Summary
- Mitsubishi Electric's reference architecture links grid connection, onsite generation, storage, distribution, rack power, and cooling.
- The designs support conventional AC and emerging 800V DC distribution.
- Direct liquid cooling is combined with air cooling in a dual-loop thermal architecture.
Mitsubishi Electric has published reference designs that combine grid connections, onsite generation, batteries, electrical distribution, rack power, and cooling into a single architecture for next-generation AI data centres.
The Chip-to-Grid DSX designs have initially been developed through the company’s US subsidiary for NVIDIA Vera Rubin NVL72 and future AI infrastructure, but Mitsubishi says it intends to use the architecture to develop data centre solutions globally.
The designs support both conventional 415/480V AC distribution and emerging 800V DC systems inside the facility.
They also combine direct liquid cooling at the chip with air cooling for other components through a dual-loop cooling architecture.
At site level, Mitsubishi’s model includes onsite power generation and battery energy storage. The company says this can reduce grid dependence during normal operation and allow a facility to switch into autonomous microgrid operation during a grid failure.
Rack architecture is reaching upstream into the facility
The significance of the design is not an individual product. It is the attempt to treat the power chain from utility connection to processor, and the thermal chain from chip to external heat rejection, as one engineering system.
AI hardware is increasing rack power densities quickly enough that traditional boundaries between IT equipment and facility systems are becoming less useful. Changes in accelerator architecture can alter the required voltage, rack distribution, busway, backup power, cooling-water temperatures, coolant distribution units, and mechanical plant.
The move toward 800V DC is part of that shift. Conventional data centres repeatedly transform and rectify power as electricity moves from the grid through distribution equipment and eventually into server components. Higher-voltage DC architectures are intended to reduce conversion stages and conductor requirements for extremely dense computing loads.
They also change protection, switching, maintenance, and safety requirements. DC faults behave differently from AC faults, while equipment availability and common technical standards will influence how quickly the architecture can move from reference designs into mainstream European facilities.
Mitsubishi’s inclusion of onsite generation and batteries also reflects the growing power constraint facing AI projects. Designers are increasingly being asked to consider grid-connected, behind-the-meter, and microgrid systems together rather than treating backup generation as an isolated emergency layer.
The cooling design follows the same integrated approach. Direct liquid cooling removes heat from the highest-density components, while air remains necessary for memory, power electronics, networking equipment, and other components that are not connected to the liquid loop.
That hybrid environment places more emphasis on coordination between rack supply, coolant distribution, facility-water systems, and air cooling. A high-density data hall cannot simply swap air-cooled servers for liquid-cooled racks without checking the mechanical plant, water temperatures, controls, leak management, and heat-rejection capacity around them.
Mitsubishi’s first reference designs target US deployments, so they are not themselves a European project announcement. The engineering questions they address are nevertheless directly relevant to European developers preparing sites for next-generation AI hardware.
The direction of travel is increasingly clear: processor roadmaps are beginning to dictate decisions further upstream, from rack voltage and cooling loops to substations, generation, storage, and grid interface. Facility design is being pulled closer to the chip.
