Despite its marketing loyalty to Nvidia, Nokia has been quietly advancing AI-RAN with its longtime chip partner, Marvell Technology.
Those following AI Radio Access Networks (AI-RAN) have noticed that Nokia has been showcasing several telecom pilot customers in recent weeks, while Nvidia has been hovering in the background like an overprotective parent. After investing $1 billion in Nokia, this major chipmaker and its graphics processing units (GPUs) have become central to Nokia’s RAN strategy. Nokia is designing its future 5G and 6G products to run on these GPUs rather than the specialized chips used in today’s networks.
Following these partnership announcements, Virgin Media O2 (VMO2), one of the UK’s three major mobile operators, appears to have been mentioned only in passing. By then, Nvidia had seemingly slipped away like a busy guardian. VMO2’s official statement regarding its AI-RAN trial notably omitted any reference to Nvidia, and Nvidia’s recent list of telecom AI initiatives, released after its GTC (GPU Technology Conference) in Berlin, did not mention VMO2 either.
Nevertheless, all previous indications suggested that the trial site in London’s Paddington, used by VMO2, relied on Nvidia GPUs running Nokia’s AI-RAN software. However, Nokia subsequently informed Light Reading that this was not the case. Far from using Nvidia’s well-known chips, this high-capacity site continues to depend on older, custom-designed chips to handle the new AI-RAN functions. (Following the emergence of these new details, Light Reading retracted its initial report on the Paddington trial.)
The VMO2 statement did not mention Nvidia, which should have been a clue. Nokia almost never omits naming this chipmaker in statements involving AI-RAN. Around the same time, competitor BT announced its AI-RAN trial with Nokia, and Nvidia was featured in the headline. Not only did Nokia fail to mention VMO2 in its AI-RAN summary, but it also remained silent on the custom chips used for AI-RAN and their suppliers.
However, this supplier is actually Marvell Technology, and NVIDIA is awkwardly pushing it out through its partnership. Today, Nokia entirely relies on Marvell’s custom chips for its resource-intensive Layer 1 software in RAN products. Light Reading learned that Nokia’s VMO2 AI-RAN trial specifically uses Marvell’s Layer 1 chip, internally codenamed ABIP by Nokia.
In the abbreviation "ABI," "ABI" stands for Airscale Baseband Indoor, and "P" appears to simply serve as a letter distinction from earlier generations. ABIP may remain in use for many years, but under the current strategy, it will be Nokia’s final custom-designed chip. Its successor, ABIG, is based on Nvidia technology, where the "G" stands for GPU.
For Nokia, the potential awkwardness lies in the fact that AI-RAN can be implemented on both existing custom chips and GPUs. Since Nokia’s partnership with Nvidia, AI-RAN has nearly become synonymous with using Nvidia chips and CUDA software technology. If existing deployed custom chips can deliver the same benefits, it may be harder to convince customers to adopt GPUs.
However, it would also be unwise to let ABIP sit idle until obsolescence before GPUs are widely deployed. Competitor Ericsson has stated that it has already implemented an AI-based successor to the standard RAN algorithm for link adaptation on its custom chips in approximately 60,000 commercially deployed base stations. The company claims this move has improved spectral efficiency by about 10%. If Nokia cannot offer a competitive software solution on its installed chips, it risks losing market share.
This does not necessarily negate the rationale for shifting to GPUs. Nokia believes that once customers deploy ABIG, they can benefit from years of AI-RAN software upgrades that continuously improve spectral efficiency, whereas custom chips require more frequent replacement. Following previous market share declines and concerns from Marvell about return on investment in the shrinking RAN industry, Nokia may have limited alternative chip options.
The issue is that if it turns out that custom chips remain competitive on AI-RAN, operators may be reluctant to spend money on GPUs—and VMO2’s trials appear to be demonstrating exactly this.
In one of the UK’s busiest areas, the trial focuses on wireless conditions between antennas and customers’ phones in the Paddington region. “Our focus with them is to use AI to better estimate channel quality,” said Rob Joyce, Director of Mobile Access Engineering at VMO2, during a recent press briefing.
Traditional algorithms rely on a set of predefined parameters for these wireless conditions. AI, however, learns by doing, like an eager apprentice, and quickly adapts to real-time factors. “You optimize based on what you observe, and it trains itself within its environment,” Joyce explained. “So the base station fine-tunes its own algorithm to that specific environment.”
VMO2 observed a 20% improvement in spectral efficiency at Paddington, meaning the site can handle more data without requiring additional capacity or spectrum. This 20% figure aligns with Nokia’s recent assertion regarding the spectral efficiency gains achievable through GPU-powered AI-RAN.
Joyce also acknowledged that Ericsson’s AI-RAN product is equally competitive. Ericsson continues to use custom chips; its CEO, Per Narvinger, previously stated that Nvidia’s “native implementation” creates lock-in. “You don’t need a GPU to achieve this,” Joyce emphasized during the press briefing. “Fairly speaking, I think Ericsson’s chips are also excellent.”
If anyone wasn’t surprised by AI-RAN being implemented on Marvell chips, it was Tommi Uitto, former president of Nokia’s Mobile Networks business group. “Years ago, someone smartly embedded certain AI and ML [machine learning] capabilities into chips—we just hadn’t used them yet,” he told Light Reading in early 2025. “Using relatively expensive GPUs for AI-RAN solely to improve network performance may not be technoeconomically viable,” he said at the time.
Perhaps the key issue is whether the normal iteration pace of custom chips can keep up with the pace of future software improvements on GPU platforms, which do not require as frequent hardware replacements as custom chips. Last summer, Nokia suggested that, regarding the 20% improvement in spectrum efficiency, a 50% gain could be achievable next year, with a doubling by 2028.
However, a white paper published by Chetan Sharma Consulting and endorsed by Nokia was far more conservative, even raising various concerns. Chetan Sharma, head of the consulting firm, wrote: “A site that is already approaching the theoretical spectral efficiency limit through conventional methods has less room for AI-driven improvements compared to a site that is significantly below the theoretical limit due to interference, channel estimation errors, or poor scheduling.”
None of the various lab tests and field trials for AI-RAN specify what baseline was used to measure the improvement in spectral efficiency. Experts believe that RAN algorithms are somewhat like Stradivarius violins—shaped and refined over decades, they are nearly impossible to improve further. It seems unlikely that AI could successfully create an instrument superior to that of this Italian luthier.
Regardless, trying ABIP on AI-RAN makes it even less likely that VMO2 will deploy GPUs at scale. Over the past two years, the company has spent approximately £1.4 billion ($1.9 billion) on mobile network upgrades, installing Nokia’s new radio equipment in southern Britain and Ericsson’s equipment in the north. The recently announced £700 million ($940 million) investment plan, confirmed by Joyce to Light Reading, does not include GPU RAN deployment.
He is also skeptical of GPU-based centralized RAN, a solution that moves baseband resources to other network facilities while keeping only wireless equipment on towers. “If you have high-capacity fiber, you can move the GPUs to a centralized location,” he said, allowing control over multiple sites. “This has advantages because you can coordinate scheduling across different technologies and further enhance performance.”
Joyce is concerned about the fiber optic requirements for these sites. The Paddington site is connected via a 10 Gbit/s fiber, which is sufficient for local needs. To centralize the baseband, VMO2 would need to reconnect the facility to the wireless site using a technology called CPRI (Common Public Radio Interface) or its upgraded version, eCPRI. “Each link is 25 Gbit/s,” Joyce says. “You could end up needing three CPRI connections, meaning 75 Gbit/s to the tower instead of 10 Gbit/s.”
Almost certainly, this means either increasing fiber or paying higher transmission costs, which also explains Joyce’s reservations about investing in centralized RAN. “My view remains that distributed RAN is the way forward; in this case, if you want to achieve this, you must add additional baseband GPU resources at the site,” he said.
For Nokia, the risk lies in customers continuing to resist upgrades that require new hardware, regardless of cost. The AI-RAN features available on Nokia’s custom chip products could further diminish customer motivation to upgrade. Conversely, if today’s custom chips cannot accommodate emerging AI functionalities already supported by GPUs, Ericsson also faces risk. Customers who have already benefited from link adaptation improvements on hardware used for years may find they need to replace their equipment sooner than expected to access more advanced AI capabilities.
Operators' concerns about the technology roadmaps of the two vendors could create an opportunity for another to expand its market share. However, due to the extremely limited number of suppliers in the RAN market, operators urgently need both Nordic companies to succeed.
More content: Europe