HOME / COMPUTER SCIENCE & DATA / AI INFRASTRUCTURE & LLM ECONOMICS / AI INFRASTRUCTURE TOTAL COST OF OWNERSHIP (ON-PREM VS CLOUD GPU) CALCULATOR
Computer Science & Data / AI Infrastructure & LLM Economics

AI Infrastructure Total Cost of Ownership (On-Prem vs Cloud GPU) Calculator

Buying the hardware

Host servers, switches and storage on top of the cards; 30–60% is common

Period the hardware cost is spread over; 3–5 years

Running it

Include its share of the server; ~0.7 kW for an H100 SXM card alone

Total facility power ÷ IT power; 1.2–1.6

Colocation or rack space, support contracts, staff time

A month has about 730 hours

Renting instead

Use your actual quote: on-demand, reserved and spot prices differ a lot

Monthly cost

Owning is cheaper by $2,857 a month

On-prem
$11,543
Cloud
$14,400
On-prem per GPU-hour
$3.61
Cloud per GPU-hour
$4.50

Buying pays for itself after about 36 months, within its 4-year life. The two cost the same at 319 busy hours per GPU a month (44% utilization); busier than that favors owning.

Cumulative cost over 4 years

0510152025303540450100000200000300000400000500000600000700000
On-prem (purchase + running costs)Cloud rental

Months on the horizontal axis.

On-prem total
$554,063
Cloud total
$691,200

On-prem breakdown

Up-front hardware$392,000
Hardware spread over 48 months$8,167
Electricity (3,136 kWh incl. cooling)$376
Fixed costs$3,000
Monthly total$11,543

Idle power, resale value, financing costs, taxes and the cost of the engineers who run the cluster are not included; add staff time to fixed costs. Cloud egress and storage fees are not included either.

Compare buying GPUs with renting the same GPU-hours from a cloud provider. The calculator spreads the hardware cost over its useful life, adds electricity (with cooling, via PUE) and fixed costs, and shows the monthly cost of each option, the cost per GPU-hour, how many months the purchase takes to pay back and the utilization at which owning becomes cheaper.

Formula

Hardware=GPUs×price×(1+overhead)\text{Hardware} = \text{GPUs} \times \text{price} \times (1 + \text{overhead})
Owned per month=Hardware÷(12×life)+GPUs×kW×PUE×hours×rate+fixed\text{Owned per month} = \text{Hardware} \div (12 \times \text{life}) + \text{GPUs} \times \text{kW} \times \text{PUE} \times \text{hours} \times \text{rate} + \text{fixed}
Cloud per month=GPUs×hours×cloud price\text{Cloud per month} = \text{GPUs} \times \text{hours} \times \text{cloud price}
Payback months=Hardware÷(Cloud per month−electricity−fixed)\text{Payback months} = \text{Hardware} \div (\text{Cloud per month} - \text{electricity} - \text{fixed})

Worked Example

GPUs8
Price per GPU$35,000
Servers, networking, storage40% of GPU cost
Useful life4 years
Power per GPU0.7 kW at a PUE of 1.4
Electricity$0.12 per kWh
Busy hours per GPU400 a month
Fixed costs$3,000 a month
Cloud price$4.50 per GPU-hour
Result
Hardware$392,000
Owned per month$11,543 ($3.61 per GPU-hour)
Cloud per month$14,400
Paybackabout 36 months
Break-even utilization319 busy hours a month (44%)

Spread over four years the cluster costs $8,167 a month, plus $376 of electricity (3,136 kWh including cooling) and $3,000 of fixed costs. Renting the same 3,200 GPU-hours costs $14,400, so owning saves $2,857 a month and the $392,000 purchase pays back in about 36 months, inside its four-year life. Below about 319 busy hours per GPU a month (44% utilization) renting would be cheaper.

Frequently Asked Questions

At what utilization is owning GPUs cheaper than renting?

When the monthly cost of owning, which barely changes with use, drops below the cloud bill, which grows with every GPU-hour. The calculator solves for the busy hours per GPU where the two match; with the example inputs that is about 319 hours a month, or 44% utilization. Clusters that sit idle most of the month are almost always cheaper to rent.

Why does the payback period differ from the monthly comparison?

The monthly comparison spreads the purchase over the hardware's useful life. Payback counts cash: you pay for the hardware up front, then each month save the cloud bill minus your electricity and fixed costs. If payback takes longer than the useful life, renting is the better deal.

What is PUE and why does it matter?

Power usage effectiveness is total facility power divided by the power the IT equipment draws. At a PUE of 1.4, every kWh the GPUs use needs another 0.4 kWh for cooling and power conversion, so it scales the electricity bill. Staff time, financing, resale value and cloud egress fees are not modelled; add staff time to the fixed costs.