Just quickly because I’m dead-ed by this late Optus/Ericsson write-up; but there’s something cooking with AI-RAN, kind of obvious, on different fires: AI for networks versus networks for AI, or AI for RAN versus AI on RAN – quite plainly. But the first is looking serious and tangible, suddenly, while the rest is really anyone’s guess – despite a lot of fun guesswork and speculation about the likeliest inward-looking gains from ‘level-five’ AI autonomy and ‘AI-native’ 6G, and the outward-looking ones from the physical and agentic AI that will be stood-up on ‘AI grids’ someday. Anyway, today is not about that; today is about the stuff in the network – and “we are not talking about front-of-house stuff”, says Optus – speaking on Tuesday at Intelligent RAN Forum. Orange said similar on Tuesday, too.
That write-up is also linked below, including references to (up to) 25% spectral gains on the Australian firm’s Ericsson setup. But ZTE and Samsung are having the same debate (separately), it seems. ZTE says the business model for RAN-for-AI remains immature, with demand limited to niche scenarios – and that carriers should put their money where they will make it back, effectively. Samsung is pushing AI-RAN projects with KT and SK Telecom in South Korea, extending the conversation towards physical AI and AI-native networks – all that future stuff, funner to talk about but less immediately meaningful. But SK Telecom also issued a warning on the session with Optus: do not repeat the 5G-era habit of building network capabilities first and looking for applications afterwards. Remember MEC, it says; remember URLLC, it might say. We will write this up properly tomorrow or next week.
But the sensible AI-RAN outlook is to worry less about the AI-native future, and more about the value in today’s network – and then building tomorrow’s.
James Blackman
Executive Editor
RCR Wireless News
RCR Top Stories
Practical AI-RAN: AI-RAN is becoming practical with Optus reporting live-network gains in link adaptation, coverage prediction, and compensation. But scaling across multi-vendor networks remains a challenge, notes SK Telecom.
On-device AI: Qualcomm’s Snapdragon 8 Elite Gen 6 comes in standard and Extreme versions, both using 2nm silicon and 5GHz Oryon CPUs. The Extreme adds NPU memory, Matrix Cores and greater headroom for on-device agentic AI.
Networks that pay: Build for the business case, says Mimosa Networks, not for fiber-first dogma. As BEAD economics shift, FWA in unlicensed spectrum offers another route to viable networks, from rural America to urban India.
Two-track AI-RAN: ZTE sees near-term AI-RAN returns in network optimization, energy savings, and experience-based monetization, while general-purpose RAN compute remains immature. Its strategy separates the two.
Rubin review: Nvidia claims its Vera Rubin NVL72 delivers up to 30x more agentic-AI throughput per megawatt than GB300, but the result hinges on a specific model, workload and operating point, making broader efficiency claims premature.
AI-Powered Telecom Infrastructure
Supermicro, in collaboration with NVIDIA, delivers AI-powered infrastructure tailored for telcos, enhancing operational efficiency, network management, and customer experiences. Explore now
Beyond the Headlines
Cambium WISPs: Cambium’s collapse leaves WISPs with a problem with their sprawling installed bases. Airspan’s partial acquisition offers relief but leaves ePMP and other lines in limbo. The challenge is securing support, spares, and migration.
NPUs for telco AI: SK Telecom is putting Korean-made NPU chips into customer-facing services at meaningful scale. Rebellions says its ATOM-Max NPUs are powering four SK services, processing 14m AI requests and 4bn tokens daily.
Eating bandwidth: Agentic AI will consume the network. AI agents could discover, request and consume network capabilities autonomously. The issue is APIs, identity, policy, security, orchestration and potentially monetisation.
Orange AI-RAN: Orange is preparing to test Nokia’s AI-RAN platform in its live network, putting claimed radio-efficiency gains to the commercial test while exploring the bigger proposition of shared compute for AI workloads.
Testing times: AI workloads are changing network testing as synchronized traffic, larger accelerator clusters, and dynamic training and inference require system-level validation, continuous assurance and real-time network visibility.
What We're Reading
AI-RAN in Korea: Samsung will deploy AI-RAN with KT and SK Telecom in Korean industrial sites, using 5G SA private networks to test welding, patrol and monitoring robots as the operators explore physical AI.
Telco GPU access: Mavenir and Neysa are combining AI orchestration, token metering, and agent capabilities with sovereign GPU infra, targeting operators, enterprises, and neo-clouds with a production-ready platform for AI services.
Texas route miles: FiberLight is extending its West Texas network roughly 300 route miles toward El Paso, adding high-capacity fiber and duct to support data center demand and create routes into New Mexico, Arizona and Mexico.
‘Orange España’: Orange has rebranded MasOrange after acquiring the remaining 50% stake for €4.25bn. The Spanish operator is now called Orange España, while its portfolio of consumer brands remains unchanged.
Alibaba in EU: Alibaba Cloud will launch its first regions in Türkiye, Finland, and the Netherlands, and expanded elsewhere, as it adds AI tools for multi-model workloads and video generation to accelerate enterprise adoption.
