this episode breaks down what AI-RAN really means, whether GPUs are actually needed in the mobile network, how 6G could become more of a software upgrade than a hardware refresh, and why Open RAN may be moving from industry hype to operational plumbing.
The conversation also tackles one of telecom’s biggest business questions: Can they avoid being relegated to dumb pipes in the AI era?
Catch the video at top, listen to the audio edition and read our transcript below, or watch this and future episodes on YouTube.
To learn more about the topics in this episode, check out:
- KT, SK Telecom tap Samsung for AI-RAN projects
- AT&T, T-Mobile and Verizon plans diverge on AI-RAN
- Nokia’s AI-RAN pitch gets reality check in new study
- 6G is more than another ‘G,’ says T-Mobile CTO John Saw
- Engineers craft 6G. It’s marketers' job to sell it
- US calls on allies to rally around 6G
This podcast is written and hosted by Diana Goovaerts. It is edited by Diana Goovaerts and Matt Rickman. Liz Coyne is our executive producer. Special thanks to guests Ericsson's Joe Constantine, Samsung's Shamik Shah and Mobile Experts' Joe Madden.
Diana Goovaerts, Fierce Network: What is AI-RAN and does it need GPUs? What's coming in the 6G era? We're tackling these questions and more as we explore what the future holds for the RAN.
Joe Constantine, Ericsson: This is like the starting point where we see how these intelligent fabrics, cloud, AI, and, and, and 6G, 5G, 6G are becoming one entity. So, if we look at this five, six, 10 years from now, this is gonna be a continuum.
Diana Goovaerts: I’m Diana Goovaerts, and this is The Five-Nine.
The radio access network, or RAN, has always been the sprawling and expensive heart of the mobile world. But as 5G matures and 6G starts to take shape, the RAN may be heading for a very different future, one defined less by hardware refreshes and more by software, AI, cloud and compute.
Today, we are looking at the future of the RAN with help from three guests, Ericsson North Americas CTO Joe Constantine, Mobile Experts founder and President Joe Madden, and Shamik Shah, who is director of RAN Systems Engineering and Performance at Samsung Electronics America.
And we're starting with the hottest, blurriest term in mobile networks right now, AI-RAN.
AI-RAN has become one of telecom's favorite new phrases, and depending on who you ask, it can mean very different things
Joe Madden, Mobile Experts: AI-RAN is, is a very confusing term because it's two major things rolled into one. You could say that AI-RAN is using AI techniques to implement better RAN performance to enhance the RAN, higher capacity, lower energy of consumption, some of those kind of things.
And you'd be right. That's AI-RAN. But you could also say that AI-RAN is edge computing, AI inference at the edge, which is a completely different concept. That's a different business altogether. And so, rolling those two things into one term is the reason why the industry is very confused right now.
Diana Goovaerts: Samsung's Shamik Shah takes the broader vendor view. AI-RAN is both network for AI and AI for network. In other words, the network has to become software-based enough to support AI applications, while AI itself starts to optimize planning, operations, capacity and performance
Shamik Shah, Samsung: You can have two different perspectives, right? Network for AI, AI for network, right? So, when I'm saying network for AI, meaning you have your base network architecture, which is a software-based network, not just for RAN, as I mentioned earlier but that basically gives you the foundation to be able to bring in AI or automation, all of those different applications on the network.
Now, the next part of it is AI for network, is when we have a cognitive NOS solution. So, what that does is it has its AI applications and automation, which looks at network planning, operations, capacity, as well as performance.
So, these are the two different angles of how the networks will evolve or is in the process of evolving, I should say
Diana Goovaerts: Once AI-RAN enters the conversation, GPUs are never far behind. And Nvidia and some of its partners have pushed a future where accelerated compute becomes part of the mobile network. But our guests were a bit more cautious. Shah said from Samsung's point of view, most AI applications can be handled by CPUs.
GPUs only come in when the application requires heavier parallel processing or large-scale data work.
Shamik Shah: So the network won't be like CPU only or GPU only. Okay? Most of it, most of the AI applications can be managed by the CPUs. Now, when it comes to very, very data processing intensive applications like parallel processing or you have to learn like vast amounts of data, that's where you would bring in a GPU.
But otherwise, a GPU will not be required. Most of it can be managed by the CPUs. And even if you... Let's say if you want, if you have certain apps or certain applications that requires higher processing power, you can always bring in a GPU, put it on the core server, move those applications through software to run on the GPU, and that's easily supported.
So, it is going to be very application specific.
Diana Goovaerts: Madden offered a similar take, arguing GPUs may matter for some AI inference use cases but won't necessarily be needed at every cell site. His example of a food delivery robot deciding whether or not to cross a busy street makes the point in plain English. Not every edge AI use case needs ultra-low latency.
Joe Madden: What we're finding is that for a security robot, a construction robot, a drone delivering a package, a food delivery robot delivering my burrito, it's okay to stop for a second, make a decision, then go ahead because there's no human involved, there's no pressure for that robot to continuously move forward.
So we're finding that, back to your question, we're finding that GPU resources at the cell site are really not necessary. You could have a rack of GPUs centrally located serving the entire country with a couple hundred milliseconds extra delay. That would be no problem.
Diana Goovaerts: Constantine said for Ericsson, the underlying hardware doesn't matter at all. What matters more is the use case.
Joe Constantine: We at Ericsson have built our software independent of any underlying infrastructure, which actually means we're the only company on the planet that can run our software on private, public and hybrid cloud. We can run on any processor technology, Intel, AMD, ARM, and we can run on any compute, GPU, TPU, CPU.
So, for us, it doesn't really matter about the underlying hardware. It's actually the use case that will manifest where that workload needs to sit, what kind of hardware we need to have. So again, the key is for a successful Western stack and trusted stack is to be able to open the door to the community society so everyone can innovate on top.
It's how we've built our software.
Diana Goovaerts: The next piece of the RAN's future is software. If operators can reuse more hardware and push upgrades through cloud-native software, the jump from one generation to the next may start to look less like a rip and replace and more like a continuous evolution.
Joe Madden: Generationless is a good word. The fact is 6G and 5G won't be that different. My view is that the 6G core is really just the 5G SA core, and they're gonna put some AI features on to optimize it better and things. But the changes are minimal.
And it's the same thing on the radio side. 6G radio will look very much like the 5G radio, and yes, we'll have some AI, and we'll be able to augment it and distort it to get better performance and things. But that's an evolutionary move, not a revolutionary move.
The operators have put their foot down and said, "You know, we don't want to do giant investments, tens of billions of dollars with every generation." They're not making the kind of growth that they used to get that would justify that kind of capital investment. So, they need to reduce their capital investment and come down to something which is more maintain and do incremental upgrades through software.
Diana Goovaerts: Shah echoed that position, saying Samsung sees 6G as more of a software upgrade than a hardware refresh. Even if new spectrum and new radios still come into play, AI-RAN and Open RAN, too, will be the foundation for that upgrade, he said.
Shamik Shah, Samsung: All of this combined lays the foundation for 6G.
So now 6G, we believe it won't be like a hardware refresh, it will be more be software upgrades, I would say. 6G can bring in new spectrum obviously, which will be-- which basically will be new radios and new technologies, beamforming. We are doing trials with the seven gigahertz band, which we call the FR3 band.
So, all of these will be on top of your existing network. So, 6G will be a software upgrade, not a hardware refresh, is how we are seeing it.
Diana Goovaerts: Constantine added an important caveat. While telcos can adopt a CI/CD approach and real-time software upgrades, some things just remain stubbornly physical. You still need things like radios and filters, antennas and power amplifiers.
Joe Constantine, Ericsson: If you look at CI/CD, that's happening today across the board with our products. A true cloud native network should be able to be upgraded real time while you're operating the software. So that's something that's happening and will keep happening.
To date, I have not seen a software filter or a software power amplifier that you need in a radio on a software antenna. So, those will actually still be physical assets. And I don't see in my lifetime this changing. It's the law of physics. You need the power amplifier, you need filters for spectrum, et cetera.
So that radio is gonna always be hardware.
Diana Goovaerts: And then there's Open RAN. Open RAN has been through multiple hype cycles, and despite the noise dying down, our guests said Open RAN is not dead. They frame it as either normalized, narrowed, or folded into the larger software and AI architecture conversation.
Joe Madden: This is a bit of a very public debate and a very loud debate, you know, has Open RAN been a failure or a success? What I would say is that it's done pretty much what we expected it to do.
We have vendors which sell radio units on top of other people's baseband, and the number of vendors that survive that sort of Darwinian process is pretty small. So we have only a few vendors that have survived that. But it's now a reality, and there are companies that are commercially using it.
At the same time, we have these non-real-time applications, they call them rApps, which allow a third-party software developer to take their software and improve the performance of the RAN in some way or automate something and that's already working quite well in the field.
So, we have two success stories out of this long Open RAN saga which I think justify the effort.
Joe Constantine: I think the fact we don't talk about it as much is because it's an integral part of what we do. I mean, at one point in time when things—When you talk about the future, we're talking about future inferencing, and we're talking about the edge, et cetera. They are about to happen. But when they happen, you start – you move on to whatever’s next.
Diana Goovaerts: All of this leads to a bigger question. What do operators actually get out of this future RAN? AI traffic may increase data volumes, but that does not automatically translate into revenue. Madden argued that GenAI alone could leave operators stuck as utilities, moving more bits without capturing more value.
Joe Madden: For 20 to 30 years the data growth has been relentless. But at the same time, their ARPU hasn't grown. AI doesn't really change that.
Gen AI just adds more and more traffic without really an opportunity for a big bump in revenue. And so I think that's a dangerous situation for the operators. They're becoming a utility that really is just trying to continue to have a slim profit margin without any prospect of growth in the future.
Diana Goovaerts: But that doesn't mean all hope is lost. Constantine and Madden both pointed to physical AI and robotics as a potential saving grace. Those use cases need widespread coverage, throughput, authentication, governance, and in some cases, deterministic latency and jitter guarantees – things telecom operators may be better positioned to provide than hyperscale cloud providers.
Joe Madden: I think if we look at robots compared with gen AI or other aspects of AI, you know, I think it's the one opportunity which really needs better connectivity. It needs certain things like a latency or a jitter guarantee, and it needs high throughput in some cases, and it needs widespread coverage.
So, it has aspects to it that are really a good fit for the mobile operators, whereas gen AI is just another source of data. With gen AI, the telecom guys have a hard time breaking out of being a dumb pipe.
Joe Constantine: Telecom networks can become actually a spatial data engine, turning sites and devices into beacons and sensors that feeds world models, which we don't have today, giving robotics, vehicles, agents rich context about the physical world around them.
If we look at robotics, to your example, where does it sit? Does it sit in a robot or does it sit in a network? I think gonna be both. You would have to have both. I don't think the device itself, without any connection to other devices and without any connection to the real world models and context awareness can actually operate in a society you would need that fabric.
The way I look at it, the intersection between 6G, AI, and cloud, that is vastly different than anything we've seen in the past.
Diana Goovaerts: The future of the RAN is still taking shape, but a few key ingredients have already surfaced. It is partly AI optimization, part software architecture, part 6G evolution, and partly a fight over who captures the next layer of value from connected machines.
One thing is clear: the RAN is not standing still.
It is becoming more intelligent, more software-driven, more tightly linked to compute, and it will remain a critical force powering the technologies of the future.