure will be capable of directly hosting enterprise edge AI workloads, i.e. a computer vision application performing flow analysis for a retail store. This will mostly encapsulate the edge IoT branch of use cases. The AI-RAN Alliance has previously demonstrated how the RAN can be used to support computer vision applications to enable smart city use cases.

2. Network slicing infused with AI: To expand the allure of network slicing, it is envisioned to be paired with AI capabilities, rather than a strictly connectivity-based proposition. For example, it could relate to network slicing infused into a service for real-time translation of multiple languages at the same time.

3. Hosting workloads for AI service providers: Akin to how telecom operators have deployed cache servers in their networks to enable CDN services delivered by hyperscalers, operators may again play a similar role in the AI service space. In this context, AI-RAN would serve as a platform to provide additional capacity for AI solution providers to deploy workloads, for example to supporting a generative AI service. At MWC 2025, ARM demonstrated how the RAN could support a digital human avatar, called James.

How will resources for these use cases be allocated?

Operators have a strong pedigree in analysing traffic profiles, and accordingly optimising network conditions. It is therefore a natural extension for operators to leverage this knowhow to develop a dynamic environment where enterprise workloads can use the idle capacity, predicated based on these traffic patterns. This constant monitoring of workloads is akin to efforts in the cloud computing industry where workloads must be constantly monitored to ensure the correct resources are provisioned. In order to achieve this ambition operators must provide a modular, scalable and software driven platform.

Dr. Alex Jinsung Choi, a South Korean luminary in the global mobile telecommunications sector, boasts over three decades of industry experience. Currently, Dr. Choi is working at SoftBank Corp.’s Research Institute of Advanced Technology and continues to champion innovation at the intersection of AI, mobile technology, and network architecture. He is also currently the Chair of the AI-RAN Alliance. He is dedicated to advancing technology frontiers, crafting strategies that enhance both business and consumer experiences, and guiding the industry toward a more connected, intelligent, and open future.