Nokia says long-term 6G is not doable without Nvidia

Held annually around this time of year in Espoo, Nokia's "Midsummer Launch" is where the Finnish vendor hopes customers will be as enchanted by its latest mobile network magic as the humans are by woodland fairies in Shakespeare's own nod to midsummer antics. This year's event marks the first since Nokia's industry-shaking deal in late 2025 to make Nvidia's graphics processing units (GPUs) the foundation of its future radio access network (RAN) products – oiled by a $1 billion investment from the giant US chipmaker. That future may just have arrived.

For many telcos, the midsummer night's dream – in fact, the one they have regardless of the season – is always about "G-upgradable" hardware, able to support a generation or even generations of mobile technology for many years. The reality has often meant cycling through all products at the same relentless pace, as new software demands a silicon swap-out. "Every three to five years, it's been new silicon," said Mark Atkinson, Nokia's head of RAN, recalling his 30 years in the sector. "Now we're going to have to think differently as an industry."

Related:Nvidia takes $1B stake in Nokia, which promises 5G and 6G overhaul

Nokia's vision, outlined during an exclusive interview with Light Reading, is that the Nvidia-based hardware products announced today and available from next year will potentially last customers deep into the 6G era. That cannot be said of the latest Nokia hardware in commercial use, based on custom silicon provided by Marvell Technology, according to Atkinson. The same would be true of the latest hardware from rival Ericsson, he believes: "It will get them into 6G, but it won't see them long into 6G."

The Nokia pivot to GPUs sets up a riveting clash between the two Nordic companies. Ericsson remains firmly attached to its own custom silicon, and its top executives think writing code for CUDA, Nvidia's software platform, could make Nokia a victim of "vendor lock-in," imprisoned by a single vendor's hardware.

For telcos interested in general-purpose processors and the greater freedom they promise, Ericsson instead offers a set of "virtual" RAN products. While based today on Intel's central processing units (CPUs), the same software can run on CPUs from other chipmakers after minimal tweaks, according to Ericsson. Only a small amount of code remains hardware-dependent, it says.

Going abstract

Yet Ericsson's virtual RAN products have a minuscule commercial footprint, and the deal with Nvidia does not make Nokia any more hardware-dependent than it already was. It merely substitutes Nvidia for Marvell in Layer 1, the physical and most demanding category of RAN software. "Today, Layer 2 plus is agnostic," said Atkinson. It can run on chips based on x86, the CPU architecture used by Intel and AMD, or chips built with blueprints from Arm, the alternative preferred by Nvidia for its Grace CPU, as well as by Marvell and various others.

Related:Ericsson and Nokia are diverging like never before on AI-RAN

Nokia's experts have also done as much as possible to develop hardware-independent Layer 1 code that can be deployed on multiple platforms without the need for an extensive rewrite. "The hardware abstraction layer that we're building helps," said Pallavi Mahajan, Nokia's chief technology officer. "There will be pieces which are going to be very Marvell-proprietary and there will be pieces which are going to be very CUDA-proprietary, but we are trying to keep it to the minimum."

Unlike the somewhat beleaguered Intel, there is no concern about the wellbeing of Nvidia – still the world's most highly valued company – among those who trade in general-purpose processors. Nokia's decision brings it into a burgeoning ecosystem of CUDA developers. As vendors and telcos try to boost performance by replacing conventional RAN algorithms with AI, that ecosystem and the parallel-processing capabilities of Nvidia's GPUs might give them an advantage over custom silicon and CPUs.

Related:Nvidia has a radical new AI-RAN plan – a 6G radio unit chip

The criticism typically leveled at Nvidia and its "AI-RAN" strategy is that GPUs are expensive power hogs, unsuited to the mobile network. But not having to replace silicon as frequently would dramatically lower costs. And Nokia is parading some very big numbers in its update. Spectral efficiency gains of 20% have already been demonstrated, it claims, envisaging a 50% capacity improvement by 2027 and a doubling or more by 2028.

Atkinson also makes the pledge that Nokia's new GPU-based products will not be more expensive than its Marvell-based products. "That's a clear message to the market – we're in an energy envelope which is competitive with custom silicon and, of course, over time we're expecting more and more GPU cycles as well," he said. "We've got something that is very performant today – it's already efficient – and we believe it will become even more efficient over time."

The claims are echoed by Mahajan. Marvell's technology works within a power envelope of 190 watts, and the Nvidia replacement will have to face the same constraints, she emphasizes. "We cannot have our customers change their power cabling, their racks and all of that stuff," she said. "But it still brings 1.5 times the capacity, two times the spectral efficiency."

End of days for custom

As promised, Nokia is launching three broad sets of products, with pilots expected toward the end of 2026 and commercial rollouts anticipated in 2027. Existing customers using an Airscale baseband chassis will be able to plug in a GPU-based card, codenamed ABIG by Nokia, to handle Layer 1 software and, potentially, the scheduling function of Layer 2, according to Mahajan. This would represent a complete alternative to what Nokia calls ABIP, its latest-generation custom silicon product from Marvell.

Telcos could alternatively deploy Nokia's GPU-based RAN on common-off-the-shelf servers, developed outside Nokia. The attraction there lies in being able to use spare capacity for other AI workloads unrelated to RAN compute – or what Nvidia calls AI and RAN. The third option appears to be a kind of Nokia appliance. "It's a compact pizza-box form factor where you have both GPU and CPU compute inside," said Atkinson. "We'll then be able to have the full benefit of GPU because we won't be limited by any power headroom anymore."

This all seems to mean the time is nearly up for Marvell as a vendor of custom silicon for Nokia's central units and distributed units – the appliances and servers responsible for baseband. "We are making now the final pivot from a custom silicon baseband strategy to a merchant silicon baseband strategy," said Atkinson. "So, ABIP will be the final custom silicon product which we will bring to market."

Marvell would still endure, however, in Nokia's advanced, Habrok-branded massive MIMO radio units, where its silicon is used for Layer 1 functions such as beamforming, a clever technique that fires mobile network signals directly at user devices. Nokia is using open interfaces defined by the O-RAN Alliance to connect the Marvell technology there to the Nvidia technology in its distributed units.

Nokia is, nevertheless, reviewing its options and not ruling out a shift to an Nvidia GPU inside future radio units. "We're currently working on the definition of the next-generation Habrok for massive MIMO, which will be in the 2028 timeframe horizon, and looking at all possible options there," said Atkinson. Whether it opts for a GPU or something else will depend on the metrics for performance and energy efficiency.

That update came weeks after Light Reading reported on Nvidia's plans to develop a GPU suitable for use in radio units. While Nvidia has neither officially confirmed nor denied those plans, it has drawn attention to the various advantages it thinks GPUs could have over other hardware platforms in this part of the 6G network. By then, the telco reception to Nokia's GPU pivot should be much clearer.

About the Author