Summary

• Nokia and Nvidia are extending AI-native radio software to improve 5G signal processing and support AI applications running alongside network functions.
• Elisa has completed a live-network demonstration, Deutsche Telekom reports typical uplink throughput gains of 10–15% in laboratory tests, and BT plans a Birmingham trial in 2027.
• Nokia is also developing distributed applications and governed operational AI agents. Wider commercial deployment and performance across different live networks remain unproven.

European telecommunications operators are moving artificial intelligence radio networks towards wider deployment, with Nokia and Nvidia reporting live demonstrations, laboratory performance improvements and new testing programmes involving major mobile service providers.

Elisa has operated voice, mobile data and video analytics on a shared AI-RAN platform within its Finnish 5G network, while Deutsche Telekom has reported typical uplink throughput improvements of 10–15% during laboratory evaluations of AI-assisted radio processing.

BT intends to test the technology on a live network in Birmingham during 2027, with Orange also preparing further laboratory and field evaluations. The programmes are at different stages, and the results do not establish that the participating companies have completed nationwide commercial deployments.

Nokia outlined the developments on 21 October as it expanded its collaboration with Nvidia. The companies are developing a software-defined platform intended to support conventional mobile network functions, AI-assisted radio processing and applications that use computing resources distributed closer to customers and connected devices.

Radio processing becomes more software-driven

Mobile networks rely on radio access equipment to connect phones and other devices with the operator’s wider infrastructure. That equipment must manage changing signal conditions, interference and competing demands for available spectrum.

Existing radio systems use complex software to estimate signal conditions, coordinate antenna activity and schedule transmissions between devices and the network.

Nokia and Nvidia are introducing machine learning into selected parts of that processing, including channel estimation, beamforming, multi-user MIMO pairing and scheduling.

Channel estimation allows the network to assess the characteristics of a wireless connection so that information can be transmitted and received more effectively. More accurate estimates may help a system use available radio resources efficiently, although performance depends on the particular conditions and equipment involved.

Beamforming adjusts how signals are directed through antenna systems, while scheduling determines how transmission resources are allocated between connected devices.

These functions interact with one another. Improving one algorithm does not automatically produce an equivalent improvement in the performance of an entire mobile network, because equipment, signal conditions and other processing stages can impose additional limits.

Nokia’s AI-native anyRAN software operates with Nvidia Aerial RAN Computer infrastructure, providing a common computing environment for radio processing and selected AI workloads.

The companies want operators to run conventional network functions alongside applications that need low-latency computing close to the source of information.

However, sharing infrastructure introduces resource management requirements. AI applications cannot be allowed to consume computing capacity in ways that compromise essential mobile services or the timing constraints associated with radio processing.

Operators report different stages of progress

Elisa’s demonstration in Helsinki provides evidence that radio and AI applications can operate simultaneously on a shared platform in a live 5G environment.

The Finnish operator ran voice, mobile data and video analytics workloads together, demonstrating the proposed architecture without claiming that every customer-facing service has already entered routine commercial operation.

Elisa expects to investigate distributed deployments and additional customer applications during 2027, when it also intends to evaluate possible commercial services.

Deutsche Telekom is pursuing a more specific performance objective through the development of AI-based channel estimation.

Initial laboratory measurements showed typical improvements in uplink throughput of 10–15%, with larger gains reported under more difficult radio conditions.

These figures apply to the evaluated algorithms and test conditions, rather than proving that the operator’s complete network has achieved the same improvement.

The partners plan to examine further spectral efficiency techniques before extending validation into field environments. That work will be necessary to determine whether measured gains remain consistent across different equipment, coverage conditions and patterns of customer use.

In Britain, BT is preparing a live network trial in Birmingham during 2027. The company has not reported a completed operational trial or disclosed measured commercial results for that programme.

Orange is also preparing further work under laboratory and real network conditions.

Although Nokia describes its wider AI-RAN platform as moving towards commercial reality, the operator programmes provide distinct kinds of evidence. A live demonstration, a laboratory benchmark and an announced future trial cannot be treated as equivalent deployment milestones.

Distributed applications extend network capabilities

Nokia is developing an ecosystem of distributed applications that use information available through the radio access network.

Its E3 interface provides selected network information to applications operating at the edge, potentially supporting sensing, location services and AI processing near connected devices.

Processing information locally can reduce the need to transfer every item to a distant computing facility before a decision is made, although the practical benefit depends on network conditions, application design and the computing resources available.

DeepSig is developing spectrum sensing software that can examine radio conditions for interference, unexpected transmissions and other potential threats using existing network infrastructure.

ZaiNar is integrating positioning technology with Nokia Cloud AI-RAN, seeking to provide network-based location capabilities without requiring modifications to connected devices or radio hardware.

These applications suggest possible services beyond providing ordinary mobile connectivity. Operators could potentially supply information or computing capabilities to businesses developing industrial, transport or location-based applications.

Commercial arrangements would still need to address service reliability, accuracy, pricing and access to network information.

Location and radio data can also reveal sensitive information about connected devices and their activity. Operators must therefore control which applications can access particular information and how those records can be processed or shared.

Nokia is working with infrastructure suppliers including Dell Technologies, Quanta Cloud Technology and Supermicro on configurations supporting distributed and centralised deployment. Red Hat is also contributing cloud software intended to host radio and AI workloads on shared infrastructure.

Such partnerships may simplify system integration, but they do not remove the responsibility of individual operators to evaluate performance, security and total operating costs.

AI agents enter network operations

The companies are also developing a separate platform for AI agents that can assist with operating telecommunications networks.

Nokia is combining operational context from its network management systems with Nvidia AI Enterprise software and Nemotron models, creating a framework for agents working across radio, core, transport and cloud domains.

The proposed applications include examining network conditions, investigating faults and coordinating actions involving several infrastructure systems.

Nokia describes its approach as glass-box autonomy, using operational guardrails and oversight intended to make agent actions observable and controllable.

That distinction is important because an error in network management can have consequences extending beyond an inaccurate software response. Incorrect configuration changes may disrupt mobile services, affect routing or reduce the availability of communications infrastructure.

AI agents therefore need defined permissions, testing procedures and mechanisms for interrupting actions that would exceed their authority.

The architecture supports deployment through on-premises, cloud and hybrid environments, potentially allowing operators greater control over where sensitive operational information and software components reside.

Such flexibility can contribute to data governance and sovereignty objectives, although the actual level of control depends on implementation, contracts and the technology suppliers involved.

Nokia identifies several operators exploring advanced uses and integration timetables, but has not established that all have deployed autonomous agents controlling production networks.

For European telecommunications businesses, AI-RAN combines two potential commercial benefits: improved performance from existing radio spectrum and additional services delivered through distributed computing.

Neither benefit is automatic. Operators must determine whether the gains justify new computing equipment, software integration, energy use and the complexity of managing shared workloads.

Elisa’s demonstration and Deutsche Telekom’s laboratory results provide evidence of technical progress, while BT’s Birmingham programme will create another opportunity to evaluate the technology under operational conditions.

Wider adoption will depend on performance measured across live networks, the reliability of new applications and whether operators can support them without weakening established connectivity services.