Verizon and other carriers push against Nvidia's AI plan for their networks

GPUs are great for training huge AI models in big data centers. But using them at cell sites or inside small telecom buildings creates new problems. They use far more power than the chips normally found in mobile networks, and that extra energy cost is hard to justify unless GPUs can either become much more efficient or help operators make more money.

Nvidia points to its ARC-Pro hardware as the answer. It's a small GPU-based platform designed for use at a mobile site. But it reportedly draws around 300 watts. That's far higher than some custom-made telecom chips, which can do similar jobs using a fraction of the power. Nvidia argues the efficiency is competitive, but the fast-growing energy use of GPU-filled data centers makes operators nervous.

And this hesitation makes perfect sense to me. Everywhere I look, operators are under pressure to trim costs as traffic keeps climbing while revenue barely moves. Every watt and every dollar is being scrutinized, and I see how no one wants to take a risk that could push bills even higher.



Some top telecom executives openly doubt the idea. Verizon's Yago Tenorio said decisions should be made based on what is best for each base station, not on trying to build extra services around expensive GPU platforms. Telus and Orange also said they see no solid business case: converting local telecom sites into mini data centers would require huge upgrades in space, electricity, and fiber links.

Why AI?

The main argument in favor of GPUs is that AI could improve spectral efficiency, meaning networks could carry more data using the same airwaves. But even here, the need for GPUs is unclear. Ericsson already showed it could get a 10% boost in efficiency using AI algorithms running on regular processors, not GPUs. Nokia itself previously claimed its existing chips were strong enough for advanced AI tasks.

If GPUs were truly essential, more vendors would be embracing them. Instead, much of the industry is looking to Intel's latest CPUs, which appear powerful enough for running 5G and even 6G features without the cost and energy burden of a GPU. Vodafone's upcoming open RAN (radio access network) rollout in Europe, for example, will rely heavily on Intel's chips.

Even companies building cutting-edge AI software for networks say they don't need GPUs today. Cohere Technology, backed by major operators, has shown real-world gains using ordinary servers.

One day soon

For now, most telecom operators simply don't see a practical or affordable reason to bring GPUs into their radio networks, but I personally think AI isn't going anywhere. The carriers will try to integrate it more and more – not in their networks per se, but in their customer services, for example. In their proprietary apps. In their offices.

Maybe that's why Verizon is about to cut 15,000 jobs as soon as next week?