Nvidia preps fresh AI-RAN pitch

• AI-RAN was a hot topic at MWC25
• But there are still plenty of sceptics due to issues related to cost and energy efficiency, among others
• As one of the founders of the AI-RAN Alliance, Nvidia is keen to dispel such concerns
• The AI chip giant’s head of telecom hints that solutions to support AI-RAN could be unveiled during Nvidia’s GTC event in San Jose this week

As most in the telecom sector are already aware, the recent MWC25 event in Barcelona was awash with AI pitches and applications, with developments in AI-RAN one of the highest profile topics in terms of AI’s impact on network infrastructure architectures. In the wake of MWC25, and in the face of some scepticism about the real-world applicability of AI-RAN for the mobile operator community, Nvidia looks set to unveil telco-specific solutions with partners during its GTC event in San Jose this week, in an effort to attract more supporters from the telco sector to the AI-RAN camp.

Before delving into the details, let’s just qualify what we mean by AI-RAN in the context of this article – we’re talking about the concept being put forward by the founders of the AI-RAN Alliance, which launched a year ago with three main focus areas:

• AI for RAN – the use of AI tools to improve the performance and efficiency of radio access networks (RANs)
• AI and RAN – the integration of AI and RAN processes on the same underlying infrastructure, so that the resources are constantly being used even if there is no mobile customer activity on the network, though with RAN processes prioritised when there is network activity
• AI on RAN – the development and deployment of AI-enabled applications at the edge of the network that can be delivered over 5G connections 

As the leading global supplier of graphics processing units (GPUs) that support all manner of AI workloads, Nvidia was one of the alliance’s founders, along with Amazon Web Services (AWS), Arm, DeepSig, Ericsson, Microsoft, Nokia, Northeastern University, Samsung Electronics, and two network operators in the form of SoftBank and T-Mobile US. Over the past year, the alliance’s membership has grown to 75, though only about 10% of the membership comes from the network operator community. 

Just ahead of MWC25, Nokia banged the AI-RAN drum loudly, using its pre-show media and analyst conference to unveil advances with Japanese operator KDDI as well as with three of its fellow alliance founders, T-Mobile US, SoftBank and Nvidia  – see Nokia makes its AI-RAN play at #MWC25.

Then during MWC25, SoftBank announced a slew of AI-RAN architecture advances, including the implementation of various radio access network functions with Fujitsu and Nvidia as part of the AITRAS AI-RAN solution it is developing in partnership with Nvidia (see this press release).

In addition, Indonesian operator Indosat Ooredoo Hutchison (IOH) announced it is working with Nokia and Nvidia to deploy what it calls “a unified accelerated computing infrastructure for hosting both AI and RAN workloads”. The three companies have agreed to develop, test and deploy an AI-RAN solution with an initial focus on managing AI inference workloads using Nvidia’s AI Aerial system and then, later on, to integrate radio access network workloads on the same platform.  

And in the wake of the show, Samsung announced it is working hand in hand with Nvidia to show how the South Korean vendor’s radio access network technology can be integrated with servers running Nvidia processors.

So lots of big names are involved in AI-RAN developments and there’s no doubt that the concept of using edge IT resources to advance telco AI business opportunities and improve network operations is a positive and appealing one. 

Yet there is plenty of scepticism in the industry, including among mobile network operator executives with whom TelecomTV discussed the topic, that the AI-RAN Alliance’s vision is suitable for most telcos, at least in the near term, mainly because of the perceived cost of deploying compute stacks based on GPUs in the radio access network and the power consumption that such deployments would require at a time when network operators are doing everything they can to improve their energy efficiency and reduce their operating costs. In addition, there’s widespread uncertainty as to how most mobile operators would build AI compute stacks into their networks – even T-Mobile US’s president of technology, Ulf Ewaldsson, admitted during the Nokia pre-MWC briefing that there was still work to be done on the optimum architecture for AI-RAN. 

With all of this in mind, what does Nvidia make of that scepticism? The concerns are not new but continue to hang over the AI-RAN concept. 

Nvidia’s head of telecom, Ronnie Vasishta, is, as you might expect, holding firm on the benefits to telcos of an AI-RAN approach, claiming that the power and cost concerns don’t often crop up in Nvidia’s AI-RAN conversations and that a lot of progress is being made on different architectural approaches. 

In a briefing with TelecomTV held at the end of MWC25 as the show floor was being dismantled, Vasishta stated that conversations about cost and energy efficiency are becoming a thing of the past and that “AI-RAN power consumption doesn’t come up that much. You’ll understand more [soon],” he noted, hinting at announcements set to be made during the GTC event. 

But there has always been the pitch from Nvidia and its partners that an AI-RAN approach is, in the grand scheme of things, efficient and opens the door for revenue-generating as well as cost-saving opportunities (see our coverage of the AITRAS pitch). 

Vasishta noted that with a centralised RAN approach, “because you’re pooling resources across an accelerated compute infrastructure, the performance per watt actually turns out to be very beneficial. When you look at a distributed RAN, [with resources deployed] all the way out to a cell site, then… an AI server that’s being utilised low on RAN is more costly than, say, a purpose-built baseband. But if you’re utilising fully the capability of the accelerated compute AI server, you’ll see that it turns out to be beneficial in TCO [total cost of ownership] performance and power consumption because it’s so capable of processing. 

“And as we move to the edge, there are different sizes of GPUs that can be used and I think [in the near future] you’re going to see how different sizes of GPUs can be used, all the way out to the distributed RAN, from some of the other suppliers as well… there will be announcements from other vendors about how they’re using GPUs and accelerated compute in a distributed RAN,” added the Nvidia man, hinting at GTC-related news to come.

“The idea that accelerated compute and GPUs are power hungry and costly depends on how much you are consuming… which GPU and accelerated compute you’re using, whether it’s large AI servers or something smaller and more power efficient,” he added. 

As for the architecture options, “we’re still going through some of that,” admitted Vasishta, as well as figuring out exactly how AI server resources would be best used. “How do you use AI on that server? And what kind of AI applications do you want to run? There’s kind of an AI track, and there’s a RAN track.” He noted that the ‘AI for RAN’ conversations have come to the fore. “The benefits are starting to become more obvious… Everyone's talking about which AI applications are going to be run – AI actually plays a very big role in improving the efficiency of the RAN [related to] spectral efficiency and power efficiency. A lot more research is starting to come to fruition.”

But couldn’t those applications run on general purpose non-GPU servers?

“When you look at the efficiency of implementing AI, even if you have embedded vector engines [on general purpose servers) they’re still not as efficient as running in the GPU. The other thing is that [Nvidia has] a huge user base of CUDA libraries that make that [GPU] implementation way more efficient and supported by an ecosystem. And then there’s the concept of closed-loop [automation] and learning. [In terms of] RAN automation… if you have a digital twin where you can train your algorithms and simulate your algorithms, then you can deploy that into the infrastructure directly [and] you can have visibility all the way down into Layer 1, Layer 2, Layer 3, and you can apply those to meet certain service-level agreements or certain scene-specific requirements, like 200,000 people in a business district by day. Maybe there’s a concert and you can optimise your network. You don’t want to have discrete little AI engines that don’t have visibility all the way up into the infrastructure and orchestration layer, and that’s what a lot of the companies that are in AI-RAN are working on.”

AI-RAN looks set to be a hot topic for the rest of 2025 (and beyond) and, if Vasishta’s hints are anything to go by, could be making some headlines again this week during Nvidia’s GTC event. 

- Ray Le Maistre, Editorial Director, TelecomTV