AI: cloud or owned hardware?
Buy or rent? An interactive calculator shows how utilization, demand peaks, and idle time change the costs.
Buy or rent? Try it.
Example: many AI requests during the day, fewer at night. Both options provide the same assumed computing capacity. Move the sliders and watch how demand peaks and idle time change the costs.
Cloud
€438/ month
€15,779 over 36 months
Owned hardware
€522/ month
€18,789 over 36 months
Owned systems needed: 1 · Upfront purchase: €12,000 · Monthly electricity: €39.
Buying would break even against renting after about 48 months.
More assumptions and calculation
Month = 730.5 hours. Buying: hardware sized for peak demand, spread over the period, plus power and maintenance. Renting: active or reserved GPU hours × hourly price. Power: active hours × working power plus remaining hours × idle power. Prices are adjustable example assumptions; taxes, financing, labor beyond the maintenance allowance, network costs, and resale value are excluded.
Start with the workload
Imagine your team uses a lot of AI in the morning and afternoon, but very little at night. One GPU can complete a certain amount of work per hour. Here, 100% demand means one GPU is fully busy. 150% needs more than one GPU. The calculator compares the same assumed computing capacity on both sides.
Buying: pay for the peak
Owned hardware needs enough capacity for the busiest hour. If you use 30% of a GPU on average, you still pay its entire purchase price. Once peak demand goes above 100%, this example needs a second system. Power and operating costs also continue while idle.
Renting: the billing model matters
With active compute billing, batched jobs use only the GPU hours they need. An always-on cloud instance also costs money at night. The calculator lets you switch between these assumptions. Autoscaling, minimum billing, loading times, and latency need separate verification in a real offer.
What goes into the calculation
The calculator spreads purchase costs over the chosen period, adding power and monthly maintenance. Working power and idle power are counted separately. Renting multiplies GPU hours by the hourly price. The break-even point compares the upfront purchase with the difference between rent and the running costs of owned hardware.
A cost model is a starting point
The sliders are example assumptions, not current provider quotes. Hardware performance depends on the model, precision, memory, batch size, and response-time requirements. Privacy, availability, staffing, networks, and maintenance also matter. Use your own measurements and prices for a concrete decision.
My notes
Notes and highlights for this post
Stored only in this browser. Nothing is sent.
Your notes and highlights appear here.
Select a passage and choose “Highlight” or “Note”. Or write a note about the whole post.