AI Server Rack Power: Four DGX H100 Systems, kW and Monthly kWh

Calculate AI server rack power and monthly electricity with explicit system, IT and facility boundaries. Four NVIDIA DGX H100 systems have a combined published maximum system power of 40.8 kW. For an explicitly hypothetical measured average of 7.0 kW per system, their average rack IT load would instead be 28.0 kW. Assuming a 30-day month, PUE of 1.40 and electricity at $0.12/kWh, that average-load example uses 20,160 kWh of IT electricity, corresponds to 28,224 kWh of allocated facility electricity, and produces an illustrative energy charge of $3,386.88. Only the 10.2 kW maximum per system is NVIDIA’s rating; the average, PUE and tariff are assumptions, not reported measurements. [1.2] [4.2] [4.4]

Reproduce these assumptions in the server-energy and electricity-cost calculator: enter 7,000 average watts, four servers, 720 hours, PUE 1.40 and a rate of 0.12 per kWh. For the measurement comparison, see the AI versus ordinary server power guide.

1. Start with the right boundaries—not just a wattage

• Maximum rating: NVIDIA lists DGX H100 system power consumption as 10.2 kW maximum. Four systems therefore total 4 × 10.2 = 40.8 kW on a maximum-rating basis. This is not evidence that the rack continuously consumes that much electricity. [1.2]
• Operating energy: Use average electrical input power over the period being estimated, not the maximum rating alone. In this original example, 7.0 kW per system is an explicitly hypothetical meter-derived average: a value we suppose a representative measurement campaign returned, not an actual test result. [1.2] [4.5]
• Facility electricity: Keep IT electricity separate from supporting infrastructure. DOE defines PUE as total annual data-center facility energy divided by annual IT equipment energy. Applying an assumed PUE to the rack’s IT energy produces a facility-energy allocation, not an additional load that the rack outlets themselves must supply. [4.2] [4.4]

The calculation boundary here is four complete DGX H100 systems only. NVIDIA’s 10.2 kW figure is a system specification, so it should not be replaced by GPU-only power or supplemented by separately adding the internal components already covered by the system rating. External switches, storage and other equipment are excluded from the numerical example. If they are part of your rack, include their measured or appropriately rated loads before calling the result the complete rack IT total. This distinction matters because a four-server subtotal and an entire populated rack are not necessarily the same electrical boundary. [1.2] [7.3] [7.4]

2. Establish the inputs and the calculation method

Use five explicit inputs for this example: four systems; 7.0 kW hypothetical average input per system; 720 hours, calculated as 30 days × 24 hours/day; PUE = 1.40; and an assumed flat energy tariff of $0.12/kWh. These are deliberately chosen illustration inputs, except for the separately retained NVIDIA maximum rating. The 7.0 kW average is not an assertion about typical DGX H100 operation, and the tariff is not a utility quote. Using a 30-day month also keeps the duration transparent rather than silently substituting an annual-average month. [1.2] [4.2]

The essential distinction is between power in kW and energy in kWh. Power describes the electrical load; energy accounts for that load over time. For a specified period, calculate IT energy = average IT power × hours. DOE’s PUE definition then gives the relationship facility energy = IT energy × PUE, provided the PUE appropriately represents the energy boundary and period. Finally, under the example’s assumed flat tariff, calculate energy charge = facility energy × price per kWh. These are unit-consistent arithmetic relationships, not workload-performance predictions. [4.2] [4.4]

For a rack with several devices, first sum their average input powers: P_rack,avg = P_1,avg + P_2,avg + … + P_other,avg. When consumption changes during the month, use time-weighted intervals instead: E_IT = Σ(P_interval × hours_interval). A period average is then total measured energy divided by elapsed hours. This preserves the distinction between changing IT loads and a fixed maximum specification. DOE explicitly notes that annual energy measurements capture dynamic IT-load trends, including power management; the same accounting principle explains why a workload-representative average is more useful for an energy forecast than a single instantaneous reading. [1.2] [4.5]

_images/dgx-h100-with-bezel.png · docs.nvidia.com

3. Work through the hypothetical average-load month

First calculate the rack’s average IT power from the hypothetical measured input: 4 systems × 7.0 kW/system = 28.0 kW. Next calculate monthly IT electricity: 28.0 kW × 720 h = 20,160 kWh. This is the electricity attributed to the four systems before adding facility overhead. It assumes that 7.0 kW is the average across the entire modeled month, including whatever workload and idle periods that month contains. It does not mean each server draws exactly 7.0 kW at every instant, nor does it replace the maximum rating for electrical-capacity review. [1.2] [4.5]

Next apply the assumed facility factor: 20,160 kWh × 1.40 = 28,224 kWh. Equivalently, the allocated average facility power is 28.0 kW × 1.40 = 39.2 kW, and 39.2 kW × 720 h = 28,224 kWh. The difference is 28,224 − 20,160 = 8,064 kWh of allocated supporting-infrastructure energy. This is 40% additional energy relative to IT energy, because the assumed multiplier is 1.40. It is not a claim that 40% of the facility total is overhead, and it is not extra power drawn through the four servers’ cords. [4.2] [4.4]

Finally apply the hypothetical tariff: 28,224 kWh × $0.12/kWh = $3,386.88. This is an illustrative facility-energy charge for the rack allocation, not a complete electricity bill or a quoted hosting price. If you want the four systems’ IT-only energy charge under the same assumed rate, the arithmetic is 20,160 × $0.12 = $2,419.20. Choose the correct boundary before quoting a cost: IT-only electricity and PUE-adjusted facility electricity answer different budgeting questions. Whether your actual commercial agreement charges either amount must be checked against its billing terms. [4.2] [4.4]

4. Compare the operating illustration with the maximum-rating case

A useful comparison holds the system count, month length, PUE assumption and tariff assumption constant, while changing only the system-power input. The maximum-rating case uses 4 × 10.2 = 40.8 kW and assumes that load persists for all 720 hours. Its IT electricity is 40.8 × 720 = 29,376 kWh; allocated facility electricity is 29,376 × 1.40 = 41,126.4 kWh; and the illustrative energy charge is 41,126.4 × $0.12 = $4,935.168, rounded to $4,935.17. This is a constant-maximum scenario, not NVIDIA’s forecast of monthly consumption. [1.2] [4.2]

The table keeps rating-based planning separate from hypothetical operating energy. Both columns exclude external rack equipment. Neither column establishes an actual site PUE, utility rate or measured workload average. The comparison is useful because it shows exactly where the energy estimate changes: the hypothetical average is below the published maximum, while every downstream assumption is unchanged. It does not demonstrate that a particular deployment will achieve the lower figure, and the maximum-based monthly calculation should not be mistaken for a substitute for a complete electrical design. [1.2] [1.4] [4.2]

5. Use PUE as an energy-allocation assumption, not a rack specification

DOE’s definition is explicitly annual: total annual facility energy divided by total annual IT equipment energy. Its guide also explains that PUE describes supporting-infrastructure efficiency, not the overall efficiency of the computing operation. For this worked month, 1.40 is simply a chosen modeling factor. Multiplying a rack subtotal by site PUE is a proportional allocation of shared facility energy; it does not independently measure the infrastructure serving that rack. That distinction is especially important when presenting the resulting 39.2 kW as a budgeting figure rather than a physical reading at a rack PDU. [4.2] [4.4]

Before replacing the example with a forecast, obtain the facility’s applicable energy data and identify whether its PUE is an annual historical value, a period-specific calculation or a design assumption. DOE notes that annual measurements account for variable free-cooling savings and dynamic IT loads. Consequently, an annual PUE can be a transparent simplifying input for a monthly budget, but it should not be presented as a measured result for that particular month. If matched-period facility and IT energy readings are available, document their boundaries and period rather than mixing unrelated readings into an apparently precise ratio. [4.2] [4.4] [4.5]

Watch on YouTube ↗

Background video on AI data-center power and cooling. · www.youtube.com

6. What to measure and verify before deployment

Measure representative electrical input. Before committing the production deployment, arrange a representative-system test or obtain applicable measurements for the intended configuration and workload. Record energy and elapsed time so the average can be calculated, and record load variation and peaks separately. Include intended busy periods, idle periods and any external rack IT equipment in the accounting boundary. Because NVIDIA documents multiple power supplies with balanced load distribution, capture total system input rather than treating one cord as the complete system. A post-installation rack measurement should then validate the proposed budget; it cannot retroactively establish that the initial hypothetical average was measured. [10.1] [4.5] [7.4]

Verify distribution capacity and redundancy independently of average energy. Give the electrical designer the four-system 40.8 kW maximum subtotal, the complete equipment inventory and the intended power-feed arrangement. Check the available circuit and PDU capacities and the required operating condition when a feed or component is unavailable. NVIDIA specifies N+2 PSU redundancy for the DGX H100 and says the system ships with six qualified locking power cords. Those facts do not establish that a proposed rack’s upstream distribution is sufficient. Likewise, multiplying a power-supply module rating by its count is not a replacement for NVIDIA’s published whole-system maximum. [1.2] [1.3] [2.1] [10.1]

Check cooling, physical fit and the actual billing inputs. NVIDIA specifies each DGX H100 as 8U, air-cooled and up to 130.45 kg. Four therefore occupy 4 × 8U = 32U and total 4 × 130.45 = 521.8 kg in specified system weight, before the cabinet and other equipment. Verify rack and floor loading, available space, airflow and operating temperatures with the site team. Separately obtain the applicable tariff, billing period and commercial treatment of facility overhead. NVIDIA emphasizes that data centers have distinct power, cooling and space constraints; the numerical electricity example alone cannot establish deployment readiness. [1.2] [1.4] [2.2] [4.2]

Electrical Specifications — NVIDIA DGX SuperPOD: Data Center Design Featuring NVIDIA DGX H100 Systems · docs.nvidia.com

Explore next

• How do you measure a multi-feed AI rack’s average and peak input power without double-counting?
• How should switches and shared storage be included in a DGX rack energy budget?
• How do you validate cooling and electrical capacity for a four-system DGX H100 rack?
• How can time-varying workloads and monthly facility energy data improve a rack electricity forecast?

References

Open a reference to inspect the original source.

Relevance scores are model estimates of how well each source addresses this question, not probabilities or fact-check ratings. Reasoned guidance is identified on hover.

1. 99NVIDIA DGX SuperPOD Data Center Designdocs.nvidia.comPrimary NVIDIA source for the DGX H100 maximum system rating and deployment specifications.The NVIDIA DGX SuperPOD with NVIDIA DGX H100 systems is an optimized system for multi-node DL and HPC. It consists of between 31 and 127 DGX H100 systems (Figure 1), with a total of 1,016 NVIDIA Hopper GPUs.›
2. 99Best Practices Guide for Energy-Efficient Data Center Designwww.energy.govPrimary US Department of Energy reference defining PUE, its annual basis and its interpretation.8 Benchmarking Energy efficiency metrics/key performance indicators (KPIs) and benchmarks can be used to track the performance of—and identify potential opportunities to reduce energy use in—data centers.›
3. 91Introduction to NVIDIA DGX H100/H200 Systemsdocs.nvidia.comPrimary NVIDIA documentation for physical specifications and qualified power cords.Power Specifications# ; 200-240 volts AC. 10.2 kW max. 3300 W @ 200-240 V, 16 A, 50-60 Hz ...›
4. 89NVIDIA DGX H100/H200 User Guidedocs.nvidia.comPrimary NVIDIA user-guide excerpt supporting multiple power supplies and balanced load distribution; compressed table formatting warrants care.The DGX H100/H200 system contains six power supplies with balanced distribution of the power load. Table 3: Table 3. Power Specifications Input Specification for Each Power Supply 200-240 volts AC 10.2 kW max. 3300 W @ 200-240 V, 16 A, 50-60 Hz›
5. 57Empirically-calibrated H100 node power models for accurate AI training energy estimationiopscience.iop.orgUseful context about node components and IT electricity, but supplied passages contain no applicable measured H100 average.suited to parallel computation hardware. Transformer architecture DNNs benefit from a variety of parallelization techniques including data parallelism, pipeline parallelism (PP), and tensor parallelism (TP) 9 [7], which unlock increased compu tational performance.›
6. 43Can Your Racks Support NVIDIA DGX H100 Systemswww.sunbirddcim.comSecondary vendor guidance about monitoring and capacity planning; primary NVIDIA evidence is preferable for specifications.The DGX H100 system's power usage is 10.2 kW maximum. each power supply is 3300 W at 200-240 V, 16A, and 50-60Hz. The DGX H100 has a maximum ...›
7. 35Planning a Data Center Deploymentdocs.nvidia.comPrimary NVIDIA planning context, but the referenced rack-power table is not present in the supplied passage.Use the values provided in Table 2 to assist in calculating the power requirements for these racks.›
8. 25H100 GPUwww.nvidia.comPrimary NVIDIA GPU information, but GPU performance claims do not establish complete rack electricity.Product Specifications ; Max Thermal Design Power (TDP), Up to 700W (configurable), 350-400W (configurable) ; Multi-Instance GPUs, Up to 7 MIGS @ 10GB each, Up to ...›
9. 20GPU Power and Cooling Calculator - SLYDslyd.comThird-party calculator with different product examples and acknowledged sourcing limitations; unnecessary for this calculation.Estimate GPU server, rack, cooling, and facility power. Use this calculator to determine power, cooling, and electrical requirements before procurement. PUE ...›
10. 8Nvidia HGX H100 system power consumption : r/HPCwww.reddit.comForum question provides no verified measurement or additional primary support.I am wondering, Nvidia is speccing 10.2KW as the max consumption of the DGX H100, I saw one vendor for an AMD Epyc powered HGX HG100 system ...›

Generated with Ithy Pro. Editorial arithmetic and primary-source checks: 2026-10-23. AI-assisted; this is not a named human expert review. References open the original sources.