"I'm excited about reinventing the telecommunications," smiled Nvidia boss Jensen Huang, seated in his customary leather jacket alongside Mike Sievert, the CEO of T-Mobile US. The occasion was the US operator's capital markets day, and the ambitious statement was typical of the man who runs the stock-market darling of artificial intelligence (AI), a chipmaker whose market capitalization has rocketed from $333 billion to almost $2.8 trillion in just two years.

Huang is rarely heard discussing the intricacies of the radio access network (RAN), but his vision, if realized, would have a massive impact on it. For several years, the telecom industry has talked at length about virtualizing the RAN. By substituting central processing units (CPUs) and other general-purpose compute platforms for the custom hardware and software of a traditional mobile network, an operator could host its RAN in the cloud, share resources with other telco and IT workloads and be a lot more efficient. The big problem? "Moore's Law has really run its course," said Huang on stage. "The CPU can't keep up with the workload of an ASIC."

Spelt out, that is the application-specific integrated circuit the likes of Ericsson and Nokia would produce for a purpose-built 5G RAN, one resistant to cloudification. Naturally, this is not what Intel – the world's biggest maker of CPUs, and the dominant force in the tiny virtual RAN market – would have the industry believe.

Yet even Intel uses custom "accelerators" for a task called forward error correction in the RAN's Layer 1, the software slice hungriest for compute resources, and its roadmap shows other computationally intensive 5G workloads hosted on these accelerators in future. Earl Lum, the president of EJL Wireless Research, reckons the hardware is likely to be a field programmable gate array (FPGA) or something even more customized. "One alternative to the FPGA could be the eASIC technology," he said. "You take an FPGA, and you harden it and make it more efficient."

Fudging virtualization

Any such customizing, though, arguably marks a retreat from virtualization, and some companies have gone much further with it. Intel's biggest partner among telecom equipment makers is Ericsson, which combines its virtual RAN software with the chipmaker's CPUs and accelerators. But Nokia, the other big Nordic vendor, prefers a different approach of putting all Layer 1 functions on an accelerator made by Marvell Technology. Software and hardware are tightly coupled in these chips customized for the RAN.

Neither approach has taken off, as Sievert noted. "There has been a trend in this industry that we're moving toward a more open RAN architecture, toward software-defined RAN, trying to get all that technology out of the radio and into the cloud for obvious financial benefits and efficiencies, but that has proved kind of elusive at scale around the world," he said.

Cue Nvidia and the opportunity for Huang to weigh in. His answer to those limitations of an Intel CPU, of course, is an Nvidia GPU (that's graphics processing unit) combined with a fully deployable Layer 1 branded Aerial, itself based on CUDA, the software platform "for general computing on GPUs," in Nvidia's words. "It took us five years to create the Aerial library, which allows us to use CUDA to accelerate 5G radio and prepare for 6G radio," said Huang.

"Once we do that the computation stack can also support AI, and now we've fused signal processing and AI into one computing platform," he continued. "Your radios are going to be extraordinary 5G radios, and be prepared for 6G in the future, as well as an AI computing platform to deploy, to host, to run AI on it. And the AI could be used to improve the efficiency of the network itself as well as host brand new AI services."

Gatecrashing the RAN

The vision was initially presented at the very start of the year, when Nvidia assembled various other RAN stakeholders in a group called the AI-RAN Alliance. But the establishment with T-Mobile of what the companies are calling an "AI-RAN Innovation Center" at the telco's headquarters in Bellevue brings it a step closer to reality.

All this would, though, require a big overhaul of existing RAN architecture and systems to exist at "scale," as Sievert put it. And Nvidia's gatecrashing of the virtual RAN party seems bound to have ramifications for both Ericsson and Nokia, the two big 5G kit makers outside China, along with their current chip partners. Aerial, as a fully deployable Layer 1, would seem to impinge on what the Nordic kit vendors do on the software side. Indeed, the algorithms developed for some of those Layer 1 tasks are considered a point of competitive differentiation.

Both Ericsson and Nokia were parties to this week's update. In the arrangement Huang seems likely to prefer, their RAN software for Layers 2 and above could be combined with an Nvidia CPU, branded Grace and based on the blueprints of Arm, which offers an architectural alternative to Intel's x86 technology (one many consider more power efficient). This would then be combined with Aerial on a GPU for Layer 1.

Yet besides giving up a Layer 1 role, Ericsson would also be retreating from its use of Intel products that integrate the CPU with the accelerator on the same die. It could, however, fall back on a lower-cost Intel CPU for Layers 2 and above, pairing that with Aerial on Nvidia's Hopper or Blackwell GPUs for Layer 1. Fredrik Jejdling, the head of Ericsson's networks unit, sounded wary in prepared remarks for an Nvidia blog. "We are now evaluating the performance and cost of NVIDIA accelerated computing in this context," he said.

Nokia, similarly, would have to abandon Marvell. Earlier this year, Tommi Uitto, Jejdling's equivalent at the Finnish vendor, noted the attractions of Nvidia technology. "Graphics processing units, because of the vector processing, just fit Layer 1 compute very well," he said. But he also said there were no plans to replace Marvell. Instead, he saw Nvidia as an option for some of those higher-layer functions. "We say Layer 2 plus with Nvidia and a SmartNIC for Layer 1, where we use our Layer 1 chip," he told Light Reading at Mobile World Congress. The SmartNIC is a card that plugs into a server, hosting Marvell's silicon and Nokia's software.

The Finnish company is sticking to the line that a partnership with Nvidia has no immediate impact. "This does not change Nokia's anyRAN strategy for cloud RAN," said Aji Ed, the head of partner cloud solutions, by email. "As a part of the AI-RAN innovation center, Nokia is evaluating the feasibility of bringing AI and RAN workloads together."

"Although not all the benefits are clear, AI-RAN has the potential to transform today's radio access network into a multi-purpose radio access and AI service platform in future," he added. "This can open new business opportunities for CSPs and introduce new innovative use cases for end users. Nokia together with its partners is exploring the potential of GPUs and AI to boost capacity, performance, efficiency and user experience in 5G and eventually in 6G with an AI-based air interface."

An Nvidia product able to handle RAN compute and AI workloads would tick the general-purpose and virtualization boxes, as Huang regards them. But there are numerous other potential obstacles besides the technology strategies of Ericsson and Nokia. While Huang thinks AI could help to reduce energy consumption, GPUs are themselves extremely power hungry when compared with CPUs and other accelerators. Edge computing, whereby applications are hosted not in large data centers but at or near mobile sites, has been a discussion point for even longer than the virtual RAN. But it has barely advanced and not everyone is convinced AI will be the catalyst.

Even so, Intel is in a torrid state and there is speculation that its recent restructuring, which seems to leave network and telco assets in a discrete unit, could be the precursor to a sale, with Marvell identified as a potential buyer. Trouble for the world's biggest CPU maker is trouble for the virtual RAN market, which has lacked Arm- and x86-based alternatives to Intel. With its apparent vision of GPUs as the general-purpose successors to CPUs, Nvidia might just have an answer.

Update: Amended to include a Nokia quote since first publication.