AI Reshapes Data Centre Rack Design Beyond Open Compute Standards
AI Data Centre Rack Design Evolves Beyond OCP Standards
Open Compute Project standards have long defined data centre infrastructure, but artificial intelligence workloads are pushing rack design beyond these traditional boundaries. Hyperscale and neocloud operators are discovering that conventional specifications no longer accommodate the thermal, power, and connectivity demands of modern AI deployments.
NVIDIA MGX and next-generation cooling solutions are enabling denser GPU configurations and improved airflow management at the rack level. These architectural innovations address critical pain points in AI infrastructure—from power distribution to liquid cooling integration—while maintaining compatibility with existing data centre ecosystems.
The rack itself is being redesigned
Artificial intelligence is changing data centre infrastructure at the rack level, and the clearest signal yet came from Raritan’s parent company. In a post published on 13 July 2026, Raritan set out how AI and NVIDIA MGX are reshaping rack design for hyperscale and neocloud environments. The conclusion is blunt: Open Compute Project standards alone no longer cover what AI deployments need at the rack.
What is NVIDIA MGX? NVIDIA MGX is a modular reference architecture for building accelerated servers and racks. Instead of one fixed design, MGX defines standard building blocks that OEMs and partners combine into more than 100 system configurations, from single-node servers to full rack-scale GPU systems. For data centre operators, an MGX-class rack means higher power density, integrated cooling, and cabinet dimensions that no longer match the fixed 19-inch assumptions older infrastructure was specified around.
Five shifts at the rack level
Raritan identifies five changes underway in AI deployments:
Legrand’s MGX/NVL cabinet is the concrete response: integrated frame construction with MGX cooling built in, flexible widths and depths, and the ability to mount 19-inch rails inside an OCP-based structure. The cabinet is no longer a fixed constant that other equipment bolts into.
Where power distribution fits
Power distribution sits inside every one of those five decisions, and the stakes are specific. Powering on a fully populated NVIDIA GB200 NVL72 rack without proper sequencing can trip upstream breakers. During training, load can double within milliseconds. Monitoring that polls every 100 milliseconds records none of it.
That is why the PDU has moved from commodity line item to design decision. The Raritan PX4 is built for exactly this rack:
When the cabinet, the cooling, and the mounting are all changing at once, the PDU stops being an afterthought on the build sheet. Specify it with the rack, not after it.
Frequently asked questions
What is an MGX-class rack?
An MGX-class rack is a cabinet built around NVIDIA’s MGX modular reference architecture for GPU systems. It typically carries higher power density than a general-purpose rack, integrates cooling into the frame, and may use non-standard widths and depths alongside conventional 19-inch mounting.
Why do OCP standards alone no longer cover AI rack design?
OCP standardised rack design and accelerated deployment at scale, but GPU systems introduce power densities, thermal profiles, and weight loads that push past standard dimensions and mounting approaches. Raritan’s position is that operators now need OCP consistency plus targeted customisation in the same environment.
What makes powering on a GPU rack risky?
Inrush current. A fully populated NVIDIA GB200 NVL72 rack powered on without sequencing can draw enough inrush to trip upstream breakers and cascade into a facility-level event. Per-outlet switching lets the PDU stage the power-on so nodes come up in sequence.
How does the Raritan PX4 differ from a basic metered PDU in an AI rack?
A basic metered PDU reports one total per circuit. The PX4 measures at each outlet: inrush current, harmonic distortion, and waveform capture at the moment of an event, with 0.5% metering accuracy. At GPU density, the circuit average hides exactly the behaviour that causes failures.
What is the Legrand MGX/NVL cabinet?
A cabinet platform designed for NVIDIA MGX and NVL deployments. It builds MGX cooling capability into the frame, supports flexible widths and depths for non-standard equipment, and allows 19-inch EIA rails to be mounted within an OCP-based structure so legacy and next-generation equipment share the same environment.
Can Enova supply and support Raritan PX4 PDUs in Singapore?
Yes. Enova Technologies is an authorised Raritan reseller in Singapore and can advise on PX4 configuration for AI and GPU rack deployments, including outlet-level metering, switching, and sensor options. Contact [email protected].
Planning GPU racks? Specify the PDU with the rack, not after it.
Ask us about the Raritan PX4Source: Raritan, “How AI and NVIDIA MGX are Changing Rack Design for Hyperscale and Neocloud Environments”, 13 July 2026.
Published by
eNOVA Technologies is Singapore's specialist distributor for data centre IT management solutions, representing Adder, Guntermann & Drunck, Raritan, Sunbird, ZPE Systems, and VuWall across Singapore and Southeast Asia. Our technical content is produced with AI assistance and reviewed by our in-house team before publication.
This article was produced with AI assistance and reviewed by the eNOVA Technologies team. All technical claims are verified against manufacturer documentation.
