AI Moves into Commercial RAN, with Spectrum Efficiency as an Early Economic Driver

On August 20, 2026, SoftBank and Ericsson announced the first Japanese trial of Ericsson's Artificial Intelligence (AI) in Radio Access Network (RAN) technology on a commercial 5G network. The companies tested an AI-native Link Adaptation Scheduler which operates directly within the RAN and dynamically adjusts downlink transmission settings according to changing radio conditions and traffic. The trial evaluated spectral efficiency, user throughput, robustness, and stability across SoftBank's commercial network.

The results showed up to approximately 25% improvement in spectral efficiency and up to 50% higher downlink user throughput compared with conventional technology. Across all evaluated locations, both measures improved by approximately 10% on average. The results indicate that AI can make real-time decisions on modulation and coding, allowing radio resources to be used more efficiently than with conventional algorithms.

The trial is significant because it moves AI in the RAN from controlled demonstrations toward operation within a live commercial network. SoftBank and Ericsson intend to continue developing the technology for 5G-Advanced and future 6G networks, making the deployment an early indication of how AI-native radio functions could be introduced incrementally into existing network infrastructure.

The commercial importance of the trial extends beyond performance improvements. Spectrum is among the most limited and costly resources available to mobile operators, and acquiring additional spectrum or densifying networks requires substantial investment. If AI can consistently improve the amount of traffic that operators can carry within existing spectrum, it provides a way to increase network capacity without relying entirely on additional infrastructure.

This creates a different economic proposition for AI-RAN from the external AI compute model being pursued by SoftBank and Nokia. Rather than requiring operators to find third-party demand for AI processing, AI for RAN generates value from an asset the operator already owns and operates. The immediate return would instead come from improved network economics: more capacity from existing spectrum and potentially deferred investment in additional sites or spectrum. The scale of the benefit will depend heavily on network conditions. The maximum 25% spectral-efficiency improvement should not be treated as a universal gain across commercial networks, particularly because the average improvement reported across the evaluated locations was approximately 10%. Network density, radio conditions, traffic patterns, and the existing level of optimization will all influence the achievable benefit.

AI-RAN also poses risks that must be weighed against the performance advantages. Operators may become too dependent on vendor-specific AI models, software, and RAN functionality, creating vendor lock-in and making scalability increasingly difficult. There are also additional costs and complexity associated with running AI-based functionality. All these aspects can limit the overall economic value, especially if there is not enough performance gain.

Taken together, the results support an evolutionary approach to AI-RAN rather than a wholesale replacement of existing RAN infrastructure. Operators can introduce AI-native functions into existing 5G networks and assess their performance before making broader architectural changes. This lowers the barrier to AI adoption as networks transition toward 5G-Advanced and, eventually, 6G.

AI-RAN could therefore develop through several commercial stages. Spectral-efficiency and throughput improvements can provide an initial business case, while autonomous optimization, energy management and external AI workloads create additional opportunities as compute and orchestration capabilities mature. For operators, this means AI-RAN economics does not necessarily depend on generating new AI-service revenue from the outset.

Operators should initially deploy AI-native RAN functions in high-utilization locations where additional spectrum or physical densification would otherwise be required. Urban hotspots, major transport hubs, and high-traffic enterprise locations should be prioritized because even relatively modest improvements in spectral efficiency can translate into additional usable capacity. Operators should set up site-level baselines for spectral efficiency, throughput, and energy consumption before deployment so they can measure the incremental value of AI against conventional RAN optimization.

Vendors should make AI-RAN functions deployable as modular software upgrades rather than requiring wholesale RAN replacement. The SoftBank trial demonstrates the value of introducing AI into an existing commercial 5G environment, giving operators a lower-risk route to test performance before expanding deployment. AI-native schedulers and other RAN functions should therefore support existing infrastructure and provide clear performance thresholds, such as minimum spectral-efficiency or throughput improvements, to determine when to expand deployment.

Operators should also link AI-RAN investment directly to avoid network expenditure. Improvements in spectral efficiency should be translated into metrics such as additional traffic supported per site, delayed capacity upgrades, and traffic carried per Megahertz (MHz) of spectrum. This would give network planners a concrete basis for comparing AI-RAN investment against adding new sites, acquiring additional spectrum, or upgrading existing radio equipment, turning AI-RAN from a technology initiative into a measurable infrastructure investment decision.