Multiple operators are testing Nokia's GPU-based radio access network products this quarter, ahead of commercial availability early next year — the biggest bet any major vendor has placed on moving RAN compute from custom silicon to Nvidia GPUs.

The stakes are high. Nokia has stopped developing new custom silicon with partner Marvell Technology and teamed up with Nvidia to design RAN products built on graphics processing units. It will continue to support its existing 4G and 5G range for years, but customers will eventually need new hardware to get new features.

Nokia argues its GPU-led strategy will support a doubling in spectral efficiency by 2028. In a RAN market not expected to grow in the next few years, such gains could justify the hardware outlay and help Nokia recover 5G ground lost to Ericsson and Samsung. But no operator has publicly commented on the trials so far, and skeptics are numerous.

Why is Nokia's GPU bet risky?

Nvidia's chips have never been used for RAN compute. If telcos balk, Nokia has no fallback. "If you don't have a companion strategy that's based on classic ASICs, you really have created jeopardy, I think," said Marcus Weldon, a former Nokia CTO who also presided over its Bell Labs research unit.

Ericsson customers face their own upgrade path — operators on its purpose-built 5G products may need newer application-specific integrated circuits for 5G-Advanced and 6G software — but Nokia's all-in GPU strategy carries more risk.

The dependency cuts both ways. Nokia's software is bound by CUDA, the Nvidia software platform often described as a moat, and it cannot port GPU-developed software to alternative silicon. "If the GPU thing doesn't catch on, they don't have an alternative," said Vincent Loncke, founder of Kenyi Technologies and a former Qualcomm developer.

What is Nokia promising operators?

Nokia has been advertising a slimline GPU that slots into an existing Airscale chassis originally designed for Marvell silicon. Executives say it must work within the same power constraints, sell at the same price, and support software upgrades and new AI algorithms for years — unlike any ASIC.

CEO Justin Hotard framed the value proposition on a call about first-half financial performance: "What we anticipate with AI-RAN is that it gives customers an opportunity to reduce the ongoing hardware investment, or to elongate that. Then they can get the benefit of ongoing enhancements once they deploy AI-RAN software."

Analyst backing exists. "ASIC refresh cycles typically run on three-to-five-year timescales," said Chetan Sharma, CEO of Chetan Sharma Consulting, in a white paper on the topic. "While ASIC-based platforms do support algorithm refinement within a given generation, the scope is ultimately bounded by fixed hardware constraints."

Sharma credits the Hardware Abstraction Layer (HAL) with insulating RAN software from silicon-level changes. His paper was endorsed by Nokia, though Sharma is a widely respected authority who has advised many of the world's biggest telecom players.

Do the numbers hold up?

The caveats pile up quickly:

• Custom silicon remains more energy efficient, Sharma concedes. A GPU RAN's total energy cost must not exceed a traditional RAN plus a separate AI inference server.
• Nvidia cycles GPU generations fast: Hopper gave way to Blackwell, now Rubin, with Feynman ahead. Blackwell brought changes that required software adaptation, Sharma notes.
• Sharma's final pages talk not of a 2x gain but of 40% to 50% improvement under favorable conditions at target deployment sites.

Sharma's own most telling sentence concerns the baseline: a site already achieving near-theoretical spectral efficiency under classical processing has less headroom for AI-driven improvement than a site operating well below its theoretical ceiling due to interference, channel estimation errors, or suboptimal scheduling.

Spectral efficiency, unlike transistor counts, faces a scientific limit in the Shannon-Hartley theorem, which caps how much information fits into a given amount of bandwidth.

What do technical experts say?

Weldon, who ran a GPU RAN project in Bell Labs, puts the realistic gains at Ericsson levels. "Those were the numbers – 20%, 30%. You can get 2x, but it is more about continuous optimization of some very specific scenarios," he said. "The 2x is real, but I would argue niche."

Nokia CTO Pallavi Mahajan has highlighted RKHS — "reproducing kernel Hilbert space" — as an advanced signal-processing method GPUs could unlock. Loncke is unconvinced: in all but "corner" cases, noise can be suppressed more efficiently with the fast Fourier transform (FFT), the approach Qualcomm reportedly chose after examining RKHS. "If you really want to do RKHS, you just dimension it to do that. It's just a matter of scaling up the circuitry to support it," he said.

The transmission time interval (TTI) imposes a hard ceiling on algorithm complexity regardless. "You have 500 microseconds, so you have to be responsive," said Loncke. A large model cannot run inside that budget; a distilled lightweight model at the edge probably would not need GPU muscle. A technical expert at a major vendor, asking not to be named, agreed TTI budgets must be met but noted some RAN functions — aspects of scheduling, channel quality adjustments, multi-user MIMO — can take 5 or 10 milliseconds or longer, where larger GPU-based models might help.

Operators themselves are split. Verizon CTO Yago Tenorio sees value in GPUs but not in the RAN. "You will not be running the RAN workloads on those GPUs," he said during a discussion last year. "You'll be using them for what they are really good at, which is to run AI inference."

Sharma's paper even hedges on the capex story: if HAL portability holds cleanly across two GPU generations, hardware investment amortizes over a seven-year window without forced replacement; if each transition requires significant operator re-engagement, the ASIC alternative's lower operational cadence becomes more attractive.

What happens next?

The trial feedback will stay private until Nokia's AI-RAN products reach the market early next year, when operators must decide whether a 2x — or more likely 40–50% — spectral efficiency gain justifies swapping out silicon that many had hoped to run well into the 6G era.