en.Wedoany.com Reported - Samsung and KDDI conducted an AI optimization trial on a live 5G Standalone (SA) network in Japan, achieving a 31% improvement in average downlink throughput during peak hours, with gains of up to 52% in dense urban areas. The trial covered hundreds of cells in the Tokyo metropolitan area, utilizing 100 MHz of spectrum in the 3.7 GHz TDD band, encompassing various deployment conditions and traffic scenarios.
For operators, field trials may be influenced by test areas, baselines, equipment combinations, and time windows, and the actual value must be assessed based on specific operational environments. The key question is whether AI can take over the aspects of mobile networks that are still considered slow, localized, and reliant on manual intervention—namely, adjusting cells one by one without disrupting neighboring areas.
Samsung's RAN Speed Optimizer (RSO) was tested on KDDI's commercial 5G SA network over several months, starting in late 2025. The network challenges in Tokyo's dense urban areas differ significantly from those at suburban coverage edges or rural propagation zones. The trial leveraged these mixed conditions to train and validate AI models against real traffic. RSO recommends customized parameters for each cell rather than using shared settings to adjust cell clusters. If effective at scale, this could represent an operational transformation.
The mobile industry has long discussed self-optimizing networks, but most tools still require engineer intervention. Samsung positions AI optimization as a more granular and adaptive approach. Its system analyzes site environmental data, predicts the effects of parameter changes, and recommends settings tailored to each cell. RSO is part of Samsung's CognitiV Network Operations Suite, which includes AI automation tools, agents, and network applications.
Operators face pressure to enhance 5G performance with limited capital expenditure. Spectrum is expensive, site acquisition is slow, and energy costs are significant. If AI can extract more throughput from existing radio assets, it becomes a capacity strategy. However, capacity strategies have consequences: improved downlink throughput does not necessarily enhance end-to-end user experience. Factors such as latency, uplink performance, handover behavior, and congestion management remain critical. Operators need safeguards, rollback controls, and audit trails to ensure models are not optimizing narrow KPIs while creating hidden issues.
The trial was conducted on a 5G SA network, which provides a cleaner architecture for advanced features and a credible foundation for AI-driven network operations, network slicing, and granular service control. However, many operators globally still operate in hybrid environments, where legacy RAN, 4G dependencies, and vendor diversity complicate network automation. Samsung and KDDI have a long-standing collaboration on fully virtualized network deployments, which facilitated this trial, but not every operator has such a foundation.
AI-driven RAN optimization may require more than just purchasing a software module; it depends on clean telemetry, consistent configuration management, and modern operational processes. For developers and network software teams, the direction is clear: wireless networks are becoming data platforms, with value shifting toward predictive models, closed-loop automation, and policy engines. Samsung needs this success—the network equipment market is highly competitive, and AI-RAN is emerging as a new battleground.
KDDI's role is also significant. Japanese operators are often early adopters of advanced network architectures. A 31% improvement in average downlink throughput during peak hours, if replicable beyond the trial scope, holds commercial value—potentially reducing congestion complaints and improving perceived 5G quality. Samsung and KDDI stated they will continue evaluating AI-based optimization for broader commercial use, and the next phase will be more critical than the trial results.
Operators should view the 31% throughput improvement as a promising trial signal, not a guaranteed network-wide result. Baselines, geographic conditions, and operational controls determine actual value. Cell-level optimization can extract more capacity from existing RAN assets while reducing manual engineering workload. However, AI-tuned RAN parameters carry operational risks; poorly governed automation could harm latency, handovers, or customer experience. The 5G SA architecture provides a clearer foundation for automation. Infrastructure buyers should ask vendors about rollback mechanisms, explainability, and multi-vendor support.
