All three of the largest US wireless carriers are moving to bring artificial intelligence into their radio access networks, but each is betting on a different architecture and vendor path, reflecting a broader industry split over whether GPUs belong inside the radio itself.
Verizon has taken the most skeptical position on GPU-based AI-RAN. Its network leadership sees GPU processing as best placed at the network edge, where it can power applications such as augmented-reality overlays, rather than inside the radio stack, where rebuilding the chain around GPUs adds complexity without a clear payoff. CPU-based radios with embedded AI inference are sufficient to optimize performance today in Verizon’s assessment, and likely to remain so for years, even as the carrier continues expanding its virtualized “Intelligent Edge Network.” That network spans more than 22,900 virtualized RAN cell sites and over 170,000 Open RAN-capable radios, with Samsung and Qualcomm integrated into a multi-vendor RAN Intelligent Controller for AI-driven automation and energy management. Verizon has also been active on 6G research, reporting real-world integrated sensing and communication and AI-native trials through the 6G Forum this month.
T-Mobile sits at the opposite end, maintaining the closest alignment with Nokia and Nvidia’s GPU-centric AI-RAN vision. The carrier runs an AI-RAN Innovation Center in Bellevue, Washington, with delivery of Nokia’s prototype GPU-based radios expected during 2026. Nokia’s roadmap targets 50 percent higher spectral efficiency by the end of 2027 and 100 percent by the end of 2028, with early demonstrations already showing gains of 30 percent or more. T-Mobile is running a parallel, GPU-free track with Ericsson at the same time: a large-scale production trial of an AI-native scheduler with link adaptation across roughly 43 sites in Los Angeles, New York, New Jersey and Salt Lake City delivered spectral-efficiency gains of up to 10 percent and downlink throughput improvements of up to 15 percent over rule-based systems, running on standard silicon instead of GPUs. Commercialization of that feature is targeted for the third quarter of 2026, alongside the longer-term GPU bet with Nokia.
AT&T occupies a pragmatic middle ground. Its RAN technology leadership defines AI-RAN broadly, covering any use of AI in network operations from troubleshooting to optimization, rather than committing to GPUs as the default path. Multiple GPU platforms are under evaluation against a power, price and performance framework specific to RAN requirements, with GPUs slated for deployment only where they fit network economics. AT&T has demonstrated integrated sensing and communication capabilities on CPU architecture and, working primarily with Ericsson after a USD 14 billion open RAN contract award in 2023 that shifted a large share of its network away from Nokia, has also tested Ericsson’s GPU-free AI-native scheduler on dedicated hardware and Intel Xeon 6 cloud RAN systems, showing similar spectral-efficiency gains of 10 to 15 percent.
The split among the three carriers mirrors a divide between their radio vendors. Nokia has partnered closely with Nvidia to build GPU-based AI-RAN hardware aimed at roughly doubling spectrum capacity, while Ericsson continues to push AI-native software on conventional silicon as a lower-cost, hardware-agnostic alternative. For enterprise and network buyers tracking 5G Advanced and early 6G roadmaps, the divergence signals AI-RAN will not arrive as a single standard architecture in the near term. Vendor and infrastructure choices each carrier locked in over the past several years, cloud-native cores, virtualized RAN builds and open RAN contracts, are now shaping how quickly, and in what form, AI reaches the radio network.
CT Bureau
Leave a Reply