Mitsubishi Electric Power Products, Inc. (MEPPI) has introduced Chip-to-Grid Reference Designs for AI factories built around NVIDIA Vera Rubin NVL72 and future NVIDIA accelerated computing platforms. The architecture connects utility power, on-site generation, battery storage, electrical distribution, liquid cooling and facility controls into a standardized infrastructure blueprint designed to help hyperscale, neocloud and colocation operators deploy high-density AI systems and scale from 250-megawatt blocks toward gigawatt-scale campuses.
The rapid expansion of generative AI infrastructure is changing the definition of an AI data center. As accelerator systems become denser and AI workloads require more computing capacity, the infrastructure surrounding the GPU increasingly determines how quickly an AI factory can be built, powered and expanded.
Mitsubishi Electric Power Products (MEPPI) is addressing that challenge with a new Chip-to-Grid Reference Design, an integrated blueprint that connects AI computing infrastructure with the electrical and thermal systems required to operate it.
The designs are intended for hyperscale, neocloud and colocation operators in North America and are built around 250 MW deployment blocks, with an architecture that can scale toward gigawatt-sized campuses. MEPPI says the infrastructure can support rack densities growing from roughly 200 kW today to more than 1 MW per rack.
The announcement reflects a broader shift in AI infrastructure: compute, power and cooling are increasingly being designed as a single system rather than separate layers.
AI Factories Are Becoming Power Infrastructure Projects
The industry’s infrastructure challenge is increasingly visible in electricity demand.
The International Energy Agency projects global data-center electricity consumption will rise from about 485 TWh in 2025 to roughly 950 TWh in 2030. AI-focused data centers are expected to grow considerably faster than data centers overall, with AI-related electricity consumption projected to triple during that period.
For AI developers and cloud operators, having access to accelerators is therefore only one part of the deployment equation. A facility also needs sufficient grid capacity, transformers, power distribution, backup systems, cooling infrastructure and controls capable of supporting high-density compute.
MEPPI’s Chip-to-Grid architecture is designed around that complete infrastructure stack.
Each 250 MW block can operate as a grid-connected facility or in an islanded configuration using on-site generation and battery energy storage, with a pathway toward utility interconnection. The design is intended to allow additional capacity to be added incrementally rather than requiring an entire campus architecture to be redesigned as AI systems evolve.
Designing Around NVIDIA Vera Rubin
The reference designs are built to support NVIDIA Vera Rubin NVL72 and future NVIDIA AI infrastructure.
NVIDIA’s Vera Rubin platform is designed around rack-scale AI computing. NVIDIA’s technical documentation describes the Vera Rubin NVL72 architecture as integrating 72 Rubin GPUs and 36 Vera CPUs within a rack-scale system.
That density changes the requirements placed on the facility.
NVIDIA’s own DSX AI Factory reference architecture treats power, cooling, networking, compute and facility infrastructure as interconnected elements. The company’s March 2026 reference design was explicitly created to provide a repeatable approach for designing and operating AI factories around Vera Rubin infrastructure.
MEPPI’s approach extends that concept into the electrical and thermal infrastructure connecting the AI rack to the wider facility and ultimately to the power grid.
800 VDC Addresses the Power Distribution Challenge
One of the technologies incorporated into the design is NVIDIA’s 800 VDC architecture.
As rack power requirements rise, traditional electrical distribution can require multiple stages of power conversion. NVIDIA says its 800 VDC approach is intended to reduce conversion complexity and improve the efficiency of delivering electricity to increasingly dense AI compute systems. The company has been developing the architecture with ecosystem partners through the Open Compute Project.
For MEPPI, supporting this architecture means the power system has to be designed around the requirements of next-generation AI racks rather than treating the rack as a conventional data-center load.
That is an important distinction. The electrical architecture of a 1 MW rack cannot simply be treated as a scaled-up version of a lower-density server rack.
Cooling Becomes a Compute Requirement
Power is only half of the physical infrastructure equation. High-density AI accelerators also generate substantial heat, requiring increasingly sophisticated thermal management.
MEPPI’s reference design incorporates a dual-loop cooling architecture. The system combines elevated-temperature direct liquid cooling for high-density AI workloads with lower-temperature chilled-water air-side cooling.
This allows different facility loads to be handled according to their thermal requirements while supporting future increases in compute density.
Liquid cooling is becoming increasingly important for rack-scale AI infrastructure. NVIDIA’s current high-density AI systems, including its NVL72 architectures, use liquid-cooling approaches designed to support dense accelerator configurations.
The result is a tighter relationship between compute architecture and facility engineering: changes in processor performance and rack density can directly affect electrical distribution, cooling loops, backup power and building design.
From Data Center Design to AI Factory Architecture
MEPPI’s reference designs are positioned as a repeatable deployment framework rather than a single facility configuration.
The company says the architecture combines utility interconnection, generation, battery energy storage, electrical distribution, cooling and facility controls. The intent is to reduce engineering complexity and allow operators to deploy capacity in standardized blocks.
That approach mirrors NVIDIA’s broader AI factory strategy. NVIDIA describes its DSX reference designs as generation-specific architectures for AI factories, covering compute, networking, storage and facility infrastructure.
The shift matters because AI infrastructure increasingly resembles an industrial production system. The objective is not simply to house servers; it is to provide a predictable environment in which enormous quantities of AI inference and training computation can operate continuously.
Power Flexibility Could Become a Competitive Infrastructure Layer
MEPPI’s support for both grid-connected and islanded operation also reflects the growing importance of power flexibility.
Grid interconnection can take years in constrained markets, while new AI campuses may require power capacity that is difficult to secure through conventional utility connections alone. On-site generation and battery storage can provide alternative pathways for bringing capacity online, although the economics, permitting requirements and energy sources will vary by location.
The IEA has identified grid connections, transformers and other energy infrastructure as emerging bottlenecks for data-center expansion. Its 2026 analysis notes that physical constraints are already affecting the pace at which new data-center capacity can be developed.
This makes the chip-to-grid concept increasingly relevant. The bottleneck is no longer confined to semiconductor availability or server manufacturing. The entire chain from accelerator to electrical grid has become part of AI infrastructure planning.
A New Layer of AI Infrastructure
MEPPI’s announcement illustrates how the definition of AI infrastructure is expanding.
The traditional model separated compute from facility engineering. AI factories are increasingly designed as integrated systems in which accelerator architecture, power delivery, cooling, networking and controls are optimized together.
For operators planning hundreds of megawatts or eventually gigawatts of AI capacity, standardized reference architectures could reduce the amount of engineering required for each deployment while providing a pathway to adopt new generations of accelerators.
MEPPI’s Chip-to-Grid Reference Designs are therefore aimed at a problem that sits below the AI model layer but increasingly determines how quickly AI models can be deployed at scale: how to reliably deliver power and remove heat from the computing infrastructure that runs them.
As AI workloads continue to expand, the next generation of AI factories will be defined not only by the number of accelerators they contain, but by how efficiently and flexibly the entire facility can turn electricity into sustained AI compute.
Market Landscape
AI infrastructure is moving toward codesigned compute, power and cooling architectures as accelerator density increases. NVIDIA’s Vera Rubin DSX reference design already treats compute, networking, power, cooling and facility controls as components of an integrated AI-factory architecture.
The energy requirement is becoming equally significant. The IEA expects global data-center electricity consumption to approach 950 TWh by 2030, while AI-focused data-center consumption is projected to grow faster than the overall market.
MEPPI’s approach adds another layer to this trend: standardized chip-to-grid infrastructure that can combine utility connections, on-site generation, energy storage, high-voltage distribution and advanced thermal management around the requirements of next-generation AI racks.
Top Insights
- MEPPI’s Chip-to-Grid Reference Designs connect AI compute with power generation, storage, distribution, cooling and facility controls.
- The architecture uses 250 MW deployment blocks designed to scale toward gigawatt-sized AI campuses.
- NVIDIA 800 VDC support addresses power-distribution requirements emerging from increasingly dense AI accelerator systems.
- Dual-loop cooling combines direct liquid cooling with chilled-water air-side cooling for different facility workloads.
- Rising data-center electricity demand is making grid access, energy infrastructure and thermal management strategic AI deployment constraints.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
