GPU Cost Calculator
Illustrative rates only. The hourly rates below are example figures for education, not live offers or quotes. Real GPU cloud prices vary by provider, region, and availability and change frequently. Use the provider directory to check current pricing, and always confirm with a provider before making decisions.
Example hourly rates used by this calculator
These are illustrative example rates chosen for teaching, not market data and not offers. They represent a plausible mid-range of what on-demand-style GPU rental might look like; spot pricing can be much lower and reserved pricing different again. Learn more in our guide How GPU Cloud Pricing Works, and check real rates in the provider directory.
| GPU | Memory | Example low ($/hr) | Example high ($/hr) |
|---|
How to use the estimate well
The calculator turns four inputs into a cost range. Two of them are usually wrong the first time, and knowing which two makes the result far more useful.
- Hours are almost always under-estimated. People enter the time a single clean run would take. Real work includes set-up, a failed run or two, evaluation and the idle hour before someone notices the job finished. For training and fine-tuning, multiply your single-run estimate by three to five; our guide on estimating GPU hours explains why. For serving, hours means hours the replica exists, not hours it is busy.
- The utilization buffer is where honesty lives. The default assumes a moderately efficient job. If your data pipeline is untested or you develop on the same instance you train on, raise it. Teams that measure their utilization are often surprised to find it below 50 percent; see why your bill is higher than your compute.
- GPU choice should follow memory, then cost. Pick the smallest card that holds your model at the precision you will actually use. The fine-tuning memory guide and the serving guide give the arithmetic for each case; quantization can move you a size class down.
- The pricing model changes the risk, not just the rate. The spot multiplier is attractive, but it assumes your job can survive interruption. If it cannot, use the on-demand figure and read spot instances and checkpointing before you change your mind.
Run the estimate three times: at your expected inputs, at double the hours, and at a higher buffer. If the worst case is still affordable, proceed. If it is not, the calculator has done its job before you spent anything. Remember that the hourly rates below are illustrative; replace them mentally with the figures from the provider you will actually use, found through the provider directory.
How the estimate is computed
- Effective GPU-hours = hours × quantity × (1 + utilization buffer ÷ 100).
- Model multiplier = 1.0 (on-demand example), 0.4 (spot example), or 0.65 (reserved example) — illustrative only.
- Low / high estimate = effective GPU-hours × the GPU's example low / high rate × model multiplier.
The buffer exists because real workloads are rarely 100% efficient: you will spend time on setup, debugging, checkpointing, and occasional idle instances. The range exists because real prices vary widely — treat the result as a planning aid, not a quote.
A fully worked sample calculation
Sample inputs only, chosen so you can check the arithmetic against the formula above: GPU = NVIDIA H100 SXM, pricing model = on-demand example, hours = 100, GPUs = 8, utilization buffer = 20%. Effective GPU-hours = 100 x 8 x 1.20 = 960. The calculator's example rate for this GPU is $2.50 to $4.50 per hour, so the estimate is 960 x $2.50 = $2,400 at the low end and 960 x $4.50 = $4,320 at the high end. Those rates are this page's teaching examples, not quotes. Before budgeting real money, replace them with a current figure from a provider's own pricing page, for example Amazon EC2 on-demand pricing or Google Cloud GPU pricing, and re-run the same arithmetic.