Power Requirements for GPU Racks: From 10kW to 120kW and Beyond
Summary
Read power at the boundary it describes: a GPU-module rating, a whole-server maximum, a rack requirement and facility energy are different quantities. Compare the actual equipment population with the applicable usable supply and required failure condition. A nominal 30 A or 60 A label does not establish that a feed can support a 20–30 kW rack.
The power question begins with the equipment boundary
A GPU-module wattage excludes other loads in its server. A complete server figure may already include CPUs, memory, fans, storage and power-conversion losses; adding a generic percentage to it can count some loads twice. Preserve the exact manufacturer, configuration and definition of each number.
The useful progression is server to rack, rack to row, then defined IT load to facility support. At each step, state what has already been included and what is additional. A maximum input specification is not an observed workload demand, and neither figure is evidence of spare facility capacity. The existing power-capacity guide explains that separate allocation question.
Use named configurations rather than GPU-family ranges
DGX H100/H200: distinguish chassis size from the power budget
NVIDIA's DGX H100/H200 guide, updated October 17, 2026, lists an 8 U form factor and 10.2 kW maximum system power. Two DGX H100 systems therefore occupy 16 U of server chassis and contribute 20.4 kW to a simultaneous-maximum planning budget. This arithmetic specifies neither a complete rack fit nor measured full-load demand. HGX identifies a different platform context; it is not interchangeable with this complete DGX specification.
EX-PWR-55 uses those two systems as a deliberately limited population. Its buyer stipulates that, after losing one complete supply path, the surviving path must continue to provide the selected 20.4 kW simultaneous-maximum server budget. This is a fictional comparison requirement, not measured demand or an approved topology. Merely remaining powered at reduced performance would be a different outcome. Rack auxiliaries outside the server boundary still need separate accounting.
Eaton's G4 input-cable table lists CBL356-10 with 208 V Delta, three-phase supply, a 60 A panel-breaker entry, 48 A derated current and 17.3 kVA maximum per PDU. Real power in kW equals apparent power in kVA multiplied by power factor; unity power factor gives the favorable upper bound. Even that bound is only 17.3 kW, below the selected 20.4 kW budget. If one complete path is lost, the other path does not inherit its nominal capacity: adding two ratings describes neither surviving path. The return therefore fails this specific full-budget requirement. This is no prediction of measured demand or an inevitable outage, and it establishes no complete DGX cord, outlet, PSU or site compatibility.
Rack and row populations: retain the same requirements
The number of servers placed in a rack cannot be decided from unused U space alone. NVIDIA's DGX H100 SuperPOD electrical guide, last updated November 19, 2025, qualifies its own density and distribution patterns. Its advice against more than four systems per rack has a thermal basis. A general rack-power overview cannot turn those named design patterns into universal approval for every H100/H200 or dense HGX deployment.
For a separate invented row containing ten racks, each assigned a 60 kW planning requirement, the rack subtotal is 600 kW. That sum does not establish the actual diversity, spare supply, operating current, circuit ratings or permitted connection scheme. Keep row services outside the rack figures separate so the responsible facility team can assess the complete population without duplicate allowances.
Phase balance is another part of that actual assessment. The same total across three phases can conceal different per-phase currents, so an aggregate rack figure is not a substitute for the applicable phase distribution and failure condition. Monitoring can inform review; it does not itself demonstrate that an arrangement meets its electrical requirements.
Named rack-scale systems: approximately 120 kW is a scoped example
NVIDIA's DGX GB rack-scale hardware guide, updated September 17, 2026, describes approximately 120 kW rack consumption and AC-fed power shelves distributing nominal 50–51 VDC internally. That is a named system description, not a universal value for every GB200/GB300 OEM configuration. Its internal DC bus does not define the facility AC interface.
Do not convert a rack kW figure into an unqualified current. Current also depends on the applicable voltage, phase arrangement, power factor and operating basis. Schneider Electric's power definitions distinguish real power in kW from apparent power in kVA. A calculated operating current would still be a different quantity from a breaker, conductor, plug or distribution-system rating.
A busway run, tap-off, rack PDU, power shelf and server PSU describe different parts of the path. Compare actual models, applicable interfaces and responsibility boundaries rather than assuming a density threshold selects the distribution architecture. The 415 V/480 V comparison develops that distinct voltage question without supplying an installation design.
Identify each proposed cooling-support device and whether its demand is already inside a rack figure, additional to it, or shared across several racks. Have the responsible owner return the exact equipment population and allocation basis with the power figure. That makes the rack-to-row account usable: a shared CDU can then appear once in the selected boundary, while its required operating condition remains attached to the facility review.
Facility-level accounting
Define the IT population before adding facility support
Include the specified servers and the network/storage IT equipment that lies outside those server or rack figures. Cooling and electrical support belong to the relevant non-IT side of the facility account. The distinction concerns defined boundaries; the word infrastructure alone does not identify which side of the IT denominator a load belongs to.
EX-PWR-56 is a separate fictional one-hour steady planning model. Stipulate 100 racks at 120 kW each: 12 MW of modeled rack demand. Stipulate another 0.5–1 MW of network/storage IT outside those rack figures. The modeled IT total is 12.5–13 MW, or 12.5–13 MWh over that one hour. These are invented steady assumptions, not measured DGX demand or a real facility allocation.
Read PUE as an energy ratio
The Green Grid defines PUE as total data-center energy divided by ICT-equipment energy. For EX-PWR-56, explicitly assume a planning multiplier of 1.25–1.30. Applying the low pair gives 12.5 MWh×1.25=15.625 MWh; the high pair gives13 MWh×1.30=16.9 MWh. Over this stipulated one-hour steady interval, the same numerical values are the modeled average MW. They are not peak-demand sizing results.
The modeled non-IT difference is 3.125 MWh at the low pair and 3.9 MWh at the high pair. It includes the support energy inside the selected facility boundary, rather than cooling alone. An assumed multiplier is not measured PUE, and a measured PUE would not by itself show available capacity for another rack. The PUE and useful-work guide treats that separate efficiency comparison.
Liquid cooling does not eliminate every server or rack fan. The named DGX GB guide retains fan-cooled networking/storage components and air-cooled power-shelf PSUs. Neither a cooling label nor this example establishes a 10–15% IT saving or a universally achievable PUE. Compare actual configuration, environment and operating evidence before assigning an energy benefit.
Specify the required continuity outcome
Required availability is a workload and operating decision. Continuing at reduced performance, retaining power to some components, preserving a recoverable checkpoint and maintaining the required training outcome are different results. No universal dollar loss or number of lost training days follows from an unspecified hour of downtime.
Keep the selected OEM's documented failure behavior and the facility's actual availability design together. The current DGX H100/H200 user guide describes 4+2 PSU redundancy and reduced performance when only three supplies remain powered. The older SuperPOD guide uses a different four-powered-PSU continuity premise in its design discussion. These are dated scopes to resolve with the appointed OEM/facility owners, not instructions to choose or change a connection pattern.
A Tier label cannot replace the customer's required maintenance and failure conditions or prove the condition of every power path. This overview prescribes neither a universal minimum Tier nor a preferred topology for all GPU work. The training-checkpoint guide explains why a saved-state artifact is also a separate question from uninterrupted equipment power.
Request the named equipment's applicable startup/load-step behavior and the actual design condition the continuity system must support. Keep steady demand, transient behavior and required continuity as separate inputs to the appointed electrical/facility review. A generic multiplier cannot supply that missing basis. The UPS and BESS guide separates energy, power and continuity claims; it does not release an electrical installation.
UPS and generation cover different parts of continuity
A UPS can provide short-term continuity while a generator supplies a complementary source; the generator does not conceptually wait for battery depletion before starting. Eaton's UPS/generator discussion emphasizes compatibility and the UPS's input-acceptance conditions. Equipment names alone do not establish the behavior of the installed combination.
Ask the responsible owners to identify the supported IT and cooling population, source-transfer arrangement, required operating interval and applicable equipment conditions. The useful return connects those requirements with the actual continuity design and its accepted evidence. Record whether the return covers continued operation, a controlled transition or a defined recovery outcome; those are different requirements for the project.
Distribution and evidence
Read voltage and compatibility together
A distribution voltage cannot be adopted merely because a higher voltage may reduce current for a comparable power transfer. Retain the named equipment's permitted input and complete interface requirements. A facility distribution label does not establish the voltage presented to an individual server PSU.
For the H100/H200 example, the current user guide identifies its200–240 VAC input scope and qualified supplied cord context. Do not infer that an arbitrary server accepts 277 V or that a familiar connector shape proves a complete approved interface. Those are questions for the actual OEM and authorized facility/electrical owners, not a cable-substitution recommendation.
Compare busway and cable-based scope on the same basis
Ask what each proposal includes: the specified rack population, distribution interfaces, expansion positions, monitoring, operating constraints and responsible parties. A modular product feature can be useful, but it does not supply project installation time, labor availability, future capacity or overall adequacy.
Align the complete proposed population and scope before comparing costs. Identify which interfaces, future positions, monitoring and operating constraints are included, which are supplied elsewhere and which responsibilities are unresolved. Then compare the resulting project totals. A subtotal reduced by moving a service into a separate facility package has not established a cheaper complete solution.
Monitoring supports a defined decision
Identify the measurement boundary, time basis and relevant units. An aggregate kW reading, per-phase current, power factor and cumulative kWh answer different questions. Compare the quantity actually observed with the requirement it is intended to support, preserving which equipment and operating condition were represented.
A monitoring stream is not automatically an integrated DCIM service, a prediction of PSU failure or confirmation of spare allocated capacity. The responsible operating owners must establish the actual integration, review and response responsibilities. Use the GPU power-cap guide for the distinct relationship between configured controls and whole-system input.
Define the permitted physical work after the power requirements
Power-cord connections to existing approved equipment and rack-mounted PDUs may be considered only within a separately agreed scope, subject to OEM instructions, facility authorization, personnel qualifications and insurance. Facility electrical work, hardwiring, new circuits and energization remain outside the standard physical installation service.
The customer and authorized facility/electrical owners must resolve the actual OEM power requirements and approved interfaces before the affected physical work is released.
Discuss a defined physical rack-and-stack package after those requirements are resolved. An initial physical installation inquiry can identify the equipment population, customer-established constraints and requested mounting or IT-cabling tasks; do not send private facility drawings or security-sensitive records through a public inquiry.