SoftBank has begun testing an AI-based scheduler from Ericsson on its live 5G network in Japan, putting machine-learning radio resource management in front of production subscribers for the first time in the partnership.

The trial, first reported by Telecoms Tech News, places the scheduler inside the operator's commercial radio access network rather than a sandbox. Scheduling determines how a base station divides time, frequency and beam resources among users every millisecond; moving that function from rule-based logic to a learned model is one of the more consequential AI deployments inside the radio access network, because a misallocation shows up immediately as throughput loss or a latency spike for subscribers.

What does the scheduler actually change?

Classical 5G schedulers allocate resource blocks using channel quality reports, buffer status and quality-of-service tags. The model Ericsson and SoftBank are testing instead attempts to learn the same allocation decisions from historical traffic, predicting short-term demand at the cell level and adjusting the grants accordingly.

That work tracks the broader 5G Advanced direction the 3GPP set in motion with Release 18, which started a formal study on AI management of the radio interface. Release 19 extends the discussion toward models that interoperate with classical schedulers during the transition window.

Where SoftBank fits in Japan

SoftBank is one of Japan's four mobile carriers, alongside NTT DoCoMo, KDDI and Rakuten Mobile. The operator launched commercial 5G in March 2020 in non-standalone mode on the 3.7 GHz and 28 GHz bands, later adding 700 MHz for coverage, and has built its footprint primarily through macro cells and street-level small cells.

Japan's 5G subscriber base remains the smallest among the major Asian markets, but the country's operators have prioritised capacity and experience over the rapid subscriber growth seen in South Korea or China. That posture — and SoftBank's willingness to host vendor research on production cells — makes the operator a recurring partner for RAN experimentation.

Why Ericsson needs this

Ericsson has spent the past three years positioning itself as the most AI-forward of the major Western RAN vendors, embedding machine-learning features across its radio software stack and front-loading them in its 2024 and 2025 product cycles. Competing work exists at Nokia and Samsung Networks, though most public AI scheduler research still comes from academic groups and Chinese vendors.

What SoftBank offers Ericsson is footprint. Most published AI-scheduler experiments rely on simulated traffic or single-cell lab setups; SoftBank's network, with its mixed device base and live traffic profile, is the only setting in which Ericsson can demonstrate that its model generalises beyond the training set.

What happens next

Neither SoftBank nor Ericsson has disclosed a commercial release date for the scheduler, and the trial's geographic scope, cell count and traffic types remain unspecified. The first externally visible data point will likely be performance results — throughput, latency and cell-edge behaviour — published once enough traffic has crossed the live cells.

If the numbers hold, the scheduler is a candidate for inclusion in subsequent Ericsson RAN software releases and for input into the 3GPP Release 19 AI/ML work item on radio-interface management.