Nokia's VMO2 AI-RAN Trial Ran on Marvell Silicon, Not Nvidia GPUs

Virgin Media O2's AI-RAN trial site in Paddington, London, delivered a 20% improvement in spectral efficiency on Marvell custom silicon — not on an Nvidia GPU, as previously assumed. Nokia confirmed the correction to Light Reading, which withdrew its original story after the new details emerged.

The revelation matters because Nokia has tied its future RAN strategy to Nvidia. After Nvidia invested $1 billion in Nokia, the Finnish vendor began designing its 5G and 6G products around Nvidia GPUs rather than the custom silicon in today's networks. Nearly every Nokia AI-RAN announcement names Nvidia. The VMO2 statement did not, and Nvidia omitted VMO2 from its own list of telco AI activities after its recent GTC conference in Berlin.

Whose chip actually powered the trial?

The chip is a Marvell Technology Layer 1 product codenamed ABIP by Nokia. The "ABI" stands for Airscale Baseband Indoor; the "P" distinguishes it from earlier generations. Nokia relies exclusively on Marvell custom silicon for Layer 1 processing in its current RAN products. ABIP is the last custom silicon product Nokia plans to develop. Its successor, ABIG, is GPU-based — the "G" stands for GPU.

The trial focused on radio conditions between the mast antenna and customer devices. "What we've been working with them on is using AI to better estimate the channel quality," said Rob Joyce, VMO2's director of mobile access engineering. "You optimize according to what you see, and it trains itself on the environment the cell site is in. So, the basestation tunes its algorithms to that environment."

The 20% gain matches what Nokia recently claimed for AI-RAN running on GPUs.

Why this complicates Nokia's GPU pitch

If installed custom silicon delivers the same AI-RAN benefits, operators may balk at buying new GPU hardware. Rival Ericsson says it has already introduced an AI successor to link adaptation, a standard RAN algorithm, in about 60,000 commercially deployed basestations using its own custom silicon, improving spectral efficiency by roughly 10%. Without a competing software product for installed silicon, Nokia would risk market share loss.

Joyce acknowledged Ericsson's competitiveness directly. "You don't need a GPU to do this," he said. "I think, to be fair, that Ericsson silicon is very good as well." Ericsson CEO Per Narvinger has previously argued a "native implementation" with Nvidia would create lock-in.

Tommi Uitto, former president of Nokia's mobile networks business group, raised similar doubts in early 2025. "Somebody was smart enough a few years ago to embed certain AI and ML capability in the chip, which we haven't used yet," he told Light Reading. "Making AI-RAN with fairly expensive GPUs, only to have network performance improvements, may not make techno-economic sense."

Nokia's counterargument is upgrade longevity. Once customers deploy ABIG, they can receive years of AI-RAN software upgrades that continue boosting spectral efficiency, while custom silicon needs more frequent replacement. This summer, Nokia claimed 50% gains would be achievable next year and a doubling by 2028.

A white paper from Chetan Sharma Consulting, endorsed by Nokia, was more cautious. "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," wrote Chetan Sharma. Field trials also rarely disclose the baseline against which gains are measured.

What it means for VMO2's network plans

The Paddington trial makes a large GPU rollout at VMO2 look unlikely. The operator has spent about £1.4 billion ($1.9 billion) on mobile network upgrades over two years — roughly £700 million ($940 million) per year — installing new Nokia radios in the south of the UK and Ericsson radios in the north. The £700 million tranche announced early this year includes no GPU RAN deployment, Joyce confirmed.

He is also skeptical of GPU-based centralized RAN. The Paddington site runs on 10Gbit/s fiber. Centralizing baseband would need CPRI or eCPRI connections of 25Gbit/s each. "You may end up then having to run three CPRI connections, so 75 Gbit/s to the tower, as opposed to the 10 Gbit/s," said Joyce. "My view is still that I think the distributed RAN is the way forward."

For Nokia, the risk is that AI-RAN on existing silicon further discourages hardware upgrades. For Ericsson, the danger is that today's custom silicon cannot host emerging AI features GPUs already support. With so few RAN vendors left, operators need both Nordic companies to succeed — and the ABIP trial timeline suggests the custom-versus-GPU question will stay unresolved at least until ABIG-based products reach the market.