By Christopher Lycett, Director of Events, Informa Connect
Network X 2026 runs 13–15 October in Vienna. Register now — operators and vendors will be there to talk about what's next for AI in the RAN.
Communications service providers (CSPs) are asking hard questions about AI in the mobile RAN — the radio access network. Which use cases actually move the needle? Where should the intelligence physically sit? And how much control are operators willing to hand to an algorithm making real-time decisions on a live network?
At Network X 2025 in Paris, CSP execs on the AI-RAN: Redefining Network Opportunities panel said AI-driven anomaly detection and predictive analytics are becoming essential to extracting maximum value from network investments.
Telcos Execs Featured in This Article
Dr. Yue Wang, Chief Technologist, Network and AI, China Telecom
Bernard Bureau, VP, Wireless Strategy & Services, Telus
Takehiro Nakamura, Chief Standardisation Officer, NTT Docomo
These executives made the case in Paris and after. Telus's Bernard Bureau, VP of Wireless Strategy & Services, laid out what's already working at Network X Americas 2026. Telus can “predict how customers are perceiving network experience,” he said, using AI for anomaly detection and network observability, then pushing remedial actions straight into the network. He also described a federated AI model layered on centralized data engineering — Telus's way of scaling AI use cases without losing control of them (Rakuten Symphony, Jul 2026).
China Telecom's Dr. Yue Wang, the operator's chief technologist for network and AI, argued in mid-2026 that today's telecom architectures weren't built for AI workloads. They need restructuring, she said, around what she calls network-compute orchestration: “the ability to orchestrate the network and compute together to support the new AI services we face today.” (RCR Wireless, Jun 2026).
Some of that same debate returns at Network X 2026 in Vienna, October 13-15, where NTT Docomo's Chief Standardisation Officer Takehiro Nakamura headlines an Operator Address titled What's Actually Scaling in AI Across the RAN Today? in the Mobile Networks track (Register here).
Here's what the latest research and these telco conversations reveal about where AI in the RAN is actually paying off, and where operators are still holding back.
What does the research show about AI's impact on the RAN?
Operators are increasingly convinced AI-native RAN infrastructure is coming, even if the timeline is fuzzy. In a Heavy Reading Platforms for Future RAN Systems Operator Survey sponsored by Fujitsu, 80% said new AI and generative AI applications will demand AI-native RAN platforms, and 82% expect shared AI-and-RAN infrastructure before 6G lands around 2030 — yet only 23% expect real deployments before the end of 2026. Asked what benefits they expect, operators put improved RAN performance and spectral efficiency first (65%), followed by cost-optimized operations (58%) and enhanced coverage (48%); 54% said AI would make multi-vendor RAN deployments more likely, not less (Light Reading, Feb 2025).
What are the Top Telco Issues with AI in the RAN?
Is AI-RAN worth the capex? Operators are split: centralized AI for energy savings and fault detection already pays off; GPU-native compute at every cell site is still an open economic question.
Can operators trust AI to act autonomously? Explainability and human-in-the-loop governance remain prerequisites before AI gets more control over live networks, particularly in Europe.
Where should the intelligence live? Edge, core, or both — vendors and operators don't yet agree, and site-level economics (power, space, security) are forcing the debate.
A separate Heavy Reading survey of 115 CSPs, sponsored by VIAVI Solutions, Google Cloud, and Broadcom's VeloCloud, found operators already running AI in production single out RAN performance, especially proactive fault detection and dynamic resource allocation, as a leading payoff area, even as edge and MEC use cases lag (Light Reading, March 2025). Dell'Oro projects cumulative AI-RAN revenue will reach $35 billion between 2026 and 2030, up from an earlier $10-billion-by-2029 estimate (Light Reading; Telecoms.com). Omdia lists AI-RAN as one of three defining RAN trends to watch in 2026, alongside 5G monetization and 6G groundwork.
Where are operators finding real value from AI in the RAN today?
Per the Network X 2025 panel, the most mature AI-RAN use cases remain centralized “AI for RAN”: energy savings, capacity planning, anomaly detection, and increasingly fast root-cause analysis, driven by the complexity of managing dense, multi-band, multi-vendor networks (Network X). Some operators are pushing further: SoftBank says a Transformer-based AI model lifted 5G uplink throughput by 30%, up from 20% with its earlier CNN model, using existing hardware and a pure software upgrade — a concrete AI-for-RAN win that doesn't require a forklift replacement (Telecoms.com, August 2025). Momentum is visible industry-wide, too: Vodafone joined the AI-RAN Alliance as its 100th member in mid-2025, pledging to use AI to optimize its Open RAN deployments, even as the two-year-old alliance is still moving from membership growth to concrete deliverables (Telecoms.com, July 2025).
Will operators trust AI enough to give it more control?
Trust remains the sticking point. At that same Paris session — moderated by James Kirby of Analysys Mason, with Nokia's Ari Kynäslahti and Intel's Udayan Mukherjee joining Wang and Bureau on stage — panelists said operator confidence in autonomous, AI-driven network changes, especially in Europe, is still limited by concerns about transparency and explainability. Operators want to know why an AI model recommended an action, and what alternatives it considered, before the change touches a live network.
Human-in-the-loop governance and digital twins that let operators test AI decisions in a safe virtual environment before deployment were cited as necessary bridges. The panel's closing consensus, as recapped on the Network X news and insights hub: AI-RAN is “transitioning from experimentation to strategic necessity,” but only if operators resolve trust, distributed-compute cost, and interoperability challenges first.
How far out on the edge network should AI-RAN be located?
Orange CTO Bruno Zerbib pushed back on the dominant AI-RAN vision of GPU-equipped compute at the base of every mast. “Having like 20 or 30 or 40 [GPU sites] deployed in Paris? I don't think that makes sense,” he said, citing GPU cost, total cost of ownership, and the security risk of parking valuable compute in unprotected roadside cabinets in a market that has already dealt with copper theft.
Nvidia has since pivoted toward an “AI Grid” strategy spreading GPUs across roughly 100,000 distributed network data centers rather than concentrating them at 7 million individual RAN sites — closer to what the industry calls AI-core than AI-RAN (Light Reading, March 2026).
What's next for AI in the RAN at Network X 2026?
These questions — how much AI belongs at the edge, and how much operators will trust it with — return to the stage in Vienna. NTT Docomo's Takehiro Nakamura opens the Mobile Networks track's AI conversation with an Operator Address on what is actually scaling in AI across the RAN today, running Tuesday, 13 October, 16:00–16:20. Network X 2026 runs 13–15 October at VIECON, Vienna, drawing more than 5,500 telecom leaders across mobile, fibre, Wi-Fi, and data centre networking. View the agenda and register to hear how operators in Europe and leading CSPs around the world are answering these questions in practice.
FAQs: AI in the RAN
Where in the network is AI improving RAN performance?
AI in the RAN is driving RAN performance. Centralized “AI for RAN” use cases, including energy savings, anomaly detection, and capacity planning, are in production today, and Heavy Reading survey data shows operators expect measurable spectral-efficiency and coverage gains. GPU-native AI-RAN, where AI workloads run inside distributed units, is still mostly pilots, and some operators, including Orange, are openly skeptical it scales economically at every cell site.
What's the biggest obstacle to broader AI-RAN adoption?
Trust and cost. Operators want explainability before letting AI make autonomous changes to a live network, and the industry is still debating whether GPU compute belongs at the base of every mast or centralized in fewer, larger sites.
Where can telecom leaders discuss AI-RAN in person?
Network X 2026 in Vienna hosts an Operator Address on AI-RAN scaling with NTT Docomo's Takehiro Nakamura on Tuesday, 13 October, 16:00–16:20, in the Mobile Networks track. View the session and register to join the conversation.
About the Author
Christopher Lycett is the Event Director of Network X and has been since the event's launch in 2022.