Overview: Why GPU Rack Power Architecture Is Different

Modern AI compute racks bear little resemblance to general-purpose server infrastructure. A single NVIDIA DGX H100 node draws approximately 10.2 kW across eight H100 GPUs, each rated at 700 W TDP. A full GB200 NVL72 rack—housing 72 Blackwell GPUs and 36 Grace CPUs—can reach approximately 120 kW per rack. These densities fundamentally change how power is delivered, distributed, protected, and cooled. Data-center builders and cabling contractors must treat GPU rack infrastructure as a specialized discipline from the earliest design stages, not as an incremental upgrade to conventional IT rack planning.

Facility Power Delivery: Starting at the Source

At H100 and Blackwell densities, 120/208 V three-phase power infrastructure is insufficient. Facilities serving these workloads standardize on 480 V three-phase distribution, which reduces conductor sizing, minimizes resistive losses over long branch runs, and supports the high per-rack kVA demands without impractical wire gauges. Power must be engineered at the facility level before a single rack is specified.

Key facility-side decisions include:

Transformer placement: Step-down transformers (480 V to rack PDU input) should be located as close to the compute row as practical to minimize voltage drop on high-current feeders.
UPS architecture: N+1 UPS configuration is the baseline for production AI infrastructure. At these power densities, even momentary outages can corrupt in-flight training jobs representing hours of compute. UPS bypass and maintenance isolation must be designed into the single-line diagram from the start.
Branch circuit sizing: Individual rack circuits must be sized with appropriate headroom above steady-state draw to accommodate startup inrush and transient GPU load spikes during training workloads.

Power Distribution Within the Row: Busway vs. Raised Floor

Traditional raised-floor power distribution becomes impractical at Blackwell densities. The volume of liquid cooling infrastructure—supply and return manifolds, flexible hoses, and containment systems—competes directly with power cabling in underfloor pathways, creating installation and maintenance conflicts. The preferred approach for high-density AI rows is overhead busway or overhead cable tray distribution.

Overhead busway (also called plug-in busduct) allows tap-off boxes to be positioned at each rack without home-run conduit runs back to a panel, significantly simplifying adds, moves, and changes. It also keeps the highest-current conductors physically separated from the liquid cooling infrastructure at floor level. Contractors should confirm busway ampacity ratings match the worst-case rack load per row, accounting for diversity factors across the deployment timeline.

For facilities that retain a raised floor, power pathways must be coordinated with mechanical engineering early to reserve dedicated zones for cooling infrastructure, avoiding the scenario where floor tiles become inaccessible after cooling hoses are routed beneath them.

Rack-Level Power Distribution Units (PDUs)

Intelligent, monitored rack PDUs are not optional at these densities—they are an operational necessity. Per-outlet or per-branch metering enables the facility team to track actual draw against nameplate capacity, identify load imbalance across phases, and detect anomalous consumption that may indicate hardware faults. For 480 V deployments, step-down transformers integrated into or co-located with the rack PDU step voltage down to the input requirements of the server hardware.

Phase balancing is a critical installation task. GPU nodes present large, relatively constant loads. Contractors responsible for rack PDU installation should verify three-phase balance at commissioning and document it per the labeling requirements outlined in ANSI/TIA-606-C, which governs administration and labeling of telecommunications infrastructure including power pathways in data centers.

Cooling Architecture and Its Impact on Power Infrastructure

Cooling method is inseparable from power architecture at these densities. The two primary approaches have direct implications for both electrical and mechanical trades.

H100-Density Deployments: Rear-Door Heat Exchangers

DGX H100 nodes at approximately 10.2 kW per node can be served by rear-door heat exchangers (RDHx), which transfer heat from exhaust air into facility chilled water without requiring direct liquid connections to each server. This approach preserves a more conventional rack form factor and allows standard overhead or underfloor power distribution. The primary electrical consideration is ensuring that RDHx units with active fans or pumps have their own power circuits, separately metered from the compute load.

Blackwell-Density Deployments: Direct-to-Chip Liquid Cooling

The GB200 NVL72 and equivalent Blackwell configurations require direct-to-chip liquid cooling and are liquid-cooled by design. At approximately 120 kW per rack, air cooling is not a viable alternative. Direct liquid cooling introduces supply and return manifolds, quick-disconnect fittings, and flexible hose assemblies that must coexist with power and signal cabling in the same rack and pathway environment.

Contractors must plan cable management systems that maintain minimum bend radius requirements for high-count fiber trunks (per ANSI/TIA-568.3-D for optical fiber cabling) while routing liquid cooling hoses through the same vertical cable managers or overhead trays. Crush hazards, pinch points, and accidental disconnection of cooling circuits are real commissioning risks that require explicit coordination between electrical, mechanical, and cabling crews.

Structured Cabling Integration at High GPU Density

Power and cabling infrastructure cannot be designed in isolation for GPU clusters. The following cabling realities directly affect pathway and space planning:

High-density fiber: 400G InfiniBand (ConnectX-7, NDR) and 400GbE Spectrum-X links use QSFP-DD or OSFP transceivers. High port-count switches require 144-fiber MPO/MTP trunk assemblies in overhead trays, planned in accordance with ANSI/TIA-568.3-D.
DAC and AOC constraints: Direct Attach Copper (DAC) cables are limited to approximately 3 meters, making them suitable only for top-of-rack or within-rack connections. Active Optical Cables (AOC) extend to approximately 30 meters for inter-rack and row-to-row runs. Runs beyond this require structured single-mode or multimode fiber infrastructure.
NVLink cabling: NVLink interconnects within DGX/HGX systems use proprietary harnesses managed within the baseboard or chassis. These are not field-terminated and should not be modified. Between nodes and racks, standard InfiniBand or Ethernet cabling applies.
Labeling and administration: Every fiber trunk, copper link, and power circuit serving a GPU cluster should be labeled and documented per ANSI/TIA-606-C to support rapid troubleshooting in high-stakes production environments.

Data center infrastructure design should align with ANSI/TIA-942, which addresses site, space, power, and cabling infrastructure for data centers holistically. Contractors should reference this standard when coordinating with facility architects on pathway routing, clearances, and redundancy tier requirements.

Commissioning and Operational Considerations

Before energizing a high-density GPU row, commissioning teams should verify phase balance and voltage under load, confirm liquid cooling flow rates and leak detection alarm functionality, validate PDU metering against known reference loads, and document as-built cabling with end-to-end continuity test results per applicable TIA standards. At these power densities, a commissioning deficiency that would be a minor inconvenience in a general-purpose server environment can result in significant hardware damage or facility-level electrical events. Treat the commissioning checklist as a safety document, not a formality.