Real-time AI RAN tuning lifted peak downlink throughput 31% network-wide across Tokyo
In sum – what we know:
- Live commercial deployment – Samsung’s RAN Speed Optimizer ran on KDDI’s live 5G SA network across hundreds of cells in greater Tokyo, not a lab demo.
- Double-digit speed gains – Peak-hour downlink throughput rose an average 31% network-wide and as much as 52% in dense urban areas, per the companies’ own figures.
- Unanswered questions – No data on uplink, latency, power draw, or multi-vendor behavior — every disclosed metric is downlink throughput.
Samsung and KDDI have completed a months-long trial of Samsung’s AI-powered RAN Speed Optimizer (RSO) on KDDI’s live commercial 5G Standalone (SA) network, and the two companies are calling it a success. The trial began in late 2025 and stretched across the greater Tokyo region, taking in dense urban, suburban, and rural areas. It covered hundreds of 5G cells, running on 100 MHz of TDD spectrum in the 3.7 GHz band. In other words, this was a genuine field deployment on a network carrying real customer traffic, not a controlled lab demo.
RSO’s job is to tune radio parameters on a cell-by-cell basis, in real time, as demand shifts throughout the day. That’s work that has traditionally fallen to skilled RAN engineers — the never-finished business of tweaking each cell by hand, which is slow, expensive, and particularly punishing in dynamic urban environments. Samsung says the system was trained on real network data spanning diverse traffic patterns rather than limited lab simulations, which matters. Models tuned in sanitized conditions have a habit of struggling when they meet actual users.
There’s history behind this, too. In 2022, Samsung and KDDI ran live network slicing trials on the same 5G SA network using a shared RAN Intelligent Controller (RIC), proving out SLA-backed slices on commercial infrastructure. The RSO trial is a direct evolution of that work, pushing the same software-defined approach to RAN control from slicing into everyday optimization.
Performance
Across the full trial area, Samsung and KDDI report an average 31% increase in 5G downlink throughput during peak hours, climbing to as much as 52% in dense urban environments. Those figures come from the companies’ own measurements — but taken at face value, they represent the kind of gain users would actually notice during evening congestion, not a rounding error.
Cell-by-cell optimization has historically meant drive testing, manual parameter changes, and long tuning cycles. RSO compresses those cycles and cuts much of the manual fieldwork out entirely. Kazuhiro Furuhata, KDDI’s Chief Network Officer, put the emphasis squarely on the per-cell piece, saying the trial proves “individual tuning for cells — a long-standing industry challenge — has now become a reality through the integration of AI.”
Samsung, for its part, has been testing and training RSO in the field since 2024, and the KDDI deployment sits alongside parallel AI-RAN work with KT in Korea, where its optimization learns from user movement paths and usage patterns.
Limitations
Of course, it’s worth noting that every disclosed metric is a downlink throughput figure. There’s no data on uplink performance, latency, handover failures, or call drop rates — dimensions that matter just as much to real-world experience as raw download speed. Nor is there any indication of how often the 52% peak actually occurred, or how suburban and rural cells fared beyond the blended average.
Samsung also hasn’t said whether RSO runs as an xApp or rApp on a RIC, as the companies’ 2022 slicing work did, or as a separate optimization engine bolted onto the RAN. And, hasn’t addressed how the system behaves in multi-vendor environments — a real question for any operator that isn’t running Samsung radios end to end. And despite how frequently AI RAN optimization gets pitched as an energy-saving technology, neither company has quantified any change in power consumption.
Samsung isn’t alone in this space, either. Nokia, Ericsson, and Huawei all have comparable AI-based RAN optimization offerings, along with a handful of specialized software vendors. What sets this trial apart is the setting — a fully commercial SA network at meaningful scale, with published performance figures.
The risks
Handing an AI system sustained control over live radio parameters raises questions around reliability. Model drift is one. A system trained on late-2025 traffic patterns will eventually face device mixes and usage behaviors it has never seen, and there’s no word on how RSO handles that. Bias is another. An optimizer rewarded for aggregate throughput could learn to favor dense, high-traffic cells over quieter rural ones, quietly steering quality toward the areas that flatter the average.
Then there’s the data itself. An optimizer that learns from live network behavior is, by definition, processing large volumes of user traffic metadata, and neither company has said what safeguards keep that processing compliant with Japanese privacy regulations. Nothing suggests a problem — but nothing rules one out either.
None of this diminishes what the trial means strategically. Quantified gains on a tier-one operator’s live network bolster Samsung’s reputation as an enterprise software and AI provider rather than a hardware maker alone. And, both companies are plainly looking past 5G. They plan to keep evaluating AI-based optimization for broader commercial use, with an eye toward 6G, where cell density and service diversity will make manual tuning physically impractical.
Table of Contents
