We Are Getting AI-RAN Wrong

GPUs don’t fit the physics, economics, or culture of telecom. AI-RAN today is driven more by FOMO than by feasibility.

We Are Getting AI-RAN Wrong

AI-RAN has become the new shiny object of telecom. Every major conference now features a panel where someone displays a slide of neural networks neatly integrated into the 5G stack. Nvidia demos a baseband on CUDA, SoftBank runs Transformers on Hopper GPUs, and analysts predict “AI-native networks” by 2030. On paper, it looks irresistible: take the hottest technology in the world and bolt it onto the most expensive part of the telecom industry.

But having spent years in both worlds, building RAN roadmaps and publishing AI research, I can tell you: we are approaching this wrong. The physics, economics, and culture of telecom don’t bend just because AI is fashionable.

Let’s take one practical example. A 5G baseband unit today costs around $10,000 in a telco warehouse. It’s a hardened appliance designed to run 24/7 with predictable latency and a power budget tuned to a cell site. Now compare that with a Hopper GPU board, which lists for $25,000–40,000 and pulls 350–400 watts. The GPU looks like a Ferrari engine dropped into a bicycle: technically impressive, but absurd in context. Telcos don’t optimize for luxury horsepower; they optimize for efficiency per watt and dollars per site.

And yet, AI-RAN is moving forward anyway. Not because the math works, but because telcos are terrified of missing the AI train. They missed the cloud. They missed platforms. They can’t afford to miss AI. This fear translates into trials and press releases, what I call the Telco FOMO Effect. Nvidia is happy to play along, but to them, the entire RAN market is a rounding error compared to a single hyperscaler order.

The result?

We’re chasing proofs of concept that won’t scale, while ignoring the deeper redesign needed to make AI fit the brutal constraints of wireless networks.