Japan's telecommunications giant KDDI is testing a novel approach to one of AI's thorniest infrastructure problems: how to power and compute AI workloads when data centers and GPUs are scarce. In September 2026, KDDI partnered with software company Morgenrot and regional network provider TOHKnet to launch a distributed data center demonstration that connects multiple facilities across Japan using advanced networking technology, allowing AI training and inference tasks to be split across geographically distant locations.

Why Are Data Centers Running Out of Computing Power?

The rise of physical AI, which powers robots and autonomous systems in retail stores and manufacturing plants, has created an urgent infrastructure crunch. Unlike traditional cloud AI that can tolerate delays, physical AI requires near-instant responses from computing systems. This means robots typically need access to GPUs and processing power located in nearby data centers or telecommunications facilities. As the number of AI-equipped robots grows, nearby computing resources and power capacity are becoming bottlenecks.

The challenge is particularly acute in Japan, where regional areas lack the concentrated computing infrastructure of major urban centers. KDDI's demonstration addresses this by creating what the company calls "Watt-Bit Collaboration," a system that coordinates power supply and demand with telecommunications networks to distribute AI workloads intelligently across multiple sites.

How Does This Distributed Data Center Model Work?

The demonstration connects KDDI's Osaka Sakai Data Center and Tama Network Center through a specialized network called an Advanced Private Network (APN). Within the Tama facility, engineers built a simulated retail environment to test how retail robots perform when their AI processing is handled by different data centers at varying distances.

The key innovation is a "point-to-multipoint" APN configuration, which allows a single location (like a retail store) to connect to multiple data centers simultaneously. This enables the system to route AI training and inference tasks to whichever data center has available GPU capacity and power at any given moment. Morgenrot developed a GPU virtualization platform that treats GPUs scattered across multiple locations as a single integrated computing resource, allowing the system to automatically balance workloads.

Steps to Implement Distributed AI Infrastructure

• Network Architecture: Deploy point-to-multipoint APN connections that link user sites to multiple data centers, enabling low-latency communication across geographically distributed facilities without requiring direct connections to each center.
• GPU Virtualization: Implement a unified management platform that pools GPUs from different locations and automatically allocates computing resources based on real-time power availability and workload demands.
• Performance Validation: Test distributed training and remote inference under multiple distance scenarios, from nearby facilities within a prefecture to remote data centers hundreds of kilometers away, to ensure response times meet physical AI requirements.

The demonstration is testing three specific scenarios. First, engineers are verifying that the point-to-multipoint APN can reliably handle the optical characteristics, traffic transmission, and power consumption requirements of distributed data centers. Second, they are measuring how much faster AI model training becomes when distributed across both nearby and remote facilities compared to using a single data center. Third, they are evaluating remote inference performance for retail robots under four GPU placement scenarios: embedded within the robot itself (0 kilometers away), nearby within the same prefecture (20 kilometers or more), remote within the prefecture (50 kilometers or more), and outside the prefecture (500 kilometers or more).

The demonstration has been selected for Japan's Ministry of Internal Affairs and Communications "Watt-Bit Collaboration Demonstration Project," indicating government backing for this approach to solving regional AI infrastructure gaps.

What Are the Broader Implications for AI Infrastructure?

KDDI's approach stands in contrast to the massive, centralized data center buildouts happening in other parts of the world. In the United States, SpaceX has rapidly constructed gigawatt-scale data centers in Memphis, Tennessee, and Mississippi, with Colossus 2 becoming the first gigawatt-scale facility. However, SpaceX's success required extraordinary measures: renting temporary power generators, purchasing power plants, and importing specialized equipment from overseas.

The challenge facing most companies is that traditional data center construction timelines are 12 to 24 months, but utility interconnection queues can stretch 3 to 7 years. New power generation and transmission equipment orders can take a decade to fulfill. SpaceX bypassed these bottlenecks by treating permitting as separate, parallel problems rather than sequential ones, and by renting temporary power capacity while waiting for permanent grid connections.

"There are still a number of older facilities that could potentially replicate the Colossus model. The challenge is finding sites that have the right combination of existing infrastructure, available power and the ability to secure the necessary permits. The scarcity of sites that check all those boxes has certainly increased their value," said Tim Comerford, senior vice president at site selection firm Biggins Lacy Shapiro & Co.

Tim Comerford, Senior Vice President at Biggins Lacy Shapiro & Co.

Japan's distributed model offers a different solution: rather than building massive new facilities, it leverages existing regional data centers and telecommunications infrastructure, connecting them through advanced networking. This approach could be particularly valuable in countries with aging industrial infrastructure and regions that lack the capital or geographic space for hyperscale data center campuses.

The demonstration is planned to expand into the Tohoku region in northeastern Japan, with inter-carrier APN connections planned for fiscal year 2027. If successful, the model could help address labor shortages by enabling more widespread deployment of physical AI robots in retail and manufacturing, while also making more efficient use of existing power and computing resources scattered across regional areas.

The infrastructure challenges facing AI are not purely technical; they are also regulatory and geographic. As Otto Lynch, vice president and head of power line systems at infrastructure software provider Bentley Systems, noted, permitting is the primary bottleneck preventing better grid interconnections. "Our grid in the U.S. is very old. Like our highways, our electric grid has a lot of 'potholes' and congestion," Lynch explained. "Permitting is the main reason we don't have better grid interconnections. It takes much longer to permit power lines than to design, procure and build them, often more than 10 years".

"Our grid in the U.S. is very old. Like our highways, our electric grid has a lot of 'potholes' and congestion. Permitting is the main reason we don't have better grid interconnections. It takes much longer to permit power lines than to design, procure and build them, often more than 10 years," explained Otto Lynch.

Otto Lynch, Vice President and Head of Power Line Systems at Bentley Systems

KDDI's distributed approach sidesteps some of these permitting challenges by working within existing infrastructure frameworks. Rather than waiting for new power plants or grid connections, the system optimizes the use of power and computing resources already available in different regions. This could prove to be a more practical path forward for countries and regions that cannot replicate SpaceX's aggressive, capital-intensive buildout strategy.