Exciting Results from SoftBank and NVIDIA's AI-RAN Trial Could Ignite a New Era in the Telecoms Industry

By Larbi Belkhit | 4Q 2024 | IN-7617

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.

In November 2024, during NVIDIA's AI Summit Japan, SoftBank announced that it had completed a successful outdoor trial in partnership with NVIDIA. The End-to-End (E2E) solution was the world's first 5G Artificial Intelligence (AI)-Radio Access Network (RAN) field trial running on an NVIDIA-accelerated computing platform and was based on a full-stack Virtualized RAN (vRAN) software integrated with a 5G core using NVIDIA's Aerial AI platform. It ran on the NVIDIA Grace Hopper superchip (Central Processing Unit (CPU) + Graphics Processing Unit (GPU)), NVIDIA Bluefield-3 (Network Interface Controller (NIC)/Data Processing Unit (DPU)), and Spectrum-X for fronthaul and backhaul networking. The trial consisted of 20 4T4R radios, as well as 100 mobile User Equipment (UEs) and utilized open, rather than proprietary interfaces.

The core software stack for the trial included:

• SoftBank-developed L1 functions (such as channel estimation and channel mapping) using NVIDIA Aerial CUDA-accelerated RAN libraries
• Fujitsu software for Level 2 (L2) functions
• Red Hat's OpenShift Container Platform as the container virtualization layer to enable different types of applications to run on the same underlying GPU computing infrastructure
• SoftBank-developed E2E AI and RAN orchestrator to enable seamless provisioning of RAN and AI workloads.

The trial resulted in some interesting findings. First, when the solution was running in RAN-only mode (i.e., solely focusing on RAN workloads with no external AI workloads), the server had 40% less power consumption than the best-in-class traditional RAN system today, although it was not announced which vendor equipment this setup was compared to. Second, when running the solution in a one-third RAN and two-thirds AI workload distribution, findings showed that the overall investment resulted in a 219% profit margin when considering both Capital Expenditure (CAPEX) and Operational Expenditure (OPEX) costs and can generate 5X the revenue of the CAPEX investment over 5 years. Even when considering a two-thirds RAN and one-third AI workload distribution, this resulted in a 33% profit margin and 2X revenue from the CAPEX investment.

Such staggering results make for big headlines, and rightfully so. The telecoms industry (especially in Western markets) has had a tough time of it since the rollout of 5G networks, so results from an AI-RAN trial such as this will definitely excite and spur operators to conduct some trials of their own. However, there are still some things to consider regarding this trial's findings:

• Key Profit Margin Assumptions: While NVIDIA and SoftBank are not hiding this fact, it is important to re-emphasize. All the CAPEX and OPEX assumptions are SoftBank-specific and use local country cost assumptions.
• Small Scale: The trial is still incredibly small, with only 20 radios and 100 UEs. Furthermore, none of these radios were Massive Multiple Input, Multiple Output (mMIMO) and all operated on a single 100 Megahertz (MHz) frequency. Commercial deployments are much more complex, which could affect the cost and energy-efficiency benefits.
• Energy Efficiency: While the energy-efficiency gains are indeed impressive, it is not clear whether these results consider the utilization of power management technology at the server level (which the energy-efficiency gains refer to in this trial) when compared to the x86 or traditional RAN systems.

The ability to run external AI workloads gives the AI-RAN initiative a much better chance to succeed commercially than the Open RAN initiative. The O-RAN Alliance's work focuses on the standardization of open interfaces within a RAN, whereas the AI RAN technology is about developing new use cases and revenue avenues on top of the RAN. This trial by SoftBank showcases three AI applications-remote support for autonomous vehicles over 5G, factory multi-modal AI applications, and robotics applications. Furthermore, AI-RAN technology does not rely on the open interfaces of the O-RAN Alliance to be deployed; due to its software-defined nature a simple software update can enable an operator to simply switch when necessary.

Another thing for the industry to consider is that while NVIDIA's AI-RAN technology can run on open interfaces, a form of lock-in will appear as it currently requires NVIDIA GPUs and, more specifically, NVIDIA CUDA, for the operators. This is not necessarily a bad thing, as recent developments indicate that a form of lock-in is considered necessary to allow for performance and cost efficiency, especially for the radio. For NVIDIA, this is yet another success story for its business strategy to further integrate themselves as a necessity for late-stage 5G and future 6G networks over the solutions currently being provided by AMD and Intel.

This is a promising development for the entire telecoms industry, with the potential to address the economic challenges it has faced in recent years. To fully capitalize on AI-RAN technology, mobile operators need to adopt a more proactive approach. Participating in initiatives like the AI-RAN Alliance and conducting trials across diverse regions can help operators better understand and unlock the revenue potential of Generative Artificial Intelligence (Gen AI) applications and Artificial Intelligence (AI) inferencing, especially as regional economic conditions influence network CAPEX and OPEX. Additionally, future trials should incorporate a wider range of radios, including technologies such as mMIMO.

When it comes to energy efficiency, future studies should directly compare the power consumption of AI-RAN solutions with and without the use of advanced energy-saving technologies to provide a clearer view to operators on the real-world energy efficiency benefits of AI-RAN. This will also bring the OPEX effects of fully utilizing all infrastructure to the forefront and, when it comes to commercially deploying AI-RAN networks, helps operators plan ahead with a much clearer picture.

From a commercial deployment perspective, it is still incredibly early in the development of the AI-RAN initiative. The AI-RAN Alliance was only established in February of this year (for more, see ABI Insight "AI-RAN Takes the Spotlight as Focus Shifts from Open RAN" on the AI-RAN Alliance). On the back of the announcement of the trial's results, SoftBank announced its AITRAS solution-a reference architecture for integrating AI and RAN workloads onto a single platform. It plans to begin offering this solution to other operators from next year and plans to commercially deploy its own AI-RAN network from 2026. This all indicates that the commercial deployment of AI-RAN networks (on a global scale) should only be expected post-2027, given that the Japanese telecoms market is typically one of the earliest adopters of new technologies.

The larger question remains-should telcos take the leap and deploy these AI-RAN networks? My opinion is that it is more a question of when, not if, in leading telecoms markets such as the United States and Europe. T-Mobile is already working with NVIDIA, Ericsson, and Nokia at their jointly founded AI-RAN Innovation Center and is the only other operator that is a founding member of the AI-RAN Alliance (alongside SoftBank). Therefore, as long as trials for operators prove to showcase the energy and revenue benefits that SoftBank found, there will be no reason to not plan a commercial deployment.

Furthermore, it also feeds into many operators wanting to make the 6G network upgrade as software-heavy as possible, and the software-defined nature of the AI RAN architecture means the radio will be the primary hardware refresh of the RAN for telcos and does not affect their new revenue streams such as Graphics Processing Unit-as-a-Service (GPUaaS). However, the total revenue opportunity for these new offerings and revenue streams will be predominantly limited by the maturity of the AI industry, especially in the enterprise segment, which could slow down the adoption of this network architecture.