AI Data Center EPC — Design, Procurement & Construction
Through the development of an AI data center cluster of up to 1GW in Satsumasendai, Kagoshima, we have accumulated hands-on expertise in site selection, substation design, liquid/air cooling and modular deployment. We have systematized this knowledge and now offer it end-to-end — from site due diligence through design, procurement, construction and operation — to companies and municipalities planning their own AI data centers.
Kagoshima Prefecture will attract clients implementing Physical AI to deploy AI demand, and on that basis build a GW-class data center cluster. Considering room for expansion, ripple effects across the prefecture, and investment in electricity, water and fiber-optic infrastructure from multiple angles, we aim to create demand within the prefecture and establish a model project for large-scale regional revitalization.
We simplify the construction process and deploy GPUs and related facilities at your site by coordinating the optimal management team.
At every step of the EPC process we listen carefully to our corporate clients, collaborate closely, and deliver customized solutions that fully meet their needs.
What matters to the client is that the deliverables of each stage are clear. We define five stages from site due diligence to operational handover, and build GPU-generation decisions into each of them. Because rack power, cooling method, floor loading and substation capacity are all determined by the GPU generation, running GPU procurement and EPC as one integrated process is the single biggest lever for cutting schedule and cost.
We have a track record handling Taiwan-built, NVIDIA-designed AI data servers, with B200 and B300 deployments planned and next-generation Vera Rubin support in preparation. Vera Rubin is manufactured by the Taiwanese supply chain including Wistron, so our direct Taiwan procurement network becomes our procurement power for the next generation. For every platform, we can design and build in either liquid- or air-cooled configurations.
| Item | B200 generationGB200 NVL72 | B300 generationGB300 NVL72 | Next generationVera Rubin NVL72 |
|---|---|---|---|
| GPU configuration | 72× Blackwell | 72× Blackwell Ultra | 72× Rubin + 36× Vera CPU |
| Rack power (approx.) | approx. 120–132kW | approx. 132–142kW (peak approx. 155kW) | approx. 120–130kW (NVL144 CPX configuration approx. 370kW) |
| Cooling | Liquid cooling required (DLC) | Liquid cooling required (approx. 90% of heat rejected via liquid) | Liquid cooling required |
| Memory (per rack) | HBM3e approx. 13.4TB | HBM3e approx. 20TB | HBM4 approx. 20.7TB |
| Floor loading (approx.) | approx. 1.5t per rack | approx. 1,580kg per rack (over ≒2,200kg/m²) | Equal or greater assumed |
| Supply status | Mass-produced, deployed generation | Current mainstream generation | Volume production and shipments start in 2026; allocations tight through 2027 |
* Figures are reference values based on public information as of August 2026. The design assumptions differ fundamentally from conventional data centers (5–10kW per rack, air-cooled). Once the GPU generation is fixed, the facility requirements are fixed — which is exactly why placing GPU procurement and EPC as one order cuts schedule and cost.
The competitiveness of an AI data center is decided by the power and cooling infrastructure that keeps its GPUs running reliably. We design and build everything from substations to cooling and power management as one system, achieving both high utilization and high efficiency.
Through Taiwanese manufacturing partners such as Wistron, we handle NVIDIA-designed AI data servers (GB200/GB300 NVL72 class) end-to-end — from procurement through installation, power connection and commissioning. AI servers exceeding 120kW per rack cannot be accommodated by the design conventions of traditional data centers. A track record with real hardware is what underwrites a facility that will actually run. All construction is carried out by Kraftia and Seiko Electric of the Kyushu Electric Power group.
Construction is handled not by a team with plans on paper, but by one that has actually built substations and deployed AI servers. A structure that can manage everything from power infrastructure construction through IT equipment deployment and commissioning as one is the foundation for executing AI data center EPC with certainty.
