GPU CAPACITY · ON-DEMAND, COMMITTED OR SPOT
Cloud GPU Pricing Comparison
Enter the hourly rate from your own quote and compare a month of on-demand, committed and spot GPU capacity for the hours you actually use.
Your experiment
Start with 8 GPUs at $4 an hour, 12 hours a day, a 40% commitment discount and the plan to commit. Raise the hours to 14 and 15, mark the workload interruptible, choose spot, then raise the interruptions and the checkpoint interval.
Every input recomputes the result immediately; there is no animation because nothing here unfolds over time. An input outside its allowed range is rejected with a message and the previous valid result stays on screen.
Computed data
Metrics
This page contains no provider prices: the $4 default only stands in for the rate in your own quote, and discounts vary by provider, region, GPU and term. A commitment is modelled as paying for every hour of a 730-hour month at the discounted rate; spot as the discounted rate with interruptions that each lose half a checkpoint interval plus the restart time. Storage, networking, minimum terms and availability are not included, and spot capacity can be unavailable for long periods.
Start from your own quote
The page has no price list: enter the rate you were quoted in Your on-demand rate, and every bill scales with it. Hours used per day and Days used per month set the useful GPU-hours, which is all that on-demand pays for. The opening case is 8 GPUs at $4 an hour, used 12 hours a day for 30 days, 2,880 useful GPU-hours. On-demand you pay only for those hours, $11,520 a month. A commitment with a 40% discount pays $2.4 an hour, but for every hour of the month, 5,840 GPU-hours, which is $14,016: PLAN OVERPAYS by $2,496 a month, 21.7% more than on-demand.
The break-even utilisation
A commitment wins only when you use the capacity often enough. The break-even utilisation is one minus the discount: 60% for a 40% discount. Twelve hours a day is 49.3% of a 730-hour month, below it. At 14 hours, 57.5%, the commitment still overpays by $576; at 15 hours, 61.6%, it becomes the cheapest at $14,016 against $14,400 on-demand. A deeper Commitment discount lowers the bar: at 50% the break-even is 50% and the opening case overpays by only $160.
When spot fits
Spot capacity is only an option for work that checkpoints and can be interrupted, which is why choosing spot in You plan to buy for the opening workload gives SPOT NOT SUITABLE. Tick Workload checkpoints and can be interrupted and spot, at the default Spot discount of 65%, costs $4,086 a month and becomes the cheapest. With Spot interruptions per 100 hours at 2, each interruption loses half of a 60-minute checkpoint interval plus the Restart time after an interruption, 10 minutes, so 40 minutes, and the 360 hours of work need 364.9 hours of wall-clock time.
When spot fails
Interruptions and long checkpoint intervals compound. At 10 interruptions per 100 hours the work needs 385.7 hours and costs $4,320. At 30 interruptions with a 240-minute checkpoint interval each interruption loses 130 minutes and the work would need 1028.6 hours, more than the 720 hours in 30 days: SPOT CANNOT FINISH. Checkpointing every 15 minutes instead keeps the loss at 17.5 minutes per interruption: the work needs 394.5 hours and the bill is $4,419. At 24 hours a day there is no slack left in the month, so even 2 interruptions per 100 hours give SPOT CANNOT FINISH, while 0 interruptions cost $8,064.
Reading the tool and its limits
The bars compare the three monthly bills, the lanes show utilisation against break-even and the spot arithmetic, and the table gives hours paid, rate, monthly cost and cost per useful GPU-hour. Real commitments run for one or three years and may cover a region or a family of instances; spot prices move and capacity can disappear for days. Use the page to see which way the decision goes, then check it against your provider's terms.