New bedfellows Nokia and Nvidia have given some fresh energy to the AI-RAN discussion. Nokia’s commitment to using Nvidia GPUs seems to be an endorsement of the AI-on-RAN proposal, but the industry is yet to be convinced. MEC did not take off as planned many moons ago, and likewise, AI compute at the edge will not be embraced by the industry.
Nvidia is buying a 2.9% stake in Nokia for $1 billion investment, and the pair are making it easier to integrate their technologies.
Nokia says it is redesigning its RAN software to be compatible with Nvidia’s compute unified device architecture (CUDA) platform. This means that Nokia will build an AI-RAN base station that will run all processing on the Nvidia GPU, called the Aerial RAN Computer Pro (ARC-Pro).
This is novel because, for the first time, Nokia is developing a product in which the most demanding Layer 1, or PHY layer, RAN processing is carried out in the Nvidia GPU only. Previously, Nokia has used a Marvell ARM-based chipset for the Layer 1 workload. Nvidia has also made capacity cards that can slot into Nokia’s existing AirScale baseband units at mobile sites.
As part of the announcement, Nokia’s major US customer T-Mobile will put the AI-RAN product into trials in 2026. “The partnership marks the beginning of the AI-native wireless era, providing the foundation to support AI-powered consumer experiences and enterprise services at the edge,” the pair said in a statement.
The alliances
Nvidia, a company with a $4 trillion market cap, has been expanding its influence into many verticals, including telecoms. It was a founding member of the AI-RAN Alliance in 2024, a group largely comprised of vendors, which promotes three principles: AI-for-RAN, AI-on-RAN, and AI-and-RAN.
AI-for-RAN uses AI to enhance the performance and efficiency of RAN activities, AI-on-RAN involves deploying AI workloads on RAN infrastructure for low-latency services, and AI-and-RAN promotes the convergence of AI and RAN systems.
The most famous and controversial of the three is AI-on-RAN, which requires MNOs to put GPUs in the RAN to run AI inference workloads at very low latencies for consumer users, also known as GPU-as-a-service (GPUaaS). See our sister unit, RAN Research’s recent report on GPUaaS.
Nokia’s commitment to the Nvidia GPU portfolio could seem like an inevitable tilt in the direction of GPUaaS. But can MNOs be coerced into adopting a new, NVIDIA-centric operating model? It looks unlikely.
Orange and Verizon have been publicly skeptical about the pitch. Verizon’s CTO Yago Tenorio said in a recent interview with LightReading that GPUs could be useful for core network software but not at the edge, and not as an additional revenue opportunity.
“AI-RAN – if that is using GPUs to do the number crunching to sell idle cycles for workloads – no, I don’t see it at all,” he said. “I think that’s a complication that I’m not sure we need, because the GPUs are very expensive.”
Likewise, Orange’s Director of Radio Innovation, Atoosa Hatefi, shared the sentiment in an interview with the same publication. “We don’t see, actually, a strong argument in favor of running RAN software on top of a GPU, particularly from a cost point of view,” said Hatefi. “We did some analysis, and we don’t see any strong driver to do centralization of RAN.”
The main problem is the low utilization rate of cell sites, which makes a costly investment in compute power hard to justify. An average cell serves around 1,000 users with demand at peak times in the morning and evening, but otherwise remains largely idle. The utilization rate is between 10-20% which makes the cost of each inferencing token, or task, much higher than if these GPUs are deployed at a central location.
If GPUs are distributed at the core of the network, they will have a much higher utilization rate, lowering overall costs markedly. By some measures, cell-site inferencing costs 10-15x more than at the core.
The other idea to consider is that the proposals for reduced latency that are central to the AI inferencing at the edge angle are undermined by the fact that the applications still need to go to the mobile core and return to the edge. This round-trip reintroduces the latency that is supposed to be reduced with edge AI compute.
Show me the money
Of course, the Nvidia-Nokia pairing is unusual in that it is a strong commitment from Nokia, which had previously seemed to be wary of the AI-RAN pitch. But no RAN vendor in the current market of slow growth will turn down $1 billion from a cash-rich investor who is happy to pay for some new research.
In its most recent earnings report, Nokia’s mobile networks business recorded an operating profit of €12 million, down from €101 million a year earlier, based on lower sales.
Nokia revamped its R&D facility in Finland this year, described as “the Home of Radio,” with a strong emphasis on AI. So, why not accept some funding to help pay the bills, even if Nokia is not completely sold on the whole AI-RAN pitch?
One angle to explain the new pairing is that traditional chip vendors are losing love for the RAN segment, and a growing role from Nvidia is inevitable.
One industry veteran, speaking to Wireless Watch off-the-record, said that poor performance by Dell and Intel in the RAN market had sent some employees over to Nvidia, but that this would not translate to a boom in GPUaaS adoption.
“I suspect that this is just a simple case of Nvidia buying a customer from Intel,” he said. “I don’t think any MNO is going to allow any random third-party applications to be run on the same platform as mission-critical RAN applications.”
Nokia and Nvidia are not purely selling the AI-on-RAN product. The new joint offering, called ARC-Pro, uses the GPU to support only RAN compute tasks.
This could make the proposition more palatable to MNOs, although not all will be convinced. Many AI-RAN tasks can be run on existing CPU systems and do not require GPU firepower.
The RAN piece has made headlines, but the more meaningful impact of this partnership may be in other networking products. As part of the partnership, Nvidia will explore the use of Nokia’s optical technologies and capabilities for its future AI infrastructure architecture.
The Nvidia investment will give it access to Infinera data center infrastructure technology, which Nokia bought last year. Nokia has been growing its business in fixed assets, especially as it has lost business with major mobile customers, such as AT&T and Verizon in the US.
The refocus on data center infrastructure follows the leadership of Nokia’s new CEO Justin Hotard who was previously responsible for Intel’s data center and AI business.
The partnership will allow Nokia to make the most of the new Infinera assets and strengthen Nvidia’s position in data center routing.
The pair said they would collaborate on AI networking solutions, including data center switching with Nokia’s SR Linux software for the Nvidia Spectrum-X Ethernet networking platform and the application of Nokia’s telemetry and fabric management platform on NVIDIA AI infrastructure.
Whatever the outcome of the AI-RAN pitch, the partnership is a huge boost for Nokia, which reached something of a nadir almost two years ago when it was excluded from the $14 billion AT&T RAN deal.