Fukushima Containerized Data Center with a Liquid-Cooled B200 Cluster

KONST Group participated in a containerized data center deployment in Fukushima, Japan, using an NVIDIA HGX B200 liquid-cooled cluster for AI training and inference. Compute and management equipment were arranged in container modules and connected to site power, external cooling, and piping, with KONST involved from design and construction planning through testing and delivery.

Racks were assembled and tested before shipment, then connected to power, cooling, and networking on site. Direct liquid cooling removes heat from the GPUs, while air-cooling systems handle network and management equipment.

Fukushima deployment: a B200 liquid-cooled cluster in container modules

The site separates compute and management functions into container modules and deploys B200 equipment as an integrated liquid-cooled rack. This gives each equipment class an appropriate installation and cooling arrangement while defining the interfaces and responsibilities between the rack and site infrastructure before construction.

The compute module contains HGX B200 servers, in-rack coolant distribution units (CDUs), and network equipment. Site infrastructure supplies power distribution, equipment that releases heat outside, backup generation, fire protection, and the required connections.

What is a containerized data center?

A containerized data center places racks and related systems inside prefabricated modules. Different modules may support compute, management, or mechanical and electrical functions, then connect to site infrastructure. Some work moves to a factory environment, but power, cooling, and network connections must still be completed and tested together on site after delivery (commissioning).

Deployment advantages shown by the case

Pre-integration reduces on-site assembly

Server installation, in-rack piping, power, and cabling can be completed earlier. Confirming rack requirements and how they connect to site systems in advance reduces rework during final setup and testing.

Liquid and air cooling handle different heat loads

Cold plates draw heat away from the GPUs. A coolant distribution unit (CDU) transfers that heat to the site's cooling circuit, which carries it to equipment that releases it outside. Air cooling removes heat from network and management devices, along with any heat the liquid-cooling system does not capture.

Modules create clearer units for expansion

Functional modules make future additions easier to scope. Each addition still requires checks on power, heat removal, networking, and room for maintenance.

KONST’s integration role: from GPU equipment to data center delivery

KONST participated in design, construction planning, testing, and delivery, coordinating compute equipment, container modules, and facility power and cooling interfaces in one project flow. This keeps hardware selection, site readiness, and delivery planning aligned.

The project demonstrates practical experience in GPU clusters, containerized data center integration, and cross-border collaboration. Enterprises can evaluate equipment integration, on-site construction, and final system setup and testing together so the facility reflects actual compute requirements.

Contact KONST to discuss GPU cluster and containerized data center deployment.