Featured Story
T-Mobile: We'll be first with 6G, and no to SpaceX MVNO #
T-Mobile reports Q3 earnings and says it aims to be first with 6G in the US. It also continues to rule out an MVNO deal with SpaceX's Starlink.
Despite its faithfulness to Nvidia in marketing efforts, Nokia has been doing some furtive AI-RAN with Marvell Technology, its old silicon partner.
People keeping tabs on the phenomenon of the AI radio access network (AI-RAN) will have noticed Nokia parading some of its telco trialists in recent weeks while Nvidia lurks in the background like a helicopter parent. After investing $1 billion in Nokia, the huge chipmaker and its graphics processing units (GPUs) have become integral to the Finnish vendor's RAN strategy. Nokia is designing its future 5G and 6G products to work on those GPUs rather than the custom silicon in today's networks.
Trailing the partner announcements, as if it were an afterthought, was Virgin Media O2 (VMO2), one of the UK's three mobile operators. By then, Nvidia seemed to have slipped away, as busy guardians do. There was curiously no mention of it in VMO2's official statement on the AI-RAN trials, and Nvidia did not refer to VMO2 in its own list of AI activities with telcos following its recent GTC (GPU Technology Conference) shindig in Berlin.
Despite this, all previous indications were that VMO2's trial site in the London enclave of Paddington featured a GPU running Nokia's AI-RAN software. But that turns out to have been untrue. Far from using one of Nvidia's famous chips, the high-capacity site continued to rely on older custom silicon to handle the new AI-RAN features, Light Reading was subsequently informed by Nokia. (Light Reading withdrew its original story about the Paddington trial after new details emerged.)
The omission of Nvidia from the VMO2 release should have been a clue. Hardly ever does Nokia involve itself in a statement about AI-RAN without naming the chipmaker. When rival telco BT announced its own AI-RAN trials with Nokia around the same time, Nvidia showed up in the headline. Not only did Nokia omit VMO2 from its AI-RAN wrap-up; it also kept quiet about the use of custom silicon for AI-RAN, as well as the provider's identity.
That provider, however, is Marvell Technology, which Nvidia is awkwardly dislodging as part of its deal. Today, Nokia relies exclusively on Marvell's custom silicon for the resource-hungry Layer 1 software in its RAN products. VMO2's AI-RAN trials specifically leaned on a Layer 1 chip from Marvell codenamed ABIP by Nokia, Light Reading has learned.
The "ABI" in that abbreviation stands for Airscale Baseband Indoor, while the "P" seems to be merely an alphabetical differentiator from earlier generations. ABIP might be around for years, but under the current strategy it is the very last custom silicon product that Nokia plans to develop. Its successor, ABIG, is based on Nvidia's technology, with the "G" in this case standing for GPU.
What looks potentially awkward for Nokia are the details about AI-RAN achievability on existing custom silicon as well as GPUs. Since Nokia partnered with Nvidia, AI-RAN has been synonymous with the use of Nvidia's chips and CUDA software technology. Persuading customers to adopt them might be harder if already-installed custom silicon appears to deliver the same benefits.
But allowing ABIP simply to molder before GPUs have been widely rolled out would be foolish. Rival Ericsson says it has already introduced an AI successor to link adaptation, a standard RAN algorithm, in about 60,000 commercially deployed basestations built with its custom silicon. The move has improved spectral efficiency by roughly 10%, it claims. Without a competing software product that it can offer on installed silicon, Nokia would risk a loss of market share.
None of this necessarily invalidates the case for shifting to GPUs. Once they have deployed ABIG, customers will be able to benefit from years of AI-RAN software upgrades that continue to boost spectral efficiency, while custom silicon must be replaced at more frequent intervals, Nokia has argued. After earlier market share losses, and apparent concern within Marvell about its own return on investment in the shrinking RAN sector, Nokia may have had limited silicon options.
The trouble, to reiterate, is that if custom silicon is shown to remain competitive for AI-RAN, operators could be dissuaded from spending money on GPUs. And that is exactly what the VMO2 trial seemed to indicate.
In one of the UK's busiest neighborhoods, it focused on the radio conditions that exist between the antenna at the mast and the smartphones used by customers in the Paddington area. "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, at a recent press briefing.
Traditional algorithms work according to predefined sets of parameters about those radio conditions. AI essentially learns on the job, like a precocious apprentice, and adapts quickly to real-time factors. "You optimize according to what you see, and it trains itself on the environment the cell site is in," explained Joyce. "So, the basestation tunes its algorithms to that environment."
The result seen by VMO2 in Paddington was a 20% improvement in spectral efficiency, meaning the site was able to carry more data without needing additional capacity or spectrum. The figure of 20% is also what Nokia recently said it had been able to achieve in spectral efficiency gains by using GPUs for AI-RAN.
Joyce has also acknowledged the competitiveness of AI-RAN products from Ericsson, which is sticking with custom silicon and whose CEO, Per Narvinger, previously argued that a "native implementation" with Nvidia would create lock-in. "You don't need a GPU to do this," emphasized Joyce at the press briefing. "I think, to be fair, that Ericsson silicon is very good as well."
One man who would not be surprised to see AI-RAN done on Marvell's chips is Tommi Uitto, the former president of Nokia's mobile networks business group. "Somebody was smart enough a few years ago to embed certain AI and ML [machine learning] capability in the chip, which we haven't used yet," he told Light Reading in early 2025. "Making AI-RAN with fairly expensive GPUs, only to have network performance improvements, may not make techno-economic sense," he said at the time.
The key question, perhaps, is whether the normal cadence of custom silicon releases can maintain pace with future software improvements possible on a GPU platform that does not have to be so regularly changed. In the summer, when Nokia talked about that 20% improvement in spectral efficiency, it reckoned 50% would be achievable next year and a doubling by 2028.
But a white paper from Chetan Sharma Consulting, endorsed by Nokia, was far more conservative, even expressing various doubts. "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, the consulting firm's boss.
What all the various lab tests and field trials of AI-RAN fail to indicate is the baseline against which those spectral efficiency gains are measured. But RAN algorithms, according to experts, are rather like Stradivarius violins – shaped and refined over decades to a point where they become almost unimprovable. An AI tasked with making a superior instrument to the famous Italian luthier's would seem unlikely to enjoy much success.
Regardless of all that, trialing ABIP for AI-RAN makes VMO2 look even less likely to embark on a massive rollout of GPUs. It has already spent about £1.4 billion (US$1.9 billion) on mobile network upgrades in the last two years, installing new radios from Nokia, its vendor for the southern part of the UK, and Ericsson, which it uses in the north. The £700 million ($940 million) tranche of spending announced early this year does not include any plans for a GPU RAN deployment, Joyce confirmed to Light Reading.
He sounds dubious, too, about the attractions of a GPU-based centralized RAN, which would consolidate baseband resources in other network facilities while leaving only radios at the mast. "If you've got high-capacity fiber, you can move that GPU to a centralized location," to control numerous sites, he said. "That's beneficial because you could then do what we call coordination of scheduling between multiple technologies and improve that even further."
Joyce's concern is about the fiber those sites would need. The Paddington site is connected to 10Gbit/s fiber, providing more than enough capacity for local needs. Centralizing the baseband would require VMO2 to connect that facility back to the radio site with a technology called CPRI (common public radio interface), or its eCPRI (evolved) successor. "Each one of those is 25 Gbit/s," said Joyce. "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."
Almost inevitably, that would mean adding fiber or paying more for transmission, explaining Joyce's reservations about investing in a centralized RAN. "My view is still that I think the distributed RAN is the way forward, and, in that case, if you're going to get this, you're going to have to put additional baseband GPU resources at the site," he said.
For Nokia, the risk is of customers continuing to resist an upgrade that would demand the rollout of new hardware, whatever it may cost. The availability of AI-RAN features on Nokia's custom silicon products could act as a further disincentive. But there would be danger for Ericsson, conversely, if today's custom silicon were unable to accommodate emerging AI features that powerful GPUs can already support. Customers now benefiting from link adaptation improvements on hardware they have used for years might discover it needs to be replaced sooner than expected for more advanced AI. Telco anxiety about either vendor's technology roadmap could be an opportunity for the other to grow market share. But with so few RAN vendors in the market, operators desperately need both Nordic companies to succeed.
Read more about: