IN Brief:
- Nokia and SoftBank will test AI-RAN in a Japanese commercial network environment during 2027, using Nokia anyRAN and SoftBank AITRAS.
- The computing platform incorporates NVIDIA Grace, RTX PRO 4500 and ConnectX-8 technologies for accelerated radio and AI processing.
- SoftBank's AITRAS Orchestrator and Nokia MantaRay will support coordinated resource management; no completed field-trial results have been reported.
Nokia and SoftBank have agreed to conduct an AI-native radio access network field trial in Japan during 2027, integrating Nokia’s anyRAN software with SoftBank’s AITRAS computing platform. The project will examine how commercial mobile radio functions and artificial intelligence workloads can operate on shared accelerated computing hardware while maintaining the timing and reliability requirements of a live network.
The agreement, announced on 20 October, extends an existing collaboration between the companies and establishes a specific commercial-network validation programme. The trial will use NVIDIA Aerial RAN Computer technology together with SoftBank’s AITRAS Orchestrator, allowing the partners to assess resource management and interoperability under commercial traffic conditions rather than relying exclusively on laboratory testing.
Nokia’s anyRAN software provides the radio processing component of the proposed architecture, while AITRAS supplies the computing platform on which radio and AI functions will operate. The integration will be examined alongside the orchestration software responsible for coordinating those workloads.
The accelerated computing configuration includes an NVIDIA Grace CPU, RTX PRO 4500 GPU and ConnectX-8 networking. These components provide general-purpose processing, parallel computation and high-bandwidth connectivity, although their combined performance within the specific trial has not yet been measured publicly.
A radio access network must repeatedly perform signal-processing operations according to the timing requirements of the cellular system. Functions such as modulation, coding, channel estimation and scheduling influence how information passes between the network and connected devices, while the available processing time is constrained by the communications protocol.
Traditional radio architectures have used specialised computing arrangements to meet those requirements. Cloud RAN approaches move selected functions into software operating on more general computing infrastructure, potentially increasing deployment flexibility while introducing additional requirements for workload isolation, networking and predictable execution.
Accelerated computing can support the parallel calculations required by radio processing, but the availability of powerful processors does not independently establish that a system will meet every radio deadline. The complete configuration must account for processing latency, communication overhead, memory access and the interaction between software components.
The integration of AI workloads adds another source of demand. An inference application may consume changing amounts of GPU capacity according to the model being executed and the quantity of incoming data. Radio processing, by comparison, must continue meeting its communication deadlines even when additional computing tasks are active.
Shared infrastructure therefore requires controls over how processing resources are allocated. The available CPU, GPU, memory and networking capacity must be managed so that an increase in inference demand does not interfere with essential radio functions. Conversely, additional AI applications need sufficient resources to provide useful performance when the radio workload is high.
SoftBank’s AITRAS Orchestrator will be responsible for coordinating AI and radio workloads within the planned trial environment. The companies are also integrating Nokia MantaRay with that orchestration layer, linking the management of computing resources to the operational requirements of the network.
The effectiveness of this approach will depend on the information exchanged between the management systems and their ability to respond to changes in demand. Commercial mobile traffic varies according to location, user activity and network conditions, while AI processing demand can change independently of those factors.
Resource allocation decisions must consequently consider the capacity required by each function and the constraints associated with changing its execution environment. Moving or resizing a workload may be straightforward from a general cloud computing perspective but still require careful management when that workload performs time-sensitive radio processing.
The trial will also examine interoperability between Nokia software and SoftBank’s AITRAS platform. The two technologies have been developed through separate programmes, making interface compatibility, monitoring and configuration management important aspects of the integration.
A successful connection between software components is only the first part of this validation. Sustained operation under commercial traffic must establish how the arrangement behaves as radio demand changes, including whether additional AI activity affects service quality or the availability of network resources.
The companies have not disclosed a specific deployment site, trial start date, number of radio cells or the AI applications that will be evaluated. Nor have they published a measured throughput improvement, power saving or resource utilisation figure for the planned Japanese installation.
The SoftBank programme is separate from Nokia’s other AI-RAN activities announced around GTC Berlin, including laboratory work with Deutsche Telekom and live-network demonstrations with Elisa. Those developments provide different forms of technical evidence but do not establish the outcome of the future SoftBank trial.
SoftBank’s wider ambition is to use distributed communications infrastructure as a platform for additional computing services. Locating AI processing within network infrastructure could support applications involving local data analysis and responsive interaction between connected systems, although the benefit would depend on where computation takes place and how information is transferred.
Industrial automation, robotics and other applications requiring closely coordinated computing and communications have been identified as possible longer-term uses. Each would require its own evaluation of latency, reliability and operational controls before the shared infrastructure could be treated as suitable for that application.
The 2027 field trial will assess whether Nokia’s radio software, NVIDIA accelerated computing and SoftBank’s orchestration architecture can operate together under commercial conditions. Its findings will provide evidence about the practical limits of sharing computing resources between mobile connectivity and additional AI functions, ahead of any wider deployment decision.
