Earlier this week, TMN spoke to Aji Ed, Head of AI‑RAN and Cloud RAN at Nokia, about the company’s launch of its new AI-RAN portfolio. Here, in his words, he addresses how Nokia got to where it is now, how it has responded to customer concerns about AI-RAN cost and power performance, and outlines the five pillars that underpin the AI-RAN portfolio.
“This has been a journey for over two years now. One of the key milestones was last year at GTC with NVIDIA; that was the starting point of a greater collaboration. At that time we announced potential paths for deployment but not everything was crystal clear.
“We were not clear about what kind of benefits it [AI-RAN] can bring in terms of the accelerated computing, and what kinds of openness and programmability it can bring.
“So for the past 6-8 months we have been working extensively with customers and partners, including T-Mobile, Softbank and IoH, to shape the portfolio, shaping real use cases and benefits for AI-RAN, why it makes sense.
“We have also been shaping the story around TCO because the moment everybody talks about the GPU, the only thing they talk about is “Oh, it’s super power hungry and it’s super expensive.” And this is not true.
“We have reached the point where we are confident to declare what we are building. We have been really working with Nvidia for quite some time now, and from an industry standpoint that’s why we have said this is an industry-first platform. We are setting the direction for the industry, and we believe this is the game changer because what is important in the years to come is getting to an AI native network.”
Nokia’s five AI-RAN pillars
First. Bringing two times spectral efficiency.
“What we are talking about, most of it requires very heavy compute: like massive multi-user MIMO pairing, deep transmitters and receivers, non-linear channel estimation, carrier aggregation. The good thing is these algorithms will keep growing and we don’t need to stop because of lack of compute. If you go with the traditional system you are limited by the CPU. There are certain things you can do – our competitors are talking about 10-15% – but we are talking about 2x performance.
“This is based on our studies and analysis: some of these are already in the testing phase in our labs so we have fairly good confidence in terms of the roadmap. Some caveats I would mention: results will depend on the load conditions, on what kind of radios are deployed. For instance, TDD Massive MIMO radios would have a higher spectral efficiency uplift than FDD standard radios because that’s the nature of the physics. And it varies from site to site. But if you look at a highly loaded site in a highly congested scenario, that is where customers operators would need the spectral efficiency compared to a rural site.
“So, what is important for our operator customers? That is exactly what our focus is – where it makes maximum sense for them, especially in dense and highly congested, high capacity scenarios, this would add significant value.
Second pillar: We are moving away from custom silicon, to merchant silicon.
“It’s important to understand because it means we are moving away from three to five year hardware cycles, waiting for silicon readiness. Now we are moving into software mode. Because it’s merchant silicon, you can just build software on top of it, so innovation happens at the pace of the software.”
TMN: If you look at Samsung, certainly Ericsson, they will say their roadmap is for their vRAN or CloudRAN software to work on any hardware or platform. That’s their end goal – that they’re not tying themselves to any particular chip platform. What’s your messaging around that?
“Well, this is exactly the same story for us. This is anyRAN software for us. Our goal is also that it’s completely open because we are not tied with any silicon players.
“If you look at our CloudRAN journey, we can work with any CPU: Intel, ARM CPUs, AMD. Our approach is also to keep all the options available for our customers. From an accelerated computing perspective, of course, we are starting with the NVIDIA GPU, but options are open.”
TMN: Can you clear up the position of your partnership with Marvell going forward? There have been some competing commentaries here.
“Marvel continues to be our one of our key partners. What I can say from our side is that we do use Marvel’s merchant silicon as well in our portfolio. They have custom silicon, they have merchant silicon. So in the baseband we will continue to use merchant silicon in our portfolio in the AirScale plug-in card.
“Every system has a CPU and GPU, when I talk about Marvel that is coming from a CPU standpoint. So where there is an off-the-shelf CPU you can have CPUs from coming from anyone. You know, it could be Marvel, it could be Nvidia, it could be any anybody, right? But the GPU is the accelerated computing, which is from NVIDIA.”
Pillar Three: We are bringing three deployment tracks.
“All of them are based on the Nvidia ARC-Pro platform, but running in three different deployment scenarios, form factors and power constraints.
“Number one offers a seamless evolution of existing AirScale system basebands with a plug-in card – and it is price and power comparable with the existing AirScale card. There is no hardware premium on it.
The AirScale AI-RAN plug-in card
“The second track is the COTS server track, which we already announced and have been trialling with a with T-Mobile US and other customers. There’s a clear plan of action with field trial readiness by Q4 this year and commercial readiness be next year. This track has maximum flexibility because because you can have multiple different types of GPUs in it. You can bring any GPUs that you want to put in there.
TMN: Nokia and Nvidia’s October 2025 release mentioned Dell PowerEdge.
“We are working with the multiple partners right now. Dell is one, we are also working with Quanta Cloud Technology. So we have multiple different options there with the GPUs. The ARC-Pro uses the RTX Pro 4500, and there are also GPUs with higher capacity for Centralised RAN scenarios. So we are in ongoing collaboration with all of these partners to make viable options for all the different scenarios. Some of them are focusing on dRAN cases, very optimised and compact. And there are several platforms with a higher capacity, which is meant for C-RAN, or even for the new physical AI use cases and other things.”
COTS server integration
“The third deployment track which is a fresh one that we will share more details about as we move along. That’s called the Accelerated AI-RAN node. It’s a standalone AirScale high capacity baseband node that would provide more AI headroom and better cell capacity than the AirScale plug-in card. Importantly, it can also interconnect with the existing AirScale system and form it as a single base station. So if you want to expand capacity you can simply add either an AirScale plug-in card or connect this all-in-one node and interconnect with a cable to the existing system.”
Accelerated AI-RAN node
Pillar Four: programmability and dApps
For more on this aspect, see this fuller article published on TMN.
“We can clearly state that we are completely O-RAN compliant with all the open fronthaul interfaces, and also O1 and O2 interfaces. And on top, what we are launching is distributed apps (dApps) which connect over a very lower layer interface to the base station called E3. We are now in the phase of partnering, or studying together with our partners, to really making sense of this.
“We need to make sure that it creates value for our customers, and that’s the journey that we take forward. We will govern it properly because this is an area where we need to be extremely tightly governed, tightly controlled. This is why it’s important that we will work with our partners and make sure that we have well-defined use cases, and make sure of how this works with the lifecycle management, with the AI workloads and dApps, with guardrails and security. All these different aspects will be taken care of during the process.
“We are working with a few of these [potential dApps] players right now, but it is too early to commend them because we are still evaluating the different aspects.
“But the good thing is, this opens up a lot of opportunities for really bringing new innovation also into the base station, not just locking it.”
Pillar FIVE Power and price
“If I talk about AirScale system, simply put there is no hardware premium because of the GPU. I want to make it very clear to the industry that is there is no hardware premium – because that was one of the concerns which we heard from the customers. We know the cost of our existing cards, it will be exactly the same. Similarly with the cloud server system, we recently submitted some RFPs together with our cloud server partners so we know the prices. It is very much comparable with the existing COTS system based on let’s say [Intel] Granite Rapids.
“Power as well is very much competitive because it’s the same power levels as for the AirScale systems. It can’t go beyond that anyway because it is constrained by the AirScale [system]. The COTS system as well has the same power [profile]. But it comes with a higher capacity so if you look at the performance per watt, it is much higher.”
