- Nokia is betting big on AI-RAN with Nvidia, promising major spectral gains by 2027 and 2028
- Analysts say the claims are ambitious, but real-world proof is still missing
- Skeptics question whether putting GPUs in base stations makes sense, especially given power and cost concerns
A significant part of the industry remains skeptical about GPUs in base stations, but Nokia is going all in, unveiling its commercial AI-RAN platform and billing its new GPU-powered AirScale capacity plug-in unit as the most efficient path forward.
Nokia is making some big claims, too, about the AI-RAN platform, which is built on its AI-native anyRAN software and Nvidia’s Aerial AI-RAN technology. The vendor says 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.
“AI-RAN is the biggest innovation in radio in decades,” Nokia President and CEO Justin Hotard declared. “AI-RAN makes the network intelligent, extends AI into the physical world, and allows telcos to get more from their existing infrastructure, including a software upgrade path to 6G.”
Nvidia CEO Jensen Huang said telecom is entering the AI era, and the radio network is the next AI infrastructure. “Together with Nokia, we are bringing Nvidia CUDA and AI into the baseband, transforming RAN into a planet-scale AI computer. This is a generational shift for operators,” he said in a statement.
Those are some bold promises. “Biggest innovation in radio in decades”? And “transforming RAN into a planet-scale AI computer”?
Fierce asked analysts if Nokia and Nvidia’s bravado is really all that.
Who’s putting GPUs in base stations?
One of Nokia’s harshest critics, EJL Wireless Research President Earl Lum said this is the No. 1 question everyone should be asking: Are Chinese vendors putting GPUs in base stations? Because Huawei and ZTE are not, he said.
“The Chinese don't think that's a good way to go, and they have more base stations than anyone on the planet. So, do you think they're concerned about power consumption? Absolutely,” he said. “They’re not idiots.”
For that matter, Ericsson and Samsung are leaving their options open. Ericsson CEO Börje Ekholm told investors this week that Ericsson’s RAN stack can run on x86, GPU or Ericsson’s purpose-built silicon and it remains agnostic from a hardware point of view.
“No one else is doing this other than Nokia, so you have to ask: Are they that much smarter than everyone else? No,” he said.
Nokia’s Midsummer update
Mike Thelander, president and founder of Signals Research Group, was among the analysts who were briefed on Nokia’s latest AI-RAN announcement. One of his reactions was “why now?” as Nokia already announced its partnership with Nvidia last year and said at Mobile World Congress 2026 in Barcelona that it would deliver AI-RAN radio units before the end of this year.
The timing could be explained by way of Nokia’s annual Midsummer update landing the same week; it’s the time of year when Nokia historically releases updates. In addition, “I think they’re kind of going from this conceptual ‘this is what we’re going to do’ to ‘we’re actually doing it,’” he told Fierce.
To that end, Nokia reiterated that AI-RAN solutions will enter pilot deployments at the end of this year and be commercially available in 2027.
In the U.S., Nokia Cloud AI-RAN is already in 5G field trials with T-Mobile, its lead trial partner. Nokia is no longer a main wireless infrastructure supplier at AT&T or Verizon.
Outside the U.S., the Finnish vendor cites collaborations with operators including SoftBank, Deutsche Telekom, Telia, Orange, BT and Indosat Ooredoo Hutchison. However, not all deployment options have been publicly confirmed as tested by all operators, a spokesperson told Fierce.
During the briefing, Thelander asked Nokia executives how widely they expect operators to use a GPU-based cell site.
“The answer seemed to suggest a high degree of ubiquity since Nokia described it as a natural evolution, plus there is its argument about total cost of ownership (TCO). Not all GPUs cost an ARM and a leg,” he wrote on LinkedIn. “That being said, Nokia also indicated the most benefit will be realized in dense, highly loaded cells so I’m a bit skeptical that the T-Mobile cell site serving SRG headquarters (Nokia infrastructure) will be on the upgrade list. Then again, T-Mobile knows where I live.”
Multiple road maps to upgrade
Any claims about performance gains should always be taken with a grain of salt, said Recon Analytics analyst Daryl Schoolar. But he added that he’s confident Nokia and Nvidia will deliver spectral gains; it’s just a matter of how much.
What’s encouraging is Nokia’s decision to provide multiple road maps for operators to upgrade, including an expansion card for existing AirScale customers.
“I think that’s important because not all operators are going upgrade the same and even within an operator, they’re not going to upgrade the same,” he told Fierce.
AI-RAN is a little bit like Massive MIMO, he said. “An operator may deploy 64x64 Massive MIMO in some areas and maybe 32 in other areas and in other areas they may stick with 16 or 4,” he said. “It all depends on what the environment is, how much capacity they need, what is the demand of the network in those areas,” he said. “I believe operators would take the same approach to AI RAN. I don’t think they’re going to deploy it at every site.”
More specifics, please
Commenting on LinkedIn, Mobile Experts analyst Joe Madden said Nokia’s going to need to provide a lot more than a press release to get the analyst world “hot and bothered.”
“Nokia needs to be much more specific about their 2X improvement claim, with field data on a highly optimized network … I will concede that their idea of focusing on MU-MIMO pairing may yield some big gains. Does that really require GPUs? Or can tensor pools, DSPs, and CPUs achieve the same thing?” he said.
“The game is not over. Nokia, to win you need to prove it,” Madden concluded.
Read Fierce Network's related coverage of AI-RAN
NTIA announces funding opp for AI-native wireless network innovation