Samsung is betting that AI-RAN will not necessitate moving the entire baseband onto GPUs, and is instead pursuing a CPU-centric architecture in which GPUs are selectively deployed to accelerate workloads such as channel estimation, senior executive Alok Shah has told The Mobile Network.
For Shah, Vice President, Network Strategy at Samsung Networks, the greater shift in importance compared to arguing over which chip platform will drive AI-RAN comes from moving to a fully virtual and software RAN architecture – something that will allow operators to prepare for 6G while still deriving efficiency gains from 5G networks.
Samsung’s Alok Shah: open to selective GPU acceleration, banking on software RAN experience to win out vs incumbency in 6G
In an interview with TMN, Shah said that Samsung’s architecture remains CPU-centric, with GPUs used as accelerators for specific workloads where the economics justify it. He identified channel estimation and channel prediction as the main GPU candidates, and said Samsung has already demonstrated channel-estimation capabilities in the lab, although there is no commercial rollout date yet.
“I think there’s still some in the industry who position AI-RAN as requiring GPU acceleration on the baseband. That’s not our philosophy. We’re very much about a software-centric architecture that is running primarily on CPUs, with on-board acceleration and an architecture that’s built to support the inclusion of a GPU when the business case makes sense – for example, when an operator is looking for an additional boost of performance in high traffic sites, or when an operator has the right business case in place to support inferencing at the edge.”
Shah added that a lot of AI-for-RAN capability can run on the CPU – “so for us it’s really about having a flexible architecture in place that’s centred around the CPU to do most of the of the compute work.”
“And then there’s a category of channel estimation and channel prediction where you can benefit from the parallel processing that you get out of a GPU platform. I would say that’s the main category that we’re looking at, where an additional boost has real value,” he added.
Shah said Samsung has demonstrated capabilities specifically around channel estimation in laboratory Proofs of Concept.
“I don’t think there’s anything we’re ready to say yet about the timing of commercial rollout of that sort of capability. But we’ve done a fair amount of work on the implementation side to be able to leverage those kinds of capabilities that can really only run on a GPU.”
While Shah said that while Samsung has proved that its RAN software can be ported for different CPU platforms relatively easily, when it comes to leveraging GPU acceleration there are a “couple of factors” to take account of.
“Clearly CUDA on the NVIDIA GPU platform presents its own layer that has to be developed on. So there’s maybe the question of, ‘Are you trying to run everything on the GPU?’ In which case, you have to port everything to CUDA, your stack. [Note that this is the path Nokia is taking, moving its stack to Nvidia’s Aerial Ran Compute platform TMN]. Or do you envision GPU as your accelerator for specific needs that you might have – for example, channel estimation? Let’s say this is where the compute requirements are going to go above and beyond what the CPU can support, in a case like that you’re talking about running a specific workload on that GPU platform as another accelerator that’s sitting on your architecture. And it’s not even porting really – it’s sort of new code that’s developed to take advantage of that: the GPU parallel processing can be built on CUDA, and then rolled out at sites where the benefit has meaningful value to the operator.”
“That’s how we’re thinking about it. That it’s likely not necessary to develop your entire stack on the GPU, right? You use it as another accelerator on the platform, where you’re building software specifically for that.”
As for deployment, that will be a matter for operators, says Shah, describing it as an operator-by operator decision. But we should expect Samsung to have that capability to move that part to a GPU if required by customers.
“We’re very customer driven, and from that respect, I think over time there may be cases where it makes sense to do that, at least at a subset of sites, and certainly we would be open to that path.”
Virtualise the RAN
However, there’s one more fundamental step that Shah counsels, before the topic of GPUs and AI-RAN arises, and that’s the foundational aspect of virtualising the RAN. And it is here that he identifies an advantage for Samsung over its competitors. Samsung is positioning its existing vRAN platform as the foundation for 6G, and also arguing that operators should begin shifting to software-based RAN architectures during the 5G cycle rather than wait for the next generation of mobile technology.
It’s great that at least some of the other suppliers are moving in the direction of software architectures – but I think the curve to get to a level of sophistication and experience takes time.
“When it comes to the competitive dynamic, and before you go deep into the kinds of gains that will come from these sorts of [GPU-accelerated] boosts, there’s a step one around having a software-driven architecture and rolling it out at scale, and that’s a heavy lift too.
“It took a while for us and for our customers who are on vRAN to learn how to deal with different tool chains and different software architectures. There’s really no way to skip steps. It’s great that at least some of the other suppliers are moving in the direction of software architectures – but I think the curve to get to a level of sophistication and experience takes time.
“The key is that I think we’re far enough along that curve, that we’re in a position to be able to bring in additional acceleration or bring in a different type of CPU because we’ve gotten past the hump.”
As for how customers make that transition, Shah says it will take time for an operator that’s got tens of hundreds of thousands of non-depreciated sites to transition.
The incumbent’s always in the driver’s seat, right?… So as a challenger we have to make a case that there are limitations in what the incumbent is offering that would justify making a change
But also, “We see quite a bit of interest and recognition from operator customers and not-yet customers, who see the value in in a software-driven architecture, recognise that it’s very much going to be what a 6G platform looks like, and see the value in in starting to plan out that transition. Ultimately it’s all about timing and comfort level with the shift.”
Although operators will make cautious noises about 6G, and recently backed a network sharing concept that would appear to give primacy to incumbents, for Shah, 6G does represent an opportunity for Samsung.
“I mean honestly, the incumbent’s always in the driver’s seat, right? That’s just the reality of the world that we live in. So as a challenger we have to make a case that there are limitations in what the incumbent is offering that would justify making a change – and we’ve had success over the years in doing that.”
“Every G has turned out to be a point of entry for us,” he added. “Historically, that’s been the way it’s worked. I do think when operators are considering the target architecture for 6G, I don’t think they’re viewing it as about keeping 5G on a on a legacy, purpose-built platform and putting 6G on AI-RAN or V-RAN. I think they’re looking at refreshing their 5G as well. It’s what we saw in in in the 5G transition from 4G in in many cases – and we’ve even deployed 2G on vRAN.
“So I think I think the operators view it more holistically than that. 5G gear will be fully depreciated and will probably have some limitations from a performance perspective that would necessitate a refresh. So I think we certainly view the next couple of years as an opportunity to grow further.”
6G thinking
Shah adds that operators are pushing for a cleaner evolution between generations, with investment driven by use cases and monetisation rather than wholesale changes to the radio. He says this is already reflected in 3GPP specifications and argues that 6G should deliver performance improvements without requiring wholesale changes to modulation and coding.
“We believe that we’ve already got the platform and software architecture in place to be 6G-ready. That’s why we think it’s logical for operators to be looking at that architecture well in advance of the 6G cycle – whenever the timing is right for you as an operator – because then you can start to take advantage of those benefits in your on 5G as well.”
This thinking mirrors comments made to journalists by Shah’s colleague Woojune Kim, President & GM, Global Head of Networks Business, Samsung, back at MWC Barcelona 2026.
At the time, Kim identified that Samsung has created a software-based RAN that can run on commodity hardware, and add a GPU accelerator into the mix where it makes sense.
Kim said, “Seven years ago, we said we are not making a radio network anymore. We’re making a data centre – all servers and software. Trend-wise we could see everything was going in that direction.
The benefit is the flexibility it brings.
“You don’t have to figure out what’s going to happen three years from now. We will ride the trends. AI is a great example. People have asked us, ‘Are you going to revamp your hardware roadmap? Are you going to create new chips. I say I don’t think I need to do anything special. And the reason is because we’ve created server-based solutions which run software on commodity systems.”
