Nokia's October 21 AI-RAN announcement brings together radio processing, AI applications at network sites and software agents for network operations. These are separate capabilities with different evidence behind them. Operator reports range from a measured lab result to a field demonstration and a planned 2027 trial, which makes the stage of each claim as important as the shared AI-RAN name.

The big change

  • What changed: Nokia and NVIDIA described an AI-RAN platform that runs Nokia 5G software on NVIDIA Aerial RAN Computer, with AI inside radio functions, inference workloads on shared infrastructure and a separate platform for governed network agents.
  • Why it matters: A buyer can now compare specific operator milestones instead of treating every AI-RAN announcement as evidence of a commercial service.
  • What to watch: Deutsche Telekom's field validation, BT's planned Birmingham sites and Elisa's planned distributed rollout will test claims that today's demonstrations cannot settle.

Three uses of AI, three different questions

The first use is AI within radio processing. Nokia says its anyRAN software on NVIDIA's Aerial RAN Computer can apply AI to channel estimation, beamforming, multi-user MIMO pairing and scheduling. Channel estimation helps the receiver account for a changing radio path. An AI model doing that work is part of the radio signal chain; an application running near a cell site is a different workload. Nokia's announcement describes the broader set of radio functions, but its cited operator measurement concerns one function.

That measurement comes from Deutsche Telekom's October 21 report. In labs in Bonn and Oulu, its test of AI-based channel estimation for the Physical Uplink Shared Channel showed typical uplink throughput improvements of 10–15% under the tested conditions. The work combined Nokia commercial 5G RAN software with NVIDIA Aerial RAN Computer. Deutsche Telekom says the gain increased with more complex radio conditions and plans a field trial next. The percentage is a lab result for uplink channel estimation, not a measured gain across a commercial mobile network.

The second use is running AI applications on infrastructure that also carries radio workloads. Nokia says Elisa ran voice, mobile data and video analytics simultaneously on one AI-RAN platform in a Helsinki 5G field environment. NVIDIA describes this as an over-the-air field trial using 5G spectrum. It is a demonstration of concurrent workloads, while Nokia places distributed deployment and evaluation of commercial services in 2027. Elisa's October 21 interim report confirms that its Nokia and NVIDIA collaboration continues, but gives no independent performance measurement for this field trial.

Orange Business offers another view of the shared-compute idea. Its October 21 release describes a distributed voice AI pilot using central cloud and regional resources, including NVIDIA Aerial RAN Computer at distributed locations. Orange separately says it is assessing AI-RAN with Nokia and NVIDIA for possible radio efficiency and shared-infrastructure benefits. The voice pilot does not establish those radio gains or a generally available enterprise service.

The third use is agents for operating networks. Nokia says it is integrating NVIDIA AI Enterprise software and Nemotron models and libraries into an agentic platform spanning radio, core, transport and cloud. It describes operator oversight, testing and execution guardrails, plus choices about where data, models and policies reside. Those are Nokia's architecture and governance claims. The release names operators working on use cases and integration timelines, but offers no measured result showing unattended agents safely running a live network.

The operator milestones to compare

NVIDIA reports that T-Mobile completed an AI-RAN call on a live commercial network in Bellevue, Washington, using Nokia software, commercial Aerial RAN Computer servers, an existing 5G standalone core and a commercial smartphone. A completed call demonstrates a connection through that setup; the release provides no network-wide performance or service-availability result from it.

BT says it plans to bring its first AI-RAN sites online on the live EE network in Birmingham in 2027. The operator intends to measure capacity, reliability and spectral efficiency there. BT cites laboratory platform results, but explicitly frames the planned live trial as the test of whether controlled results translate to customer traffic. The 2027 sites are a plan, not an operating deployment today.

For network buyers, the common platform has a concrete dependency: Nokia's 5G anyRAN software runs with NVIDIA Aerial RAN Computer and accelerated compute. NVIDIA announced MGX 2U and 4U configurations for higher-capacity sites and says server partners will offer ARC-ready systems. The public releases do not give a common deployment bill of materials, operator access terms, prices or a comparable field benchmark across these programs. A procurement decision therefore still needs the exact site configuration, radio performance under local conditions, capacity available to AI applications and the operator's control of data and agent actions.

Sources & further reading