"Reproducing kernel Hilbert space" (RKHS) is not a sequence of words that would be strung together in many conversations, even by someone immersed in technology at Nokia. It describes an algorithm used in machine learning, little known outside expert circles, and has been largely ignored by Nokia's radio access network (RAN) people simply because the infrastructure couldn't handle it. "We never used it because it's very mathematically compute-intensive," said Pallavi Mahajan, the Finnish company's chief technology officer. "It has a lot of vector computation."
That all changes now. "With GPUs, you can do that, and you get efficiency," said Mahajan. Those GPUs are the graphics processing units designed and sold by Nvidia, the world's biggest company by market capitalization. They have underpinned the generative AI boom, providing a much more powerful platform than older central processing units (CPUs) to train large language models (LLMs). Nokia's experts believe they could be equally valuable in the RAN as the industry begins work on 6G.
Around ten months have now elapsed since Nokia accepted a $1 billion investment from Nvidia. That came with a sector-shaking announcement of plans to design RAN software based on CUDA, the chipmaker's platform, and make GPUs integral to Nokia's future 5G and 6G products. For a time, these will coexist with the custom silicon that Nokia obtains from Marvell Technology, its longstanding 5G partner. But Nokia has been emphatic: the Marvell chips in its newest products will be the last custom silicon it ever develops.
An Nvidia GPU will provide a hardware foundation able to withstand anything 6G throws at it for a long time – or so Nokia believes. "Whatever we do from now on, it's just like six years of software upgrades," said Mahajan. The ability to use RKHS is merely one attraction of moving to a GPU. "It enables us to go out and run different algorithms, which previously we could not," she explained.
Nokia's claim that all this could double spectral efficiency by 2028 has provoked varying responses. Some have shown outright disbelief. A few have reacted with drooling tongues, like ravenous diners impatiently awaiting a banquet. The promised improvement dramatically and pointedly exceeds what rival Ericsson previously said was possible on custom silicon running AI software. Like Marvell's chips, that will carry Ericsson only so far in 6G before it needs replacing, Nokia reckons.
What the FEC?
Yet the idea that 6G will ultimately demand a GPU is soundly rejected by the Swedish vendor. "As previously communicated, GPUs are one option for 6G and AI RAN, not the only one," said a spokesperson for the company by email before harping on the appeal of custom silicon. "Purpose-built silicon remains relevant where operators need performance, lower energy consumption, and the ability to scale RAN workloads efficiently. We expect future networks to use a mix of compute architectures, depending on the use case and operator requirements."
Ericsson is not totally hostile to the concept of a GPU RAN and would probably not wish to be too critical of a chipmaker that excites stock markets like no other. Even if it has done so without much enthusiasm, it has for years offered customers a set of virtual or cloud RAN products based on CPU architecture, rather than custom silicon. Running an energy-hungry task called forward error correction (FEC) on those CPUs gobbles too much power to satisfy telco demands. Intel stumps up a hardware accelerator to handle FEC and integrates that with its latest CPUs. An alternative that Ericsson has investigated would replace Intel's Xeon processor with an Nvidia Grace CPU and rely on a GPU as the FEC accelerator.
But Ericsson's experts evidently disagree that GPUs would confer AI-based RAN benefits unrealizable with CPUs or custom silicon. Largely, that is because there are constraints outside the data center to do with power consumption and latency, a measure of the time a network signal takes to complete its journey. RAN models must necessarily be more lightweight, and that means they do not require a GPU, according to Ericsson.
As an analogy, picture a building site where objects of various weights must be moved from one point to another. Only the strongest two or three workers can lift the heaviest items. But employing them to carry lighter goods would be a wasteful misallocation of resources.
A RAN constraint that no one can avoid is called the transmission time interval (TTI). "We have 500 microseconds, half a millisecond, to get this done," explained Michael Begley, Ericsson's head of RAN compute, when he met with Light Reading at this year's MWC Barcelona tradeshow. "So, you can't do a million-parameter model. You need to compress it. Otherwise, there's no point because if you miss the TTI, radio conditions have changed and whatever you calculated isn't relevant anymore."
AI models that include millions, billions or even trillions of parameters might need the muscle of a GPU. But when they are limited by hard science to only tens of thousands of parameters, Ericsson's AI-RAN models do not. "Those kinds of rules of physics and specifications are the same for Nvidia," said Gabriel Foglander, Ericsson's head of RAN strategy, during a more recent meeting at the company's Stockholm headquarters. "Inferencing will need to be met, whatever compute you want to put in there."
Sports cars and SUVs
Similar opposition to Nokia's argument that 6G will ultimately necessitate a GPU RAN comes from Samsung, the world's fifth-largest 5G vendor. "Samsung does not believe that GPUs are a prerequisite for 6G, nor that a CPU-only or GPU-only approach is the right choice," a representative said by email. "We see CPUs continuing to serve as the primary general-purpose foundation for RAN, with GPUs or other accelerators added where additional performance is needed and the workload can benefit from it."
But there are some important nuances that distinguish Samsung from Ericsson. The Swedish vendor's AI-RAN strategy is geared heavily to its own custom silicon, which accounts for nearly everything it sells today. Samsung, by contrast, is now mainly invested in virtual RAN products based on Intel's CPUs. That could explain why Samsung has expressed slightly more interest in using GPUs for AI-RAN.
The South Korean vendor likens a CPU to a sports utility vehicle (SUV) and a GPU to an actual sports car. "A sports car may be faster, but an SUV is more versatile for everyday use – carrying people, traveling long distances and operating efficiently across different needs," said the company in its email. A CPU provides a general-purpose foundation for the everyday needs of RAN compute. But certain "intensive" RAN workloads might benefit from the high-performance compute offered by a GPU.
The limits imposed by the TTI make CPUs a more "natural fit" for most tasks, in Samsung's opinion. At the same time, it echoes the views of Nokia's Mahajan about the usefulness of GPUs for computationally demanding workloads. Nvidia's chips are, it elaborates, "fundamentally optimized for highly parallel processing and can execute highly parallel workloads very quickly."
The challenge seems to be the constant need to cram enough data into a GPU's hungry mouth to make system-wide use of them look economical. More data equals more throughput, and more throughput risks increasing system latency and overstepping the TTI mark. "This creates a trade-off between maximizing GPU performance throughput and meeting tight TTI requirements," said Samsung in its email. "To meet the required TTI, some of that throughput advantage may therefore need to be sacrificed."
Adding intrigue is Nokia's claim that it has been able to realize the dramatic improvements in spectral efficiency and other measures without boosting energy consumption. Its forthcoming GPU products include a card designed to slot into an existing Airscale-branded baseband chassis and execute the same tasks as today's Marvell cards for newer releases of the mobile standard.
"They work with a power envelope of 190 watts. Now that we're bringing in the new card, it has to work in exactly the same power envelope," said Mahajan. "We cannot have our customers change their power cabling, their racks and all of that stuff. But it still brings 1.5 times the capacity, two times the spectral efficiency."
According to one industry view, a GPU capable of addressing later-stage 6G needs would necessarily be over-specified for today's requirements. Yet Nokia's roadmap includes an array of products, from the power-limited card usable with today's Airscale chassis to a more advanced appliance. "It's a compact pizza-box form factor where you have both GPU and CPU compute inside," said Mark Atkinson, the head of Nokia's RAN business. "We'll then be able to have the full benefit of GPU because we won't be limited by any power headroom anymore."
Running with ASICs
Still, if GPUs are superior to CPUs for running certain AI algorithms, the claim they will beat an application-specific integrated circuit (ASIC) is more controversial. Google, notably, has been shifting from GPUs to what are known as tensor processing units (TPUs) for the training of LLMs. Co-designed with chipmakers Broadcom and Marvell, those TPUs are a type of ASIC optimized for AI compute. To Google, they clearly hold attractions compared with Nvidia's GPUs. Both Anthropic and OpenAI, the leading US developers of frontier models, also reportedly now use TPUs for LLM training.
Besides developing custom silicon for its baseband equipment, Ericsson has recently started introducing an ASIC it calls a "neural network accelerator" into its advanced, massive MIMO radios. That allows it to support features including an AI-native scheduler for link adaptation, AI-managed beamforming and AI-powered macro positioning, said Foglander. "It's mathematics-heavy processing in the radio," he explained.
There is certainly nothing in the emergent 6G standard that mandates use of a GPU, according to knowledgeable sources. Whatever the pros and cons of ASICs, CPUs and GPUs, market forces and persistent concerns about "vendor lock-in" may be more significant factors in decision-making. In 5G, Nokia had always looked more dependent for its custom silicon on an external player – namely Marvell – than Ericsson ever did. Justifying the investment would have become harder after Nokia's loss of market share in the lucrative US. Eager to find a client for its GPU RAN pitch, Nvidia might have been the only viable option.
The criticism is that Nokia has traded lock-in with Marvell for lock-in with Nvidia and CUDA, deemed incompatible with other silicon platforms. "If the GPU thing doesn't catch on, they don't have an alternative," said Vincent Loncke, the founder of a startup called Kenyi Technologies and a former Qualcomm executive.
By contrast, the ability to move software written for Intel's CPUs to other x86- or Arm-based chips, with only minor changes, seems to give Ericsson and Samsung some degree of independence in virtual RAN. So far, however, nearly all deployments are based on CPUs from Intel, a business that reported a $15 billion net loss for the first half of 2026 and one whose workforce shrank by 46,800 employees, more than a third of the earlier total, between 2022 and 2025.
Besides Huawei, Ericsson is the only RAN vendor still investing in its own custom silicon. Per Narvinger, its incoming CEO, sounds committed to that strategy and has repeatedly rejected assertions from outside that Ericsson will eventually have to bow to economics and embrace general-purpose hardware. As the world's biggest RAN vendor excluding China, the Ericsson of 2026 can still make the numbers add up. But if telcos spend less on 6G than they have on 5G, or Nokia can stage a comeback, that formula could start to look shaky.