How Nvidia could threaten Ericsson and Nokia in 5G and 6G

Telecom executives are looking at Nvidia with a mixture of wonder and trepidation. The chipmaker's share price seemed like an ever-ascending helium balloon when it hit $974 on March 8. While it has subsequently dropped to less than $880, Nvidia's stock is still up 216% in just two years, valuing the company at about $2.2 trillion today. It has profited from the insatiable hyperscaler appetite for its graphics processing units (GPUs), chips serendipitously suited to be the trainers of generative artificial intelligence (AI). But it also has telecom in its GPU sights.

The basic idea is that an Nvidia GPU could be used for both AI and telco workloads at the network "edge," a catchall for smaller facilities that would bring IT resources closer to end users. In the old, pre-AI world, an equipment vendor like Ericsson or Nokia would provide all the purpose-built 5G products for a particular site. As a potential usurper, Nvidia is perhaps the most muscular challenge those incumbents have faced.

A couple of trends now favor Nvidia and other companies that hail from an IT rather than classical telecom background. The first is virtualization, the ability to run network functions on standard computing chips – shared with other workloads – instead of dedicated hardware. The other is the concept of the open radio access network (RAN). Older, proprietary interfaces forced an operator to buy all the products for a site from the same supplier's system. That prevented specialists from catering directly to a telco. With open RAN, they can theoretically be slotted in next to other suppliers.

Nvidia seemingly has no interest in building the radios that gobble up more than half of telco spending on the RAN. But it is developing both the chips and software that could be installed in an operator's central and distributed units (CUs and DUs), the server boxes that host network functions. In a virtual RAN, a telco would run the software on a mixture of general-purpose central processing units (CPUs) and specialist silicon. Through a superchip branded Grace Hopper, Nvidia has both.

Aerial takes flight

The Grace part is essentially a CPU based on the blueprints of Arm, a UK-based chip designer that represents the main architectural alternative to Intel's x86 platform. In the Nvidia setup, this would look after data link functions, commonly referred to as Layer 2 or the MAC in sector parlance, as well as some higher-layer software. Hopper, the GPU part of the superchip, would deal with Layer 1 or the PHY, the software hungriest for IT resources.

What differentiates Nvidia from other companies targeting the same opportunity – apart from its growth trajectory – is the comprehensiveness of its approach. Intel stumps up CPUs along with custom Layer 1 "accelerators" but leaves most of the software to partners such as Ericsson and Samsung (indeed, it has previously said Ericsson should focus on this rather than designing chips). The telecom forte of Marvell, a critical supplier to Nokia, is custom silicon for Layer 1. Arm-based Graviton chips from AWS can now host some RAN functions. But AWS has stayed out of RAN software design.

Nvidia, by contrast, seems to be expanding across these domains. It has previously claimed to have a complete set of Layer 1 software branded Aerial, which intrudes on the software role normally played by an Ericsson or Nokia. Qualcomm has a similar Layer 1 strategy, offering both silicon and software on a smart network interface card (a SmartNIC) that plugs directly into a server. But Nvidia now presents Aerial in a much grander and more ambitious frame.

"Aerial is the brand name for Nvidia platforms for wireless (5G/6G)," said Soma Velayutham, the general manager of Nvidia's AI, 5G and telecom business, by email. "It includes SDK [software development kit], hardware system etc." Layer 1, then, is just one component currently pitched as "Aerial CuPHY." The "Cu" in the name is a nod to CUDA, an Nvidia software platform compatible with popular programming languages such as C++ and Python.

On the RAN software side, Nvidia has begun moving beyond these Layer 1 boundaries. "The software-defined nature allows for the integration of additional functionalities within the RAN stack for end-to-end performance optimization," said Velayutham. It is on the lookout for expertise to advance its RAN ambitions, too. In the last few days, it has advertised for a senior software engineer to work on Aerial. The ideal candidate will have expertise at both PHY and MAC layers, according to the advert. Nvidia is offering a basic annual salary of between $180,000 and $339,250, depending on experience.

Having already been instrumental in setting up a new group called the AI-RAN Alliance, Nvidia this week unveiled what it calls a "6G research cloud platform." It includes various Nvidia tools for the incorporation of AI into 6G-like technologies, along with a "full RAN stack" for researching 6G, a standard that has yet to be nailed down by official bodies. Ecosystem partners include Fujitsu, Nokia, Samsung and SoftBank, among others, said Nvidia.

Velayutham downplays talk of Nvidia as a competitor to Ericsson, Nokia and other network equipment providers (NEPs). "Nvidia does not compete with NEPs but prioritizes collaboration," he said. "NEPs can integrate specific functionalities from Aerial into their existing solutions." Options for them also include taking only hardware or only software from Nvidia, he added.

Monopoly's not a fun game

Given the virtualization trend, some NEPs would probably welcome Nvidia as a hardware alternative to Intel, still the dominant player in the CPU market. "It should not be a monopoly and there should be diversity," said Tommi Uitto, the head of Nokia's mobile business group. "If you look today at Layer 2 and Layer 3 and transport functions, you can do it with Intel x86, you can do it with AMD CPUs, Arm core, and now we've even announced you can do it with Nvidia CPUs."

What's more, Nvidia's GPUs look well suited technologically to the specific demands of Layer 1, according to Uitto. "Graphics processing units, because of the vector processing, just fit Layer 1 compute very well," he explained. Vector processing can essentially crunch through an entire array of data in one go and therefore differs from scalar processing, which deals with one data item at a time.

Yet Nokia is sticking with Marvell as its Layer 1 silicon supplier and has no plans to introduce Nvidia in this part of the network. Instead, it has announced a deal where Nvidia could theoretically support 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," said Uitto. This differs from Marvell's merchant silicon offer, he points out. "Our chip is of course better in performance because the chip has some leading IP [intellectual property] from us, some IP from Marvell and some third-party IP."

Ericsson, meanwhile, is known to be dismissive of Nvidia right now as a Layer 1 option. Several years ago, the Swedish vendor examined the pairing of its own RAN software with Nvidia's GPUs. But the performance was not deemed good enough. And Nvidia's GPUs continue to be seen by various industry experts as power hogs compared with other silicon choices. Today, Ericsson's attraction to Nvidia, and involvement with the AI-RAN Alliance, appears mainly driven by interest in bringing AI applications to the RAN.

Nvidia has acknowledged that putting GPUs into the RAN will be economical only if they have an AI role and are not simply used for RAN workloads. In other words, its success hinges on the need for GPUs to support AI applications at the telco edge. Bruno Zerbib, the chief technology officer of France's Orange, is dubious and thinks lower-cost chip alternatives are on the horizon. "You have to be very careful with your investment because you might buy a GPU product from a famous company right now that has a monopolistic position," he said, avoiding a direct reference to Nvidia.

Concern about that monopoly may galvanize telcos and other companies in their hunt for alternatives. After all, the original goal of open RAN was to broaden the vendor ecosystem beyond Ericsson and Nokia as governments were turning against China's Huawei and ZTE. Other chipmakers slotting into open and virtual RAN deployments are relatively sanguine about Nvidia, too.

"I think what they want is for a lot of the RAN traffic to go through their GPUs, because there is a lot of volume, but the functions they provide don't need to be anything related to Layer 1," said Joel Brand, a senior director of product marketing at Marvell. "They could be related to security or routing or application network data."

Still, if this vision of AI at the telco edge pays off commercially, operators will ask why they should invest in parallel RAN architecture and not simply rely on Grace Hopper for everything. If this sounds too much like dependency on a single vendor, the likes of AWS, Google Cloud and Microsoft Azure might shoulder the investment, integrating Nvidia into the technology stacks they carry into a telco's own facilities. The potential concern here is about dependency on a hyperscaler instead. When it comes to network matters, there are no easy decisions.