AI-RAN a Game Changer for Mobile Networks: Interview with SoftBank Corp. Principal Fellow Alex Jinsung Choi

Since it first announced its AI-RAN concept in 2023 as a foundation for next-generation mobile networks to support a society that coexists with AI, SoftBank Corp. (TOKYO: 9434) has made significant progress in bringing its vision to reality.

One person instrumental to this progress is Dr. Alex Jinsung Choi, a Principal Fellow at SoftBank's Research Institute of Advanced Technology and Chair of the AI-RAN Alliance.

SoftBank News sat down with Choi to learn more about the transformative power of AI-RAN, and how it can benefit mobile network operators, enterprises and consumers.

AI-RAN can reduce costs for operators and enhance the customer experience

What is AI-RAN and why is it significant?

AI-RAN actually means the introduction of "AI-native" radio access networks (RAN). That means we're trying to reinvent the traditional RAN from various aspects - architecture, operation, design, deployment - to make it AI-native. So we're reinventing the wheel in a way where we innovate how we design and operationalize RAN while trying to generate new AI services and applications, which will give us new sources of business revenue.

AI-RAN is all about incentives. The first incentive is about bringing AI-RAN to mobile network operators (MNOs). There are three key benefits for them. One has to do with the total cost of ownership (TCO) of their networks. TCO has been a major burden for MNOs for a long time, and it's getting bigger. Operators typically spend billions every year on TCO. Our goal with AI-RAN is to significantly reduce those costs.

The second incentive for MNOs has to do with new revenue generation. In the telecommunications industry, top line revenue is not growing that much. So when looking at how we can break through this top line, we need a new business. To create a new business, we need a new value proposition and new enablers. AI-RAN comprises three parts: "AI for RAN," "AI on RAN" and "AI and RAN." The "AI and RAN" part is particularly geared to new business models. One example is AI inferencing on the edge of mobile networks (known as Multi-access Edge Computing, or MEC), which could be large language model (LLM) inferencing or any kind of generative AI inferencing. Our belief is that with the adoption of this kind of AI, and with inferencing services at the edge of networks, we can create a new business model and a new stream of revenue.

The third incentive is to see if we can significantly improve the customer experience. AI-RAN is a very customer-oriented value proposition. Customers are still struggling today with traditional key performance indicators like network capacity and coverage, and there's still a lot of work to be done. AI-RAN has advantages over the traditional RAN with improved performance, coverage and capacity.

At the same time, we want to introduce really low latency services. One example is a real time chatbot, a voice-based chatbot, not a text-based one. AI assistants or chatbots have all these features, and you'll be able to use them with a voice-based interface. In voice mode, everything has to be in real time because there may be urgent situations when a person needs help. So every millisecond counts. With AI machine learning and local edge AI inferencing, end-to-end latency can be dramatically reduced.

Imagine you need a ticket urgently. You're standing in line and you want to buy one, but have no information on the spot about what tickets you want to purchase. So you just quickly use your voice to say, "I need information on this right away." And you get an immediate response. But if you use the typical centralized cloud services, you might have to wait as there may be long periods of latency.

These days chatbots are being equipped with what we call reasoning features. Reasoning means they can provide much smarter, intelligent answers and responses, but at the cost of higher latency because machines need time to think about what the best answers are. In view of that trend, latency is increasing.

What the customer wants is the best quality answers with low latency. It could be more than just a voice response. Materials like video, voice or text could come to your device in real time. That's the ultimate future experience. Only with the support from edge AI inferencing can we make these types of top-quality, AI-enabled customer experience services happen.

AI-RAN will definitely be a game changer. But at the same time, we have to understand that hyperscalers also have similar visions and strategies. They want to deploy their inference engines as close as possible to their customers and users. Their approach is to just build more and smaller AI data centers.

As MNOs, we have a window of opportunity, and we shouldn't miss it. SoftBank is the front runner in AI-RAN, and it has a great opportunity to take the lead.

Utilizing unused network capacity to provide new AI-driven services

Could you tell us in more detail how operators can lower their costs and grow their revenues with AI-RAN?

The logic is simple. The average utilization of a RAN is below 30%. Mobile traffic patterns are predictable in most cases, and seasonality is predictable, unless something unexpected happens. This means operators can predict how much capacity will be available and when and where. So why not monetize this extra capacity?

Operators have a choice. They can completely switch off their network to save energy or monetize unused capacity. The former can make sense in countries that are energy sensitive, like in Europe, for example, as money can be saved immediately. But from a technology standpoint, some part of the network needs to be kept on in case there is sudden traffic due to an emergency of some kind. So the energy savings are not really as high as desired.

The alternative is to monetize the unused capacity, which leads to the question, "What kind of traffic should be hosted?" AI traffic is a good candidate.

Imagine a situation where you have a car, and it needs a software update. If you're using a tablet and it needs to be updated, you're bothered every once in a while with an update request, and you have to download a gigabyte or so of software. That means software needs to be downloaded frequently. These kinds of updates can be done overnight when traffic isn't heavy. That's one example.

Another example is with enterprises. They deploy thousands of IoT devices, including cameras and sensors, which collect raw data on agriculture, for example. But raw data volumes are very large, so some pre-processing needs to be done with that data before it's sent back to the software. That's another compelling use case: the RAN collects the raw data, and edge computing does all the processing by utilizing unused capacity. This kind of work can be done when the RAN workload is low. The data stream is sent to the server after pre-processing, and enterprises can take care of the remaining processing or store data.

There could be many more examples. For example, CCTVs collect data, but the analytics could be done at the edge of a network, during the day or offline, unless there is some kind of urgency.

You mentioned voice-based chatbots with low latency as a potential use case. Are there any other compelling applications for the consumer?

With AI-RAN, customers will benefit from a significantly enhanced AI-driven experience. I think robotics is going to be the new market category. There's still a way to go with humanoid robots, but it seems like large tech companies are really betting on them. Robots need a commander, and edge AI inferencing will be that commander.

SoftBank's demonstration conducted at Keio University Shonan Fujisawa Campus ("SFC") in November 2024 combined LLM inferencing and low latency. In the demo we showed two different robots - one using a traditional RAN and our robot enabled with AI-RAN. There was a difference in terms of latency, which would be important in the case of security applications. For example, if a thief steals something, the robot should recognize the thief and immediately give chase. But if latency is high, the robot cannot do that as the camera would lose visibility. But a robot operating at extremely low levels of latency can access a camera continuously and follow the thief. In this case robots powered by AI-RAN can contribute to human safety, protecting them like bodyguards.

Fast-growing AI-RAN Alliance developing an ecosystem to enable AI-RAN

These use cases sound really exciting. But to make AI-RAN a reality, is it true that different industry players need to work together?

AI-RAN was first conceptualized by SoftBank, Arm and NVIDIA, but they realized they could not achieve their ambitions on their own and needed more ecosystem partners. That's what led to the AI-RAN Alliance. We're very young and just celebrated our first anniversary.

The Alliance membership is well balanced. We have MNOs in the US like T-Mobile US and Boost Mobile, three Korean operators, Globe Telecom in the Philippines, Indosat Ooredoo Hutchison in Indonesia, Turkcell from Turkey, and we expect more to join. We also have network vendors - Ericsson, Nokia and Samsung - and there are academic institutions like Northeastern University and The University of Tokyo, and startups like DeepSig. I joined the AI-RAN Alliance as Chair in July 2024. Things have accelerated since I joined and we established all the necessary structures and operational principles, such as charters for working groups, milestones, deliverables and use cases. The individual working groups have already discovered interesting use cases.

In addition, we have two supplementary committees: the Technical Steering Committee and Marketing Steering Committee. We also recently announced the launch of two flagship initiatives, Data-for-AI and Test Methodology, and the creation of four AI-RAN Alliance-endorsed labs. So I would say the organization is quite stable now and up to speed. Membership-wise, we started with 11 members and now we have 84. Our awareness has increased significantly, so we have more than 70 companies in our application pipeline. If we approve and finish the membership approval process, sooner or later, that would give us close to 160 members.

The AI-RAN Alliance also faces challenges. Some people still don't fully understand what the AI-RAN Alliance is trying to do yet, despite our efforts to proactively engage industry stakeholders, especially in the operator community. We actively share our vision, goals and working group activities but there is some confusion because other organizations like 3GPP and the O-RAN Alliance - which I chaired for the last two years - are engaged in similar activities.

I clarify any confusion by stating that the telco industry has a different technology life cycle. Our industry lifecycle is simple - you do the research, which can be called pre-standardization. And then you standardize based on the outcome of your research. This is standardization. After developing and publicizing the standard, you enter implementation, or the post-standardization phase. Then there's testing certification and commercial rules. So these are the five steps of the cycle.

3GPP and the O-RAN Alliance are taking care of steps two and five - standardization and testing and certification. This is the job of standards development bodies (SDOs) like 3GPP. We have no intention of duplicating their work. Our focus is on step one, pre-standardization, which has more to do with research. That's why the AI-RAN Alliance has so many academic members, which represent a large portion compared to other industry forums. This makes sense because our focus is on pre-standardization and post-standardization. Post-standardization means we implement and develop AI-RAN reference implementation. And we showcase it in the form of technology demonstrations or blueprints. Others, like third parties, are responsible for the official testing and certification process so the technology can be commercialized.

Before you became Chair of the AI-RAN Alliance, you headed the O-RAN Alliance. You've played pivotal roles in other industry associations as well. What did those experiences teach you?

I follow three principles based on what I've learned. One is "don't try to do it by yourself." In some cases there are leaders who are very smart and want to drive many initiatives, and they expect that others will automatically follow. That works well within a company because a team has to listen and follow. This is not the case with an industry alliance, however. You have to be patient, and even if you believe you have the capability to drive something and take the initiative, you better let others do that. It has to be about collaboration, and partnership has to come first. That's the hard lesson I learned.

And the number two lesson is that the telecommunication industry technology lifecycle is pretty long. That means you have to make sure you get the proper support from industry stakeholders. In many cases, this means CTO- and CxO-level commitment and support from MNOs. Without their support, it won't go well, no matter how brilliant the technology is. You have to get them to buy-in.

That leads to another principle, "socialize and entertain." As a technologist, that's not my strong point, but customer buy-in requires preparation, and real decisions on commercialization come from the top. For an innovation project, it's possible to convince working level executives that something is a great idea. But when it comes to actual commercialization and the procurement phase, it needs to be explained to decision makers. Sometimes that means going back to the starting point, which often happens. But if you know this in advance and set the strategy accordingly, things should work out OK. MNOs tend to have larger budgets than companies in other industries, so in many cases they spend money, at least on pilots or trials. But this generosity may disappear if the telecommunications industry is continuously struggling. That seems to be the case as our market is close to saturation. Vendors in the industry are struggling because their customers are struggling.

The final lesson I learned is that you have to have a very clear vision. In the telecommunications industry the life cycle is so long. Once technology generation can last for 10 years, which can feel like forever. So you have to be very clear about your vision and demonstrate leadership. Without a vision, no one will follow you. Having a vision means you set the direction.

Are there any challenges in the short-term for the AI-RAN Alliance?

In the short term it's all about efficiency. Our industry has been growing so fast. If you look at all the different corners, you'll find a lot of redundancies. There's a lot of room for efficiency improvements: network efficiency, operation efficiency, financial efficiency. You have to discover these issues and remove them. At the same time, you have to introduce digitalization and actively adopt AI. Some businesses are reluctant to use chatbots and LLMs because they have some weaknesses, like hallucinations.

But people need to understand that no single technology is perfect. They need to find workarounds and be aware of the issues. To avoid LLM hallucinations, for example, work needs to be done and investments need to be made. The good news is that the technology is advancing fast, so the likelihood of hallucinations is now much lower, and the accuracy of the latest LLMs are around 95-97%. This is a good number, and much more needs to be done. But there's no need to wait. Companies should embrace AI technology and integrate it into their daily business.

What are your thoughts on the AI-RAN Alliance's future?

The AI-RAN is new and can step up to take the initiative. AI-RAN will take the lead in the future for RAN evolution and transformation, and bring added and new value to MNOs, the technology customers. With our AI-RAN contribution, telcos can avoid the fate of being a "dumb pipe" service provider.

Related Articles

(Posted on April 3, 2025)
by SoftBank News Editors