Intel and telcos left in virtual RAN limbo by rise of AI RANIntel and telcos left in virtual RAN limbo by rise of AI RAN
A multitude of general-purpose and specialist silicon options now confronts the world's 5G community, while Intel's future in telecom remains uncertain.

Throughout the 5G era, Intel has been urging the telecom sector to abandon the custom silicon used in the radio access network (RAN) and switch to its general-purpose chips instead. Still dominant in the PC and server worlds, Intel's x86 architecture enjoys economies of scale that custom chipmakers catering to specific industries must envy. Piggybacking on it would liberate valuable resources within RAN developers. Operators could host other workloads on the same infrastructure.
Yet this phenomenon of virtual or cloud RAN has reached an awkward crossroads in the last year or so. Progress in the PC and server markets by AMD, an x86 rival, and chipmakers using the blueprints of Arm, an alternative architecture, has not been mirrored in virtual RAN, which Intel continues to dominate. But money that once poured into Intel's central processing units (CPUs) has instead been flooding into chips for artificial intelligence (AI), and especially Nvidia's graphical processing units (GPUs). Rather than build a multitenancy virtual RAN, able to host telco and plain IT workloads, telcos are being asked to consider a multipurpose AI RAN, good for 5G and more advanced AI applications.
Intel's current predicament is yet another reason for telcos to proceed with caution. Results published last week show a net loss of $19.2 billion in 2024, compared with a profit of $1.7 billion the year before, after sales fell about 2%, to $53.1 billion. That is almost $26 billion less than Intel made in 2021. Surging costs, including a bill for restructuring and other charges of about $7 billion, have torn into margins.
Markets fear the once-great company may succumb to challenges on multiple fronts. Efforts to set up an advanced foundry that can rival Taiwan's TSMC and South Korea's Samsung are gobbling funds. Hyperscalers that previously bought from Intel are building their own chips, based on Arm's designs, and spending more heavily on Nvidia's GPUs. Apple made a similar shift from Intel processors to internally developed, Arm-based silicon for its MacBook computers. In x86, AMD has clearly outperformed Intel, growing annual sales by almost $13 billion or 132% between 2020 and 2023 while Intel's fell by $23.7 billion.
The latest upset seems to have come in AI, with reports seizing on Intel's decision to scrap Falcon Shores, an AI chip previously set for launch this year. Intel now plans to use it for internal testing only and instead concentrate resources on "a system-level solution at rack scale with Jaguar Shores," said Michelle Holthaus, the company's interim co-CEO, on the analyst call about recent quarterly results.
Falcon Shores was originally supposed to be a GPU, Intel's response to Nvidia. But there may be alternatives to GPUs, and particularly Nvidia's GPUs, for AI workloads of the future. Industry talk grows louder about the move from training to inference, the less computationally intensive deployment of fully trained models to support AI applications. The possibility China's DeepSeek came up with a much more efficient AI model has encouraged the industry to think about substituting other silicon for Nvidia's GPUs. Holthaus is eyeing "a significant opportunity for CPU-based inference on-prem and at the edge as AI-infused applications proliferate," she told analysts.
The Nvidia factor
So far, Intel has had nothing specific to say about AI RAN, a term that seems to have been coined by Nvidia as a rescoping of virtual, cloud or "open RAN," the combination of different vendors at the same mobile site. The idea is to put less resource-hungry network software on Grace, an Arm-based Nvidia CPU, and run the demanding "Layer 1" functions on a Hopper- or Blackwell-branded GPU. Extra capacity on that GPU could be leased for inference at the edge. Nvidia reckons a telco could earn $5 over a five-year period for every $1 it invests, although there are few signs that telcos are biting.
What seems bound to deter some telcos is the risk this nascent AI RAN market ends up as an Nvidia enclave, just as the small virtual RAN subsector has been monopolized by Intel. In 2023, analyst firm Omdia, a Light Reading sister company, reckoned virtual RAN accounted for just a tenth of all the installed baseband worldwide. Omdia forecasts a doubling of this share by 2028. But that would still leave purpose-built compute with 80% of the baseband market.
The lack of commercial alternatives to Intel is not the only or even the main reason for the slow march of virtual RAN. In its latest filing with the US Securities and Exchange Commission, Intel said it had shown that 5G basestations "can be almost entirely built from software running on Intel Xeon processors with Intel vRAN Boost," its name for an accelerator that fuses hardware and software to provide added oomph for some Layer 1 functionality.
Yet an industry debate has continued to rage about the attractiveness of virtual RAN. Nokia, one of the world's biggest RAN vendors, prefers to keep Layer 1 on a separate custom chip supplied by Marvell, arguing this is more efficient. Using the same compute platforms for both RAN and other telco or IT workloads would probably require a telco to centralize its baseband resources, normally located at mast sites. In many countries, that would demand a considerable investment in fronthaul, including fiber links between mast sites and the facilities that house the IT resources. For many telcos desperate to reduce capital expenditure, it may not seem worth the effort.
The performance of Intel's telecom business has continued to look disappointing. Last year, the network and edge group (NEX) that includes that business reported sales growth of just 1%, to about $5.84 billion, and an operating profit of $931 million. While that was a big improvement on the $203 million Intel managed a year before, it was down from $1.5 billion in 2022, when NEX also had revenues of $8.4 billion. In 5G, moreover, Intel complained that customers had still been "tempering purchases to reduce existing inventories."
No doubt, Intel has been hurt by the downturn affecting the whole RAN market. When Omdia issued a forecast in late 2024, revenues generated globally from RAN product sales were expected to fall 12.5% last year at the midpoint of its range, to about $35 billion, after they had already dropped 11% in 2023. But Intel's focus seems to be wavering. It is known last year to have sought a buyer for some of its telecom assets and been rebuffed by at least one candidate. Under the latest restructuring plans, it will seemingly demolish NEX, moving the edge part into its client computing group and networking into its data center and AI business.
But a retreat by Intel would hardly suit telcos calling for optionality or the RAN vendors, such as Ericsson and Samsung, that have developed strategies based partly on Intel's product roadmap. Existing foundries are dedicating a bigger share of capacity to GPUs. The view of one prominent telco executive, speaking on condition of anonymity, is that custom chipmakers serving relatively niche markets (like RAN) will be squeezed. Riding in the slipstream of AI chips could be the safest move.
If there is "a significant opportunity for CPU-based inference on-prem and at the edge," as Holthaus believes, AI just might bolster the case for building an Intel-based virtual or cloud RAN. There are considerable doubts about the economic viability of using expensive and power-hungry Nvidia GPUs at edge sites in the RAN. Developers would have to rewrite software for compute unified device architecture (CUDA), the Nvidia platform, whereas much of that work has already been done for Intel's x86 processors.
Forthcoming products should also be able to cope with more than just the RAN. That, at least, is the expectation of people involved in rollouts. Granite Rapids, Intel's successor to the Sapphire Rapids product currently in deployment, will "free up those existing platforms for new applications, or create more bandwidth, if you will, in the processor space to bring new applications into those nodes," said Paul Miller, the chief technology officer of Wind River, a developer of cloud computing infrastructure and Intel partner.
A confusing array of options
Intel, though, is not the only potential AI RAN alternative to Nvidia. In China, the DeepSeek model has reportedly been deployed for inference on the Ascend range of GPUs made by Huawei. The Chinese kit maker was last year showing off an "A-RAN" strategy that may conceivably unite those GPUs with Arm-based CPUs, as Nvidia has done. Arm's role as both AI and CPU developer could be critical. Press reports have suggested it is collaborating with SoftBank on the development of an AI chip for inferencing to challenge Nvidia.
Then there are the tensor processing units (TPUs), AI accelerators and Arm-based CPUs from hyperscalers, many of which could theoretically handle RAN workloads, too. AWS, notably, has positioned its Graviton-branded CPUs as a virtual RAN option, although it still had nothing to show for its efforts last July, when Light Reading caught up with Ishwar Parulkar, the company's chief telecom technologist. "The challenge is porting software to Arm," he said at the time. "Today, it is all primarily Intel-based, and there is work to be done, and it is engineering work."
A multitude of chips with different architectures would introduce complexity and confusion, potentially hampering development. Sachin Katti, the general manager of Intel's NEX group, has even argued that Arm is not a single system or product. "Every variant of Arm is different and so Ampere's Arm product is different from AWS's Arm product," he told Light Reading in late 2023. "And, frankly, I don't think software you write for Ampere is easily portable to Arm running on AWS."
The same criticism could, of course, be levelled at AMD as a kind of x86 fork in the road. Having seemingly made almost no commercial progress in virtual or cloud RAN against Intel, it had nothing "concrete" to say about any plans for AI RAN, it told Light Reading when recently approached on the topic. AI data center specialists like UK-headquartered Nscale are now buying GPUs from AMD as well as Nvidia. But for developers, AMD's GPUs mean dealing with a different instruction set from CUDA.
Even more specialized silicon could prove more suitable for AI inference. Jonathan Ross, the man behind Google's TPUs, left the hyperscaler in 2016 to launch Groq, a chips startup building a type of application-specific integrated circuit it calls a language processing unit (LPU). Large quantities of those LPUs were recently shipped to Saudi Arabia, said Tareq Amin, the former CEO of Aramco Digital who before that ran telecom businesses for Japan's Rakuten, in a LinkedIn post. Amin's current role is described as "confidential" on LinkedIn, although he is thought to be working in a senior AI position for the Saudi Arabian state.
The conundrum for telcos and their suppliers is balancing the demands for optionality and competition with the need for a platform that can tap into a broad ecosystem and global economies of scale. With Intel, Arm variants, Nvidia, other GPUs and more specialist silicon, the scales are in danger of tipping heavily toward fragmentation, plunging RAN developers and their customers into complexity. In a market that could benefit from a shove, the outcome threatens to be stasis.
About the Author
You May Also Like