Buying hardware vs renting cloud GPUs

Renting is cheaper if you use a GPU now and then. Owning is cheaper if you use it a lot. This guide shows how to find your own break-even point.

The break-even sum

Divide the purchase price by the hourly rate of a comparable cloud GPU. The result is the number of hours of use after which owning has paid for itself.

A worked example with made-up numbers: a $20,000 card, compared with a cloud GPU at $3 an hour, breaks even after about 6,700 hours. Running around the clock, that is roughly nine months. At 40 hours a week it is a little over three years. Put in the real rate you are quoted, in Australian dollars and including GST, because the answer moves a lot with that one number.

Costs people forget when buying

• The workstation or server the card goes into.
• Electricity and cooling.
• Your own time to set it up and maintain it.

Costs people forget when renting

• Storage and data transfer charges on top of the hourly rate.
• Paying for idle time when an instance is left running.
• Exchange rate movements if you are billed in US dollars.
• Availability: the GPU you want is not always free when you want it.

Reasons that are not about cost

• Data control. With your own hardware, sensitive data never leaves your premises.
• Predictability. A one-off purchase is easier to budget for than a variable monthly bill.
• Flexibility. Cloud lets you scale up for a short burst without buying anything.

A simple rule

If your GPU would be busy most of the working week for a year or more, price up owning. If your use is occasional or very bursty, renting usually wins. Many teams own a baseline and rent for the peaks.

Want help with the sums for your workload? Book a consultation.