Jensen Huang loves to watch his dogs snatch balls from the air, sometimes mid-somersault. "My puppies don't know Newtonian physics," said the leather-jacketed boss of Nvidia, on stage at T-Mobile's capital markets day this week. Training and practise have given the animals a predictive ability he likens to the artificial intelligence (AI) running on Nvidia's chips. "It doesn't understand the causality, but it predicts it wonderfully," he said.

The tech man of the moment was talking of plans to make AI an expert weather forecaster. But the same logic could also be applied to the operation of a radio access network (RAN). Exposed to the public for the first time this week, a new Nvidia computer, called ARC-1, includes all the baseband hardware and software a telco needs to run a 5G network. It also doubles up as a host of AI applications that could predict network outages or support a better service for customers.

Nvidia's network ambitions first came to attention several years ago when it began talking about Aerial. Based on CUDA, the software platform for Nvidia's graphics processing units (GPUs), Aerial at first covered all the functions needed for Layer 1, the slice of RAN software hungriest for compute resources. Nvidia's original plan was to host this on one of its Hopper-branded GPUs and combine it with a more humdrum central processing unit (CPU) for the less computationally demanding RAN functions, provided by the likes of Ericsson and Nokia. But Nvidia, it transpired this week, has now gone much further.

The full monty

In its latest guise, Aerial covers the full stack, including all the functions in Layers 2 and above that Nvidia would previously have left to the Nordic kit vendors or another third party. ARC-1 effectively packages this up with compute and memory into a sophisticated appliance or server that can be slotted into a data-center rack or placed at a mobile site. It is perhaps the most serious challenge the traditional RAN vendors have faced in years.

Anyone dissecting this sophisticated machine would encounter various state-of-the-art components, explained Ronnie Vasishta, the senior vice president of Nvidia's telecom business, on a call with analysts and reporters. Besides its latest Blackwell GPU is a Grace-branded CPU, based on the blueprints of Arm, an architectural rival to Intel's x86 technology. These are linked using a high-speed interconnect system called NVLink. ARC-1 also includes one of Nvidia's Bluefield-3 data processing units to support fronthaul, the connection between baseband and a physical radio in the field.

"It has the CUDA libraries, the OS layer for accessibility to the Blackwell GPU," said Vasishta. "Because of the capabilities of the fronthaul, and because of the capabilities of the Grace-to-Blackwell NVLink, which is a unified memory architecture, you can now run high-performance RAN algorithms across that NVLink with very low latency."

As in Nvidia's original strategy, Aerial's Layer 1 functions are handled by the whizzier GPU, with the Grace CPU used for most other RAN software. "Many of the L2 functions will run on Grace CPU," said an Nvidia spokesperson, answering follow-up questions by email. "However, some compute intense functions like MAC Scheduler could benefit from GPU and AI acceleration."

Scandi noir

All this naturally seems like a big threat to Ericsson and Nokia. Besides developing purpose-built RAN products, each vendor has already teamed up with other IT players on a "virtual" RAN offer. Ericsson relies heavily on a partnership with Intel. Under that arrangement, most RAN software is hosted on an Intel CPU, leaving only one or two of the most demanding Layer 1 functions to an Intel accelerator, a more customized type of chip. Ericsson says one of its objectives is to ensure the same software can be deployed on a CPU from AMD, another x86 chipmaker. But it has conceded that moving Layer 1 software to an Arm-based chip would be harder.

As for Nokia, it also relies on Intel CPUs for virtual RAN software at Layers 2 and above. For Layer 1, though, it prefers to use customized silicon made by Marvell Technology. This software is tightly coupled with the hardware. Nokia could not simply transfer the code to another silicon platform without an extensive rewrite.

Both vendors, however, are a part of the AI-RAN Alliance, the group Nvidia formed at the start of the year, and both featured in this week's update about ARC-1 and the opening of an AI-RAN innovation center at T-Mobile's US headquarters. "Ericsson and Nokia are more than capable of leveraging the platform, the hardware and CUDA libraries, to build their own stack and deploy their own software on top of that as well," said Vasishta, when asked how Aerial could fit into the strategies of the Nordic kit vendors with their existing arrangements.

In other words, Ericsson and Nokia could rely on Nvidia as a hardware partner and reference architecture, combining their software with its chips. But this sounds like it would require a significant new investment to develop a CUDA-based Layer 1. As an alternative, they could potentially modify their software for Layers 2 and above to run on a Grace CPU, a seemingly easier task, and relinquish the Layer 1 job to Nvidia. Yet that would mark a big retreat from their normal role.

Both vendors, of course, have the advantage over Nvidia of building radio units. Nvidia says all its new products are fully compatible with new "open RAN" specifications, allowing telcos to pair one company's baseband (the ARC-1 computer, in this case) with another supplier's radios. But integration remains a challenge and the performance of massive MIMO, an advanced 5G technology, is better when all the products come from the same vendor's system, said Tommi Uitto, the head of Nokia's mobile business, at the start of the year.

There is nothing to say this cannot be a win-win for all parties. Nvidia is likely to value the radio expertise of Ericsson and Nokia just as the Nordic vendors prize Nvidia for its GPUs and AI smarts. But 5G expenditure has recently dried up. Spending fell 11% last year, to about $40 billion, according to Omdia, a Light Reading sister company, which forecasts another decline of between 7% and 9% this year.

In that environment, ARC-1 will sound to companies struggling to recoup their investments like an expensive purchase. Of course, with revenues for its most recent quarter of more than $30 billion, up 122% year-over-year, and a gross margin of 76%, Nvidia could always stump up the money.

Update: Amended to include an email comment from Nvidia.