IN Brief:
- Mitsubishi Electric Power Products has released chip-to-grid reference designs based around repeatable 250MW AI factory deployment blocks.
- Each block can operate grid-connected or islanded using on-site generation and battery storage, with a pathway to later utility connection.
- The architecture supports high-density computing including NVIDIA's 800VDC approach and rack power moving from around 200kW towards more than 1MW.
Mitsubishi Electric Power Products has released chip-to-grid reference designs for AI data centres, using repeatable 250MW deployment blocks that combine utility interconnection, on-site generation, battery energy storage, electrical distribution, cooling, and facility controls.
The designs have been developed around NVIDIA Vera Rubin NVL72 and future high-density computing infrastructure and are intended for hyperscale, neocloud, and colocation operators in North America. Mitsubishi says individual 250MW blocks can be repeated as campuses grow towards gigawatt scale.
Each block can operate connected to the utility network or in an islanded configuration using on-site generation and battery storage, while retaining a pathway to later grid connection. The electrical supply strategy is therefore part of the reference architecture rather than an assumption that sufficient utility capacity will always be available before computing equipment is installed.
The approach differs from a single component launch. Mitsubishi is not presenting one converter, switchboard, or battery as the answer to AI power constraints; it is defining how several layers of electrical and thermal infrastructure can be arranged around a repeatable computing block.
Power availability has become a major constraint on large AI developments because campus demand is rising while individual racks become considerably denser. Mitsubishi’s architecture assumes rack loads moving from around 200kW towards more than 1MW, forcing changes throughout electrical distribution and cooling systems.
A megawatt-class rack concentrates electrical demand into a footprint historically associated with much lower power levels. Current rises sharply unless distribution voltage increases, while every conversion loss becomes additional heat that cooling equipment must remove. Electrical and thermal design therefore become tightly linked.
Mitsubishi’s architecture supports NVIDIA’s emerging 800VDC distribution approach. The company has previously announced work around high-voltage DC for AI infrastructure, so the new development is broader than the voltage level itself. The 250MW blueprint extends from rack distribution to utility connection, generation, batteries, cooling, and campus controls.
That wider scope distinguishes it from recent supplier announcements centred primarily on DC distribution equipment. Recent work on hybrid AC/DC architectures has concentrated on conversion and distribution closer to the computing load; Mitsubishi’s reference design adds a site architecture capable of operating before, during, or after the availability of a large utility connection.
Islanded operation changes the engineering requirements significantly. On-site generators and batteries have to maintain frequency, voltage, reserve, fault response, and power quality without relying continuously on a strong external grid. Controls must coordinate those resources as computing demand changes, equipment trips, or the site moves between operating modes.
Battery storage can respond quickly but cannot sustain an islanded campus indefinitely without sufficient energy or continuing generation. Generators can provide longer-duration supply, but fuel infrastructure, emissions, start performance, maintenance, and redundancy have to be engineered around the availability expectations of a mission-critical facility.
The grid-connected version creates another set of constraints. A 250MW block is already comparable with a major industrial load, and repeating several blocks can take campus demand towards gigawatt scale. Utilities may need new transmission capacity, substations, transformers, protection schemes, and generation resources before they can provide that power continuously.
The modular approach gives developers a standard starting point while allowing the supply arrangement to evolve. A block could begin with substantial on-site generation and later connect to a reinforced utility system, or operate primarily from the grid while batteries and generation support resilience, peak management, and ride-through requirements.
Cooling follows the same modular principle. The reference architecture uses a dual-loop approach combining elevated-temperature direct liquid cooling for high-density computing with lower-temperature chilled-water systems for remaining air-side loads. The design recognises that not every electrical component and computing load operates at the same thermal conditions.
Standardisation could reduce repeated engineering where operators intend to build several similar halls or campuses. Once a block has been validated, parts of its electrical topology, protection philosophy, controls, and cooling arrangement can be replicated rather than designed from first principles for every expansion.
Site-specific engineering does not disappear. Utility fault levels, grid codes, generator technology, environmental permits, fuel availability, climate, water constraints, land, and redundancy requirements can all change the final system. A reference design can reduce repetition without making a 250MW electrical installation an off-the-shelf product.
The architecture also creates interfaces between suppliers that must be controlled carefully. Utility switchgear, generators, batteries, converters, busways, rack power equipment, chillers, and supervisory controls may come from different manufacturing groups even where Mitsubishi provides the overall blueprint.
Protection coordination becomes especially demanding when a site can operate both grid-connected and islanded. Fault current levels, grounding, relay settings, transfer sequences, synchronisation, and restart procedures can change with the operating mode, requiring the controls and protection scheme to recognise how the campus is being supplied at any given moment.
The 250MW block gives Mitsubishi a defined unit around which that engineering can be organised. Actual projects will show whether operators adopt the full architecture or select individual elements according to utility capacity, deployment speed, and existing facility standards.
The next substantive milestone is therefore a named deployment. The blueprint establishes how Mitsubishi proposes to combine utility power, generation, storage, high-voltage DC distribution, and cooling; live projects will determine how readily that 250MW model survives contact with local grids, planning rules, equipment lead times, and the availability requirements of AI computing.
