Skip to content
DCIM ProfessionalsDCIM Professionals

Cost & Energy Tool

AI / GPU Infrastructure Power & Cost Calculator

Facility power, cooling and running cost for GPU/AI workloads — GPU wattage × count × server overhead × PUE, rack density with an air/liquid-cooling verdict, and annual cost at your own electricity rate. Scoped for a handful of GPU servers up to a multi-rack GPU cluster, not gigawatt hyperscale.

Updated: July 9, 2026Planning reference

Quick Answer

GPU wattage alone underestimates power draw by roughly half — and rack density decides whether air cooling still works.

Enter your GPU model (or custom watts), count, utilization and PUE. The calculator builds up from raw GPU power to node power (server overhead) to facility power (PUE), then checks kW per rack against the point where air cooling stops being enough.

GPU & Workload

GPU deployment

Scope: a handful of GPU servers in a server room up to a ~20-rack GPU cluster — not gigawatt-campus hyperscale.

= 32 GPUs total

GPUs rarely sustain 100% during training, but AI inference workloads commonly run near-constant. Overhead ~1.8× covers the non-GPU node power — CPUs, NVLink, RAM, PSU losses — an 8×H100 node draws closer to 10 kW than the raw 5.6 kW of GPU wattage alone.

Facility & Density

PUE and rack layout

1.3–1.5 is typical air-cooled; ~1.15 is typical for direct liquid cooling. Measure yours with the PUE Calculator.

Electricity

Running cost

Your all-in business rate incl. taxes & grid fees — it's on your electricity bill. Same rate logic as the Energy Cost Calculator.

Not a hardware-pricing tool

This calculates power, cooling and running cost — not GPU purchase price. GPU model presets set typical wattage (TDP) only; hardware prices change too fast and depend on vendor and contract, so enter your own figures elsewhere for a purchase-cost comparison.

Formula / Method

How the calculator works

Raw GPU power is GPU wattage (TDP) × GPU count × utilization. That's multiplied by a server overhead factor (~1.8, covering CPUs, NVLink, RAM and PSU losses) for node power, then by PUE for facility power — the number that actually shows up on your electricity bill. Node power divided by rack count gives kW per rack, checked against the point where air cooling stops being enough (~30 kW/rack). Facility power × hours/year × your rate gives annual running cost.

Facility kW = TDP × count × utilization × overhead × PUE

Inputs

Required inputs

GPU model or custom wattage, GPUs per server and server count, utilization, server overhead factor, PUE, rack count, your electricity rate and currency, and hours of operation per year (default 8,760 = 24/7).

Uses

Where this helps

Sizing power and cooling capacity before a GPU server or cluster purchase, checking whether existing racks and cooling can host a planned deployment, budgeting the annual electricity line for an AI project, and deciding early whether liquid cooling needs to be part of the build.

FAQ

AI / GPU power notes

Why multiply GPU wattage by ~1.8?

An 8-GPU server isn't just the GPUs — CPUs, NVLink (the high-speed link between GPUs), RAM and power-supply losses add roughly 45% more on top. An 8×H100 node draws closer to 10 kW continuously, not the 5.6 kW the raw GPU wattage alone would suggest.

When do I need liquid cooling?

Air cooling can't remove much more heat than about 30 kW per rack, even with good containment. Dense modern GPU racks (NVIDIA GB200 NVL72-class systems, for example) reach 120–140 kW per rack and require direct liquid cooling. This calculator flags the transition automatically from your rack count and node power.

Does this include the GPU purchase cost?

No — this is a power, cooling and running-cost tool, not a hardware-quote tool. GPU purchase prices change too fast and depend heavily on vendor, volume and contract terms to bake into a planning calculator.

Learn More

Reference links

Project

Planning power and cooling for a GPU / AI deployment?

ITCOREOPS can support power and cooling sizing, rack density planning and documentation for on-prem GPU infrastructure projects.

Open project formVisit ITCOREOPS →

Disclaimer

Planning reference only

Planning-level model — real GPU draw varies with workload, and actual server overhead and PUE depend on your specific hardware and facility. Validate against vendor specs and a facility power audit before committing to a build.

Workflow

Check power and cooling before committing to a GPU server or cluster purchase.

Use this calculator before signing hardware quotes, so rack density and the cooling verdict — not just GPU count — drive the facility decision.

Use exact inputs ↑

Support

Need help planning power and cooling for GPU infrastructure?

For GPU deployment power sizing, cooling capacity checks and operational infrastructure support, visit ITCOREOPS.

Visit ITCOREOPS →Email ITCOREOPS ✉

Feedback

Found a bug or calculation issue?

Report display problems, calculation issues or suggestions for improving this calculator.

Report a bug ✉

Worked example

Inputs: 32× NVIDIA H100 (700 W, 8 GPUs × 4 servers), 100% utilization, 1.8 server overhead, PUE 1.4, 4 racks, €0.30/kWh, 8,760 h/yr.

Result: GPU power 32 × 700 W = 22.4 kW → × 1.8 overhead =40.3 kW node → × 1.4 PUE = 56.4 kW facility. Density: 40.3 / 4 racks ≈10.1 kW/rack → Air cooling OK. Cost: 56.4 × 8,760 × €0.30 ≈ €148,000/yr(≈€12,300/mo). A denser single-rack build — e.g. 8× B200 in one rack — still lands at only ~14 kW/rack here, but real dense AI racks (NVIDIA GB200 NVL72-class, ~72 GPUs) reach 120–140 kW/rack, where this calculator's verdict correctly flips to liquid cooling required.