Back to Data Center
Data Center tools

Data Center Power Cost & TCO Calculator

A real data center total cost of ownership (TCO) model has to go beyond hardware cost to capture the full operating picture: electricity/energy cost, cooling load, PUE overhead (the energy spent on cooling and power delivery beyond the IT load itself), idle or underutilized capacity, and multi-year operational cost. Many hardware-focused cost estimates miss these, which is why AI infrastructure TCO is often dramatically higher than the server sticker price suggests — the racks are only the start of what you pay for. This calculator captures energy cost, cooling overhead via PUE, and operating cost over a multi-year horizon to give a defensible total-cost view. It starts from rack density and count to size the IT load, applies PUE to account for the cooling and facility overhead the grid actually has to deliver, prices that draw at your negotiated electricity rate, and projects the result across the asset life with annual escalation. The output is the energy-and-cooling operating cost that dominates a running facility’s TCO — the line that decides where AI capacity gets built and whether a site pencils out beyond hardware.

IT load
1,200kW

racks × kW

Grid draw
1,560kW

IT × PUE

Year-1 power cost
$1,024,920

8,760 hours

Total over horizon
$5,441,439

with escalation

Results update live as you type. For planning and field-check estimates — always verify against applicable standards and equipment ratings.

How we calculate this →

Why the power bill decides where AI gets built

A defensible total cost of ownership (TCO) model for AI infrastructure goes beyond hardware cost to capture the full operating picture — and the core equation is short: rack kW × rack count × PUE × 8,760 hours × rate. It compounds fast. Twenty racks at 60 kW with a 1.3 PUE draw 1.56 MW from the grid — about 13.7 GWh a year — so every cent per kWh in the rate is worth roughly $137,000 annually. That is why AI capacity concentrates where industrial power is cheap and abundant, and why a facility's PUE and its negotiated rate matter more to lifetime economics than most capex line items.

PUE overhead is the energy spent on cooling and power delivery beyond the IT load itself — chillers, CRAH/CRAC units, UPS losses, humidification, and building services. Multiplying IT load by PUE scales the server draw up to the power the grid actually delivers and you actually pay for, which is why a hardware-focused cost estimate that ignores cooling and PUE overhead understates a data center's true operating cost. Idle capacity compounds the gap further: racks provisioned for peak training runs still draw power and still need cooling when they sit underutilized, so the energy and cooling bill accrues against capacity you are not always using.

Over a 5–15 year horizon, escalation dominates. Utility industrial rates have historically risen 2–4% a year, and constrained markets are moving faster; modeling a flat rate flatters the pro forma. Large loads can push back — through fixed-price PPAs, hedges, time-of-use scheduling of deferrable training jobs, or co-located generation — but every one of those levers gets negotiated against the baseline this calculator produces.

Power and cooling together are typically 40–60% of an AI data center's total cost of ownership once the facility is running, ahead of staffing and maintenance — operational cost, not hardware, is where the money goes over a multi-year horizon. If you are underwriting a site, run the sensitivity: rate ±2¢ and PUE ±0.15 usually swing the model more than anything else on the sheet.

Frequently asked questions

What should a data center TCO model include besides hardware?

A defensible data center total cost of ownership (TCO) model goes beyond hardware cost to capture the full operating picture: energy cost (the electricity the IT load actually consumes), cooling load and the PUE overhead it creates (the energy spent on cooling, UPS losses, and power delivery beyond the IT load itself), idle or underutilized capacity that still draws power and cooling, and multi-year operational cost including staffing, maintenance, and rate escalation. Server capital is only the starting point — for an operating AI facility, energy and cooling typically run 40–60% of total cost of ownership, ahead of hardware refresh and staffing. A model that counts the racks but not the power bill, the cooling overhead, or the idle capacity understates what the site actually costs to run.

How does PUE overhead affect data center total cost of ownership?

PUE (Power Usage Effectiveness) overhead is the energy the facility spends on cooling and power delivery beyond the IT load itself — chillers, CRAH/CRAC units, UPS losses, humidification, and building services. A PUE of 1.3 means the grid delivers 30% more energy than the servers consume, all of it billed to you. In a TCO model you multiply IT load by PUE to scale server draw up to the power the utility actually meters, so PUE overhead directly inflates the energy cost line — and because cooling load scales with rack density, high-density AI training racks make PUE the single most leveraged assumption in the model. Dropping PUE from 1.5 to 1.2 cuts the energy and cooling bill by 20% every year for the life of the asset, which is why liquid cooling and efficient mechanical design show up as TCO wins, not just engineering choices.

Why is AI infrastructure TCO higher than the hardware cost alone?

Because the hardware is only a fraction of what an AI data center costs to run. AI infrastructure pushes rack densities far above general-purpose computing — 40–100+ kW per rack for GPU training clusters — which multiplies both the energy cost and the cooling load required to keep the racks within thermal limits. That cooling load shows up as PUE overhead on every kWh the servers draw. GPU racks also spend significant time idle or underutilized between training runs, yet still consume power and still require cooling, so idle capacity inflates the operating bill against capacity you are not always monetizing. Add multi-year rate escalation, staffing, maintenance, and facility amortization, and the operational cost over a 5–15 year horizon routinely dwarfs the server sticker price — which is why a hardware-focused estimate misses most of the true TCO.

How do you build a defensible TCO model for AI data center infrastructure?

Start with the IT load — rack kW × rack count — then apply PUE to capture the cooling and power-delivery overhead the grid actually delivers, price that draw at your negotiated electricity rate, and project it across the asset life with annual escalation. That gives you the energy-and-cooling operating cost that dominates a running facility’s TCO. Layer in idle-capacity assumptions (how often the racks sit underutilized and still draw power), staffing and maintenance as recurring operational cost, hardware refresh cadence, and facility capex amortized over its design life. Then run the sensitivity: rate ±2¢ and PUE ±0.15 typically swing the model more than any other line, so a defensible TCO is one that stress-tests those two assumptions rather than quoting a single point. This calculator produces the energy-and-cooling core of that model; pair it with the PUE, cooling-load, and data-center-load calculators to fill in the components.

How much electricity does a 1 MW data center use per year?

At a PUE of 1.0 a continuously running 1 MW IT load would consume about 8,760 MWh (8.76 GWh) per year — 1 MW × 8,760 hours. Real facilities draw more: apply the PUE to capture cooling and overhead, so a 1.3 PUE site pulls closer to 11.4 GWh a year for the same IT load.

Why multiply by PUE in a cost model?

The utility bill covers total facility draw — IT equipment plus cooling, UPS losses, and building services — not just the servers. Multiplying IT load by PUE scales it up to the power the grid actually delivers, which is what you are billed for.

What share of data center TCO is electricity?

For an operating AI facility, power is typically 40–60% of total cost of ownership, ahead of staffing and maintenance. That share rises with rack density and higher energy rates, which is why the rate and PUE assumptions dominate a site's lifetime economics.