Nokia has started shipping cell-site equipment that carries a graphics processor beside the radios that move mobile traffic, a wager that the same silicon running artificial-intelligence data centers can also run a phone network.
The Finnish vendor began delivering GPU-based radio access network gear to several operators this quarter, Light Reading reported on October 12. The systems are built with Nvidia, pairing Nokia’s anyRAN software with Nvidia’s Aerial AI-RAN platform so that one network runs both AI inference and the signal processing that keeps calls and data moving. Carriers are testing the hardware now, and the industry is split over how much it will earn them.
The argument behind the gear answers an imbalance telecom operators know well. Carriers own towers, spectrum and power connections that sit idle much of the day, while AI companies are hunting for exactly those resources. If a base station can run models in the gaps between traffic, Nokia contends, a radio network becomes a distribution channel for computing rather than a pure cost center. Operators, in that telling, would sell access to intelligence over the same pipes they built to sell connectivity.
The two companies have worked toward the idea for two years. Nvidia and Nokia announced a strategic partnership in October 2025 to add Nvidia-powered AI-RAN products to Nokia’s line, with T-Mobile US as an early partner folding the technology into its 6G work. In July 2026 the pair introduced what they described as the first commercial GPU-based AI-RAN platform, promising gains of more than 100% in spectral efficiency by 2028, roughly doubling the capacity operators can draw from airwaves they already license. Field tests have since spread across Europe, Asia and North America, including work with T-Mobile, Indosat and SoftBank.
Nokia is not alone in chasing the market. The AI-RAN Alliance, formed at the industry’s Barcelona trade show in 2024, counts Ericsson, Samsung, SoftBank and T-Mobile among its members, and each has an interest in pulling computing toward the edge of the network. Nvidia’s role as the supplier of the accelerators underneath gives it a stake on every side of the competition, much as it profits whether AI models are built by open or closed labs.
The technical case is narrower than the marketing. A base station’s traffic load rises and falls through the day, and carriers have long wanted to monetize the quiet hours. Running AI inference on the same hardware that handles radio scheduling could trim the power bill, a sizable cost for operators who buy electricity by the megawatt. But the two workloads make different demands. Radio processing is unforgiving about latency, and a dropped packet is a dropped call. Splitting one machine between a real-time network and a batch AI job is an engineering problem the industry has not fully solved.
Analysts are divided on the payoff. Supporters see a new revenue line for carriers that have watched average revenue per user stagnate for a decade. Skeptics note that operators have heard similar promises before, around edge computing and private 5G, and that few have produced meaningful income. The cost of the GPU hardware, and the question of whether an operator or an AI customer pays for it, remains unsettled.
What Nokia is really selling is a hedge on the shape of 6G. If the next generation of mobile networks is built to mix connectivity and computation, equipment makers that supply both gain an advantage over those that supply radios alone. Nokia’s software is designed to run on cloud infrastructure and commercial off-the-shelf servers, letting operators keep existing baseband gear and add AI capacity through an upgrade rather than a rip-and-replace.
The stakes are larger than one product line. Nokia has spent years reworking its business after losing ground in the smartphone era, and its network equipment unit competes against larger, better-funded rivals. A technology that gets operators to buy more hardware, and to buy it from Nokia, would be worth more than the margin on a single contract.
There is also the question of what the AI demand does to the network itself. If carriers start renting out their radio sites as compute nodes, they take on customers whose needs may conflict with those of ordinary subscribers. A model training on a tower’s GPUs could compete for the same power and cooling budget that keeps a town’s phones working during a storm. Nobody has yet said how that trade-off would be governed.
For now, the deliveries mark a shift from demonstration to product. Nokia spent the past two years showing that AI-RAN was technically possible. Shipping it puts a harder test in front of the company: whether operators are willing to pay for the version that exists today, and whether the savings and revenue it promises arrive before the next generation of equipment makes the point moot. The answer will come one cell site at a time.
