KDDI, Morgenrot and Tohoku Network start distributed data-center demonstration for Physical AI
DataNews Editorial Team
KDDI, Morgenrot and Tohoku Network have announced the September start of a distributed data-center demonstration using retail robots for Physical AI workloads. The project links KDDI’s Osaka Sakai Data Center and Tama Network Center over an All-Photonics Network (APN) to test distributed training and remote inference on virtualized GPU resources.
- The initiative is a demonstration project, not a commercial distributed data-center deployment.
- Osaka Sakai Data Center and Tama Network Center are connected through APN infrastructure.
- The trial combines point-to-multipoint APN with GPU virtualization developed by Morgenrot.
- Remote inference will be compared across four GPU-placement scenarios, ranging from 0 km to locations from 500 km away.
- The proposed APN expansion to the Tohoku region has not been confirmed as completed.
KDDI, Morgenrot and Tohoku Network have started a demonstration of a distributed data-center architecture for Physical AI, using retail robots as the workload. The companies announced that the demonstration began in September and said it was selected under Japan’s Ministry of Internal Affairs and Communications “Watt-Bit” coordination demonstration programme.
The deployment connects KDDI’s Osaka Sakai Data Center and Tama Network Center through an All-Photonics Network. At Tama Network Center, the participants have established a simulated network linking data-center infrastructure with user sites using a point-to-multipoint APN configuration. GPU resources at the two facilities are pooled through a GPU-virtualization platform developed by Morgenrot.
The trial compares robot-training approaches using a nearby data center, a remote data center and both sites in a distributed configuration. It also evaluates remote inference for retail-robot tasks across four GPU-placement scenarios: GPU embedded in the terminal at 0 km; an in-prefecture nearby location from 20 km; an in-prefecture remote location from 50 km; and an out-of-prefecture location from 500 km.
For operators and AI-infrastructure planners, the work is relevant as a test of whether low-latency optical networking and GPU virtualization can make geographically remote compute resources usable for workloads that would otherwise need to run at the closest facility. If the trial confirms the required operating characteristics, the model could improve flexibility in AI workload placement and reduce dependence on a single local site where power or GPU availability is constrained.
The companies intend to extend the APN segment used in the demonstration to the Tohoku region. That extension remains a stated intention rather than a completed deployment.
The demonstration tests whether geographically distributed GPU capacity can support latency-sensitive Physical AI workloads while providing greater flexibility to shift compute away from locally constrained sites. Its relevance to data-center infrastructure depends on whether the optical network can deliver the required performance and energy efficiency; no test results have been disclosed.
- Impress Watch — クラウド&データセンター / September 25, 2026 · JapaneseSource →
This article was prepared by the DataNews editorial team based on the sources listed above.