T-Mobile Runs World's First AI-RAN Call Over Nvidia GPUs

Swapping custom telecom silicon for graphics processors

Cellular base stations normally depend on custom application-specific integrated circuits (ASICs) to decode radio frequencies. Those chips do baseband processing efficiently, but they cannot do anything else. T-Mobile's test in Bellevue swapped out that dedicated silicon for servers powered by Nvidia's Aerial RAN Computer platform running Nokia's CUDA-based anyRAN software.

Instead of connecting test equipment inside an isolated lab, the site routed a standard smartphone call straight into T-Mobile's commercial 5G Standalone core. The primary milestone here is simple survival: the GPU-based stack handled production radio signals without dropping the user.

T-Mobile AI-RAN Bellevue trial setup

The missing performance numbers behind the test

While Nokia claims AI-RAN could eventually double spectral efficiency, neither T-Mobile nor Nvidia published latency, throughput, or power consumption data from the Bellevue trial. The call confirms the platform works, but it does not prove that running basebands on heavy GPUs delivers a better mobile connection.

The comparison gets murkier when looking at recent carrier tests. Just before this announcement, T-Mobile revealed a 16% downlink throughput bump during busy hours in New York using an AI beamforming model. That trial, however, ran on Ericsson's proprietary silicon—not Nvidia hardware.

Telecom operators now face an awkward question: why pay the heavy energy and cost premium for Nvidia accelerators when telecom equipment makers can add machine learning algorithms to existing, lower-power custom chips through software updates? Meanwhile, mobile hardware continues to push forward at the consumer end, where flagship devices like the Find X10 Pro Max depend heavily on cellular efficiency to justify their footprint.

Availability and timeline

Subscribers will not see any immediate change on their devices. Voice calls sound identical, and the underlying radio infrastructure remains an experimental deployment. Nokia plans pilot rollouts for its AI-RAN hardware toward the end of 2026, targeting commercial network gear for 2027.

Broader operational deployments will likely wait until carriers prepare for early 6G standards around 2030, meaning this architecture is a foundational trial rather than an impending consumer rollout.

Connecting a commercial phone across GPU-driven radio software is an impressive engineering feat for Nokia and Nvidia. Still, carriers will demand hard numbers on power efficiency and operational costs before ripping out battle-tested telecom silicon across thousands of towers.