―Addressing Power and GPU Shortages in the AI Era Through Point-to-Multipoint APN and GPU Virtualization―
KDDI Corporation
Morgenrot Inc.
TOHKnet Co.,Inc.
KDDI Corporation (Head Office: Minato-ku, Tokyo; President and CEO: Hiromichi Matsuda; hereinafter "KDDI"), Morgenrot Inc. (Head Office: Chiyoda-ku, Tokyo; President and CEO: Masamichi Nakamura; hereinafter "Morgenrot"), and TOHKnet Co.,Inc. (Head Office: Sendai City, Miyagi Prefecture; President: Atsushi Ito; hereinafter "TOHKnet") began a distributed data center demonstration (hereinafter, the "Demonstration") utilizing retail robots in September 2026, with the aim of addressing power and GPU shortages in the AI era.
The Demonstration has been selected for the Ministry of Internal Affairs and Communications' "Watt-Bit Collaboration Demonstration Project" (
Note 1).
In this Demonstration, KDDI's Osaka Sakai Data Center and Tama Network Center are connected through APN. In addition, a network that simulates connections between data centers and user sites are constructed within the Tama Network Center using a Point-to-Multipoint APN (
Note 2). By combining GPUs installed at each center with the GPU virtualization platform developed and built by Morgenrot, distributed training and remote inference of physical AI using retail robots are conducted. Through this initiative, the optimal utilization method and effectiveness of computing resources distributed across multiple locations are verified.

KDDI promotes the "Digital Belt Initiative," under which assets such as data centers and submarine cables are integrated with technologies such as APN and 6G to build a nationwide low-latency network and AI computing infrastructure spanning land, sea, and air in an AI-centric society.
Going forward, the three companies will expand the APN connectivity area in this Demonstration to the Tohoku region and promote the creation of an environment that enables the effective utilization of computing resources and power available in regional areas. Furthermore, through the development of distributed data centers, they will accelerate the social implementation of Physical AI and help address labor shortages.
■Background
In recent years, with the growing adoption of AI, the social implementation of Physical AI has been advancing, particularly in industries such as retail and manufacturing. Because physical AI requires low-latency and stable processing, it is common to utilize computing infrastructure located in data centers and telecommunications facilities close to customer sites.
On the other hand, as the number of robots equipped with physical AI increases in the future, shortages of computing resources and power supply capacity at nearby data centers and telecommunications facilities are expected to become a challenge.
To address these challenges, achievement of "Watt-Bit Collaboration," which coordinates power supply and demand with telecommunications networks, is expected. To distribute physical AI processing to geographically remote data centers, low-latency and high-capacity communications are required. In addition, mechanisms are needed to flexibly switch processing destinations according to power supply-demand conditions and computing workloads while utilizing computing resources distributed across multiple sites as an integrated resource.
To achieve these mechanisms, APN providing low latency, high capacity, and low power consumption is important in ensuring the response performance required for physical AI even when remote data centers are utilized. Furthermore, technologies such as point-to-multipoint APN, which enables connections from a single site to multiple sites, and GPU virtualization platforms that allow integrated utilization of distributed GPU resources are also required.
■About the Demonstration
1. Overview
- (1) Verification of point-to-multipoint APN for use in distributed data centers
Using a Point-to-Multipoint APN, network simulating connections between data centers and user sites are built within the Tama Network Center. The optical characteristics, traffic transmission performance, operational management functions, and power consumption of the Point-to-Multipoint APN are verified.
- (2) Verification of the effectiveness of distributed training using retail robots
At a demonstration field within the Tama Network Center that simulates a retail store, training data related to robot operations are collected. The reduction in training time are evaluated for:
- 1Training at a single nearby data center
- 2Training at a single remote data center (Osaka Sakai Data Center)
- 3Distributed training using both nearby and remote data centers

<Illustration of Distributed Training Effect Verification>
- (3) Verification of the Effectiveness of Remote Inference Using Retail Robots
Remote inference processing for retail robot operation is comparatively evaluated under four GPU placement scenarios: embedded within the device (0 km), nearby within the prefecture (20 km or more), remote within the prefecture (50 km or more), and outside the prefecture (500 km or more).

<Illustration of remote inference performance evaluation>
2. Roles of each company
KDDI
- Provision of the field verification environment
- Construction of the edge data center environment
- Remote inference and distributed training of physical AI using retail robots
Morgenrot
- Development of the GPU virtualization platform
- Execution of distributed GPU processing
TOHKnet
- Study of inter-area connectivity for inter-carrier APN connection demonstrations planned for FY2027
- Joint demonstration of point-to-multipoint APN
(Reference)
■About Morgenrot
Morgenrot provides cloud-based GPU resources and a unified management platform with virtualization capabilities. In this demonstration project, which targets multi-site connectivity via Point-to-Multipoint APN and the practical application of physical AI, Morgenrot is responsible for developing and building the GPU virtualization infrastructure and executing distributed GPU processing. This will enable flexible management of computing resources that accounts for training and inference execution pipelines, contributing to the social implementation of distributed data centers.
■About TOHKnet
As a company responsible for the development and advancement of information and telecommunications infrastructure across the six prefectures of the Tohoku region and Niigata Prefecture, TOHKnet verifies the potential utilization of regionally distributed data centers through participation in this Demonstration. By realizing high-speed and low-latency wide-area connectivity using APN, TOHKnet promotes the creation of an environment where AI computing resources and electric power available in regional areas can be effectively utilized. Looking ahead, TOHKnet aims to contribute to the development of AI and digital infrastructure extending from the six Tohoku prefectures and Niigata Prefecture to the entire country, thereby helping realize a sustainable regional society.
- (Note 1)
- (Note 2)A network configuration that connects a single site to multiple sites via APN by utilizing Open XR Optics technology and optical splitters, enabling the transmission and identification of signals for multiple destinations within a single optical signal.
- *The information contained in the articles is current at the time of publication.
Products, service fees, service content and specifications, contact information, and other details are subject to change without notice.
Downloads
