Nvidia rules out an AI-FAN replay of its AI-RAN push into fixed access

Nvidia has told Telecom Gazette's sister coverage it will not extend its AI-RAN strategy into the fixed access network, insisting fixed broadband is "fundamentally different" from mobile and already well served by dedicated silicon. The verdict arrives just as its $1 billion investment in Nokia pushes the Finnish vendor out of custom RAN chip development and toward GPU-based software.

The contrast matters commercially. Nokia's fixed networks unit was the worst performer of the six businesses inside its mobile infrastructure and network infrastructure groups last quarter, with revenues down 22% year-over-year to €386 million (US$432 million). The company said OLT sales were stable, which implies the decline came from optical network terminals at customer premises. Declines at the unit are squeezing NI's margins.

Why does Nvidia say fixed access resists GPUs?

The argument is architectural. Wireless RAN does massively parallel, compute-intensive signal processing — MIMO, beamforming, channel estimation, interference cancellation — that suits GPU acceleration, an Nvidia spokesperson explained by email. Fixed access, by contrast, relies on "deterministic functions such as framing, multiplexing, FEC, and packet scheduling, which are already highly optimized in dedicated silicon."

Not everyone finds that airtight. Ericsson makes the same optimization case for its RAN custom silicon, yet its GPU trials have focused only on forward error correction — the very function Nvidia labels deterministic — alongside non-RAN workloads, with the rest of the virtual RAN running on CPUs.

The industry has also been here before. Under the Broadband Forum's Open Broadband Initiative, the standards body tried virtualizing OLT functionality into a software layer, Stefaan Vanhastel, Nokia's VP of marketing and innovation for fixed networks, said last year. It stalled because moving real-time functions from customized equipment onto general-purpose hardware proved too hard.

Where does the fixed-line opportunity actually sit?

Dean Bubley of Disruptive Analysis broadly agrees the FAN is simpler than the RAN — there's no fixed-line equivalent of beamforming, for one. But Nvidia hasn't abandoned fixed networks entirely. It sees value higher up the stack, in "intelligent bandwidth allocation, traffic prediction, network optimization, and autonomous operations," plus DPUs in virtualized broadband gateways, security and programmable networking.

That's a more nuanced position than it sounds. A virtualized broadband gateway would shift OLT functions onto commodity servers. Nvidia's BlueField DPUs, already part of its AI-RAN offer alongside Arm-based CPUs and GPUs, would handle the networking, storage and security chores.

"The opportunity is not necessarily to replicate AI-RAN (PHY acceleration) in fixed networks, but to bring AI and accelerated computing to the areas (higher up in the communication stack) where they create the greatest differentiation and economic value," the spokesperson said.

Is the real prize the buildings, not the boxes?

Nvidia's Berlin GTC press release pitched networks as "distributed intelligence platforms." The arithmetic explains the ambition. A company now reporting quarterly revenues of almost $100 billion is chasing a RAN market worth only $35 billion a year in product sales, and a fraction of that in silicon — a mismatch that still mystifies many experts.

The likelier play is real estate. Cabinets and central offices suit conversion into small AI data centers far better than mast sites do. Nvidia's chips may not replace the Quillion silicon inside Nokia's OLTs, but fixed-network facilities could still host the GPU fleets. Nokia hasn't commented publicly on any of this, and the RAN transition, with Nvidia GPUs destined for baseband units and possibly radio units, remains its stated path for now.