Methodology · updated October 6, 2026

How the estimates are built—and where they stop.

This is a transparent first-pass capacity model. It applies arithmetic to power, heat, rack dimensions, and limits you supply; it does not simulate a workload, electrical distribution system, cooling plant, or code requirements.

Choose one power boundary

A hardware preset is either power per accelerator, input power for a complete system, or power for a complete rack system. The calculator multiplies only per-GPU values by the GPU count. A complete-system value already includes its internal components, so adding its GPUs or a second server-power allowance would double count them. Network, storage, and site equipment outside the quoted system still need a separate allowance.

For per-GPU presets, the forward calculation is IT watts = (GPU watts × GPUs per system + auxiliary watts per system + other IT watts per system) × system quantity. For a complete-system or rack-system preset, its stated power replaces the per-GPU term. The separately entered additional IT load is available in the GPU-only path.

Translate IT load to facility capacity

Facility design kW = IT design kW × PUE. PUE is a ratio of total facility energy to IT equipment energy; it is not a percentage added to the GPU alone. The calculator uses one supplied PUE as a constant for this estimate. A real facility's PUE can vary with load, weather, equipment, and measurement boundary.

Capacity reserve is treated as the fraction deliberately left unused. With reserve r, usable capacity is limit × (1 − r), while the required capacity for a known load is load ÷ (1 − r). A 15% reserve therefore means dividing by 0.85; it does not mean adding 15% of the load. Reserve applies to power only. Supply cooling and rack limits that already include the headroom you need.

Worked comparison: GPU-only vs complete system

Eight 700 W accelerators plus 1,200 W of server auxiliary load and 800 W of other IT load produce 7.6 kW IT. At PUE 1.2, that is 9.12 kW facility design load; with 15% reserve, plan for 10.73 kW of facility capacity. The 7.6 kW IT input corresponds to about 2.16 refrigeration tons of heat in this model.

A DGX H100 preset instead uses the manufacturer's 10.2 kW maximum complete-system input for one 8-GPU, 8U system. With the same PUE and reserve, its example is 12.24 kW of facility design load and 14.4 kW of required facility capacity. Do not add another 5.6 kW for its eight 700 W GPUs: those accelerators are already inside the complete-system boundary. The calculator's preset table links each model to its manufacturer reference and shows the date its value was last checked.

Heat-equivalent cooling estimate

The estimate treats the entered IT electrical load as heat: BTU/h = IT watts × 3.412142. Refrigeration tons are BTU/h ÷ 12,000. For the example above, 7.6 kW is about 25,932 BTU/h, or 2.16 tons. This is a heat-equivalent rate, not the electrical input or rated capacity of a chiller. It also does not model airflow, supply-water temperatures, heat recapture, liquid-cooling loops, distribution losses, redundancy, or ambient conditions.

For some products, the vendor publishes a separate thermal-output figure. For example, NVIDIA's DGX H100/H200 system guide lists 38,557 BTU/h of heat output, while the 10.2 kW input preset converts to about 34,804 BTU/h with this generic formula. Treat those as different source-specific inputs and boundaries to investigate; use the exact model's validated thermal data for cooling design rather than assuming this simplified conversion replaces it. The guide was last updated January 26, 2026.

Rack packing uses whole systems

Systems per rack are limited by both usable rack units and the IT kW/rack limit. Each limit is rounded down to a whole system; the tighter limit determines the plan. Rack systems occupy their configured rack units as one unit. The power-density value is an IT-side limit, separate from PUE and facility input. A rack-system preset should be compared with a rack limit large enough to contain that rack's stated IT power.

Reverse sizing: the tightest supplied limit wins

Reverse mode converts the available power into usable IT power (including reserve and, for facility input, PUE), then independently calculates whole systems permitted by power, cooling, rack space, and rack power density. It returns the smallest whole-system count and names every constraint tied at that minimum. An empty rack or unused cooling margin does not relax a separate power limit. A zero result means at least one supplied limit cannot fit a whole system.

Electrical-current estimate

For a single-phase load, the estimate is I = P ÷ (V × PF). For balanced three-phase load, it is I = P ÷ (√3 × V × PF). These equations assume the entered power, voltage, and power factor represent the same system boundary. The result is not a breaker or conductor selection: phase balance, harmonics, continuous-load rules, redundancy, voltage tolerance, and local electrical code require a qualified design.

Energy and cost estimate

Average facility kW is multiplied by 730 hours for an average month or 8,760 hours for a year. Energy-only cost is the resulting kWh multiplied by the entered flat rate. Real bills may include demand charges, time-of-use rates, taxes, fees, power-factor penalties, and billing periods that differ from these averages. No price, utilization, or PUE preset is a vendor quote or measured operating result.

Source checks and change limits

Hardware figures link to manufacturer product pages or documentation. Each preset identifies its power basis, displayed note, source type, and last-verified date; a maximum or design figure is not a measured workload average. Vendor configurations change, so confirm the linked model documentation and the exact equipment variant before using a figure in a budget. The calculations intentionally avoid filling in specifications that the cited source does not give.

Results are estimates for early planning. Validate input boundaries and the selected equipment configuration, then have qualified electrical and mechanical engineers verify capacity, protection, cooling, redundancy, and applicable code requirements before procurement or installation.