AI Network Demands and Uses Still Taking Shape

Carriers have seen a rising curve in data demand since LTE, when video streaming became possible on smartphones, Hatt said. No one knows yet how much AI will affect data traffic, he said, noting that the wireless industry also needs to understand the difference between direct AI traffic, from an AI-native app like Gemini or ChatGPT, and indirect traffic. “Indirect means activity you might be doing anyway that’s augmented by AI -- could be video, could be gaming.”

Operators will have to manage that traffic but also want to monetize it, Hatt said. GSMA is working with technology providers and carriers to get a better understanding of the value of “edge” AI, he said. “A lot of operators are looking at new revenue, leveraging GPU [graphics processing unit] capabilities,” he added. “We’re seeing a lot of interesting experimentation in Korea, the U.S. and parts of Europe, and we expect that to expand.”

Guy Turgeon, senior principal industry specialist at Red Hat, said carriers have been using “predictive AI” in their networks for years. With 5G and the open radio access network, “the RAN has become more software defined,” which allows for Generative AI and greater network automation.

Carriers are embracing AI for power savings and finding ways to cut the cost of the network, said Rob Hughes, head of wireless marketing at Fujitsu. “Those are fairly easy solutions to implement that don’t necessarily cost a lot and require big changes to the network,” he said. “That’s what’s moving the fastest.”

Other providers are starting to put RAN software and AI applications on the same hardware, Hughes said. “We’ve got some pioneers that are moving in that direction,” including Softbank and T-Mobile.

UScellular views AI within the RAN architecture as still “very early” in development, said Michael Irizarry, its chief technology officer for engineering and information technology. Those kinds of changes will happen in 6G or the late stages of 5G, he said.

AI on top of existing network architecture is “moving along briskly,” Irizarry said. “From our perspective, we’re applying it very judiciously” to analyze handoff and network security logs. “We’re also using it for site selection, to be more precise,” he said. In addition, UScellular is piloting an AI platform to optimize power consumption by putting cellsite components to sleep when they’re not in use, he said. “Things are moving along nicely but at different paces” for various uses of AI.

Qualcomm Technologies is working with customers on a “bottom-up analysis” of what network engineers do daily, said Ofir Zemer, vice president of product management. Carriers are using AI to improve network performance, but that doesn’t always translate into lower operating costs, he added.

'Army' of Engineers

The bigger carriers have “an army” of engineers “crunching” key performance indicators from systems “all day long,” Zemer said. Carriers want AI “to take that load away,” he said, because even a 10% savings in operating costs is worth billions of dollars to the larger companies.

AI in RANs is moving from “concept” to “experimentation” in the real world, said Alex Jinsung Choi, chair of the AI-RAN Alliance, which has grown quickly to 94 members in 18 countries. This year’s Mobile World Congress in Barcelona featured 10 AI technology demos, said Choi, also a research fellow at Japan’s Softbank.

While carriers are following different strategies, networks are evolving to use a “hybrid” computing platform combining GPUs and central processing units (CPUs), Choi said, noting that both have different strengths.

CPUs are “well-suited” to handling the control of the control plane, Choi said, while GPUs “make a lot of sense” to handle complicated tasks like 5G signal processing or massive multiple-input, multiple-output beamforming, which require complex computations. “We will very likely see hybrid CPU and GPU platforms emerging as the dominant architecture” for the RAN, he said. There’s “no one-size-fits-all solution.”