Skip to content
100% local

GPU compute cost calculator

Calculate GPU run costs and compare configurations side by side.

Input
0
100

GPU compute cost calculator

Paste one or more GPU configurations and this tool calculates the total run price, the price per hour and the price per GPU-hour for each, then lines them up cheapest first for an easy comparison. Each line follows name;GPU count;hours;hourly rate, for example "A100 run;8;10;2.50" for eight A100 GPUs running ten hours at $2.50 per GPU-hour. Use it to budget a training job, check a cloud invoice, or weigh an on-demand instance against a spot one.

The options model the levers that move a GPU bill. Spot/preemptible discount cuts the hourly rate by a percentage. Average GPU utilization doesn't change what you're billed — you still pay for wall-clock time — but it recomputes the price per GPU-hour of useful work, so a job whose GPUs are only 60% busy shows its true cost per productive hour. The storage & transfer surcharge adds a flat amount for the data movement a run also pays for. Knowing a job's throughput, set "Convert price to" and enter steps or tokens per second: the tool reprices the run per 1,000 steps or per 1,000,000 tokens, the number that predicts what a model actually costs to train or serve.

Switch between a compact comparison table and an itemized breakdown per configuration; both respect the currency and decimal places you choose. A line that doesn't parse — a missing field, a name with no price — is skipped and counted separately, and the tally reports how many configurations and invalid lines were found.

Everything runs in your browser. Instance names, prices and run details are never uploaded — the calculation happens locally, so it's safe to paste real budget numbers. Copy the result, download it as a .txt file, or send it back into the input to refine a configuration and recompute instantly.

FAQ

What format does each configuration line need?
Four fields separated by semicolons: name;GPU count;hours;hourly rate. For example "H100 spot;4;10;4.80" prices four H100 GPUs running for ten hours at $4.80 per GPU-hour.
How is the spot/preemptible discount applied?
It reduces the hourly rate by the chosen percentage before the total is calculated — a 20% discount on a $5.00 rate prices the run at $4.00 per GPU-hour.
What does the utilization percentage change?
It doesn't change your bill — you still pay for the full wall-clock time — but it recomputes the price per GPU-hour of GPU work that was actually productive, which is the number that matters when GPUs sit partly idle.
How do I get a price per token or per step?
Set "Convert price to" to steps or tokens and enter your run's throughput per second. The tool multiplies that by the total run time to get the total units processed, then divides the total cost by that.
Is my pricing data uploaded anywhere?
No. Every calculation runs entirely in your browser — the configurations and prices you paste never leave your device.