U.S. carriers split on GPU role in AI-RAN deployments

Diverging paths to AI-powered radio networks

The three major U.S. wireless carriers are pursuing artificial intelligence in their radio access networks with notably different technical philosophies, particularly around the role of graphics processing units.

Verizon CTO Yago Tenorio drew a sharp distinction between using GPUs for applications versus embedding them in base stations. While praising GPUs for edge computing tasks—such as powering prototype smart glasses that overlay player statistics at sporting events—he questioned their necessity in radio equipment itself.

"Do you need a GPU for doing just RAN? No, today, you don't. You can do that with a CPU," Tenorio told Fierce Wireless. He argued that embedding AI inferencing to improve radio performance can be accomplished with CPUs, and that this will likely remain true for years to come. Using GPUs to reprogram the entire radio stack is "unnecessarily complicated," he said.

T-Mobile commits to GPU-accelerated infrastructure

T-Mobile has taken the opposite stance, working closely with Nokia and Nvidia at its AI-RAN Innovation Center in Bellevue, Washington. The carrier expects to receive AI-RAN radio prototypes from Nokia later this year for extensive testing.

T-Mobile Network CTO Ankur Kapoor acknowledged Nokia's projections of achieving 50% spectral efficiency gains by late 2027 and 100% by 2028. Early demonstrations are already showing improvements above 30%, he said during a recent industry summit.

For T-Mobile, these efficiency gains translate directly to expanded capacity for fixed wireless access customers and higher broadband speeds. Kapoor dismissed concerns about GPU power consumption and cost, framing the decision around customer benefits rather than infrastructure metrics. He indicated deployment will be selective rather than universal, depending on demand and service requirements in specific locations.

AT&T evaluates multiple silicon options

AT&T appears to be charting a middle course. Rob Soni, vice president of RAN Technology at AT&T, said the carrier defines AI-RAN broadly as any technology bringing AI into radio operations, including troubleshooting, debugging and network optimization applications.

The carrier is evaluating GPU platforms alongside its primary vendor Ericsson, which won AT&T's open RAN business but did not receive the $1 billion Nvidia investment that Nokia secured. AT&T has successfully demonstrated Integrated Sensing and Communication use cases with Ericsson using CPU-based systems.

"We have an ability to consume any silicon," Soni said. "If a GPU shows up and it fits the network and it makes sense from a business perspective, we'll do it. But we won't just do it just for the sake of doing it."

AT&T is not currently interested in running third-party workloads at cell sites unless a clear business case emerges, Soni added.

Why it matters

The divergence among major carriers reveals that AI-RAN is not a monolithic technology path but rather a collection of approaches with significant cost and architectural implications. Verizon's CPU-first stance could influence vendor roadmaps and challenge the GPU-centric narrative that Nokia and Nvidia have promoted. For equipment vendors and chipmakers, the lack of consensus means they must support multiple implementation strategies rather than converging on a single standard, potentially slowing widespread deployment and complicating the business case for specialized AI-RAN hardware investments.

These details were first reported by Fierce Wireless.