Wireless

Nokia’s AI-RAN pitch gets reality check in new study

By Monica Alleven Sep 8, 2026 12:01pm

Nokia Nvidia T-Mobile Chetan Sharma

Earlier this summer, Nokia said its AI RAN work with Nvidia put it on track to deliver 50% spectral gains by 2027 and more than 100% by 2028
Though impressive, the claims were met with a high degree of skepticism from some analysts who said they’ll believe it when they see it
A new study by Chetan Sharma Consulting breaks down the claims and concludes that putting GPUs at the right sites – rather than everywhere – is the best response

Nokia’s claims earlier this summer about dramatic gains in spectral efficiencies thanks to its AI-RAN platform, developed with Nvidia, were met with great fanfare and a healthy dose of skepticism.

To recap: Thanks to its AI-native RAN platform, Nokia said it’s on track to deliver 50% spectral gains by 2027 and more than 100% by 2028, helping telecom providers carry significantly more traffic in dense cells while reducing the cost per bit.

That, understandably, created a lot of buzz in the industry, with some analysts lauding Nokia for its leadership in AI RAN and others saying they’ll believe it when they see it.

Today, a new study released by Chetan Sharma Consulting – commissioned by Nokia – delves into some of Nokia’s claims and answers the industry’s burning question: Where does it make sense to put GPUs in wireless networks?

Both issues – Nokia’s claims and the GPU question – include some caveats. For example, the 50% uplift in spectral efficiency is a “best-case envelope, not a uniform network gain,” the report said.

That 50% mark is achieved “only where several conditions hold at once: a TDD Massive-MIMO deployment with high-order arrays and reciprocity-based beamforming, a heavily loaded sector with many simultaneously active and spatially separable users, sustained high-volume traffic in both directions, a good-to-moderate-SINR [signal-to-interference-plus-noise ratio] interface-limited regime, and mobility low enough to preserve channel coherence,” the study says.

Whew! That’s a lot of conditions. Suffice it to say: In dense, urban environments, the full force of the GPU is felt. In rural and less populated regions? Not so much.

The study points out that network traffic is not uniform – about 20% of sites carry 60%-70% of total traffic and those sites are overwhelmingly where the full AI-RAN algorithm portfolio can deliver maximum value.

So, where do GPUs belong?

“The correct answer is GPU-native AI-RAN at the sites where the economics justify it, ASIC-enhanced platforms at the rest and upgrade optionality built into every platform that sits near the quadrant boundary,” the study says. “The operators who map their network’s heterogeneous economics most precisely, and deploy accordingly, will outperform both the all-GPU and all-ASIC alternatives.”

Nokia RAN CTO: Deploy in stages

Nokia CTO of RAN and Core Udayan Mukherjee acknowledged that they’re not talking about these great spectral efficiency gains across the entire network.

Places that are densely populated with a mix of mobility and fixed wireless usage will be the areas where the spectral efficiency gains are felt the most. “It is not across the network,” he told Fierce.

Importantly, they’re not telling operators to use this technology in rural areas. “It’s not going to give that kind of benefit,” he said. “To be honest, that’s why we are doing it in stages. We are not telling operators that you have to replace everything… Start in those highly dense and highly profitable areas” and go from there.

By the way, there’s a good bet Mukherjee will be sharing more about the findings later this week. He’s scheduled to be on a panel at Chetan Sharma’s Mobile Future Forward (MFF) event in Seattle on Thursday. Fierce will be there too, so stay tuned.

Nokia, Nvidia and T-Mobile

T-Mobile was an early partner in the Nvidia/Nokia endeavor and it’s been testing AI-RAN at the T-Mobile AI-RAN Innovation Center in Bellevue, Washington. That’s where they’ve successfully tested GPU-accelerated AI-RAN workloads in an over-the-air lab environment, running Nokia’s AirScale Massive MIMO radio on the 3.7 GHz (n77) band on a single Nvidia Grace Hopper 200 server.

What they’re evaluating runs the gamut: performance capabilities, the ability to run telco and AI workloads simultaneously on a single platform and total cost of ownership, with earlier lab work also covering real-time machine learning and spectrum efficiency methods, according to the study.

“The AI workloads being tested are a video captioning AI application and physical AI scenarios with computer vision, robotics and drones for industrial/urban use cases,” the study notes.

Looking ahead, “Nokia’s product timeline calls for field trials in Q4 2026 featuring 5G software with TDD Massive MIMO, multi-user MIMO and higher-order QAM, with commercial 5G offerings slated for Q4 2027,” Sharma wrote.

Mukherjee said some of this stuff is already running in Nokia’s labs in Dallas and Finland. Later this year, they’ll be doing field trials with T-Mobile in three different sites in the U.S.: Seattle, Dallas and possibly Chicago.

Nokia CTO: Monetization will come

Mukherjee acknowledged that monetization is the tough part. Recall, there was a lot of skepticism in 5G when the industry talked about mobile edge compute and guess what? “Nothing really happened,” he said.

But when you talk about spectral efficiency, that’s another story. “If I can improve spectral efficiency … that is directly equivalent to money being saved,” he said. “If I can free up some of the spectrum for more users, more fixed wireless users, more broadband users … it's a form of monetization.”

In addition to giving the operators more spectrum to serve home broadband users via fixed wireless access (FWA), the AI-RAN framework provides the means to serve robots, cameras and other physical AI use cases that are heavy users of uplink data.

“Freeing up the spectrum that you have to actually absorb that additional traffic is going to be very useful,” he said.

Early AI-RAN bird gets the worm

Of course, there’s a sense of urgency in all of this, as one might expect. Nokia is, after all, trying to sell operators on its strategy and time is of the essence.

The white paper outlines three waves, beginning with the Wave 1 portfolio of algorithms now in field trials. Wave 2 encompasses AI-RAN exclusive algorithms, many of which are theoretically well-understood but not deployable on ASIC-based baseband. Wave 3 represents the AI-native 6G destination of it all.

Taken together, they represent opportunities for operators that get on board early.

Operators that move in the 2026-2029 timeframe will gain operational experience, capture early non-RAN AI revenue and arrive at the 6G transition with two to three years of accumulated training data and a software update path to AI-native features, the report states.

“Operators who defer risk missing that transition window and may encounter a hardware upgrade cycle at a less favorable time,” the report concludes.

One more thing: T-Mobile CTO awaits results

During a recent interview, Fierce asked T-Mobile Chief Technology Officer John Saw about its AI-RAN work with vendors and the prospect of spectral efficiencies. Earlier this year, T-Mobile conducted AI-RAN trials with Ericsson where they reportedly achieved close to 10% increase in spectral efficiency and up to 15% boost in downlink throughput compared to legacy methods.

“I think one of the biggest benefits of having AI natively built in RAN is going to be things like special efficiency. That's something that we really look forward to because it allows us to support our customers better,” he said. “It’s substantial.”

As for the spectral efficiencies that Nokia is talking about, “the proof is going to be in the pudding,” he said, noting that T-Mobile will be starting field trials with Nokia very soon. “We’ll see whether we see that.”