A GPU run costs the time between launch and stop at the VM's hourly price, and that time is your hours of work plus the minutes the VM takes to boot. The calculator applies QuantaCloud's billing rules to live prices and gives you the cost of one run, the cost of a month of runs, the credit you need to launch, the same run on every other GPU at the same count and, if you have a reserved quote, the monthly hours below which on demand costs less.
- Cost per run
- $5.31
- 2 h 3 min from launch to stop at $2.59 per hour
- Cost per month
- $5.31
- 1 run, 2 h 3 min billed
- Credit to launch
- $2.59
- One hour at this configuration's price
| GPU | Per hour | Per GPU-hour | Cost per run | Cost per month |
|---|---|---|---|---|
| 1x RTX 6000 Ada | $0.79 | $0.79 | $1.62 | $1.62 |
| 1x L40S | $1.09 | $1.09 | $2.23 | $2.23 |
| 1x A100 SXM4 80GB | $1.50 | $1.50 | $3.08 | $3.08 |
| 1x RTX PRO 6000 Blackwell | $2.17 | $2.17 | $4.45 | $4.45 |
| 1x H100 PCIe (selected) | $2.59 | $2.59 | $5.31 | $5.31 |
Prices checked 14 Nov 2026, 09:16 UTC
How the calculator works#
The billing rules set the formula. The first hour is charged at launch, each further hour when the previous one is used up, and the unused seconds of the current hour are refunded when you stop. The clock starts when you click Deploy, and the provisioning minutes count once the instance is running, so boot time is billed like any other minute. The net effect is per-second billing with a one-hour prepay: a run costs its exact time at the hourly rate.
You set the GPU and how many of them, the hours of work in one run, the minutes each launch takes to boot, the runs per month and, if you have one, the monthly price from a reserved quote. Boot starts at 3 minutes for 1 to 4 GPUs and 10 minutes for 8: across QuantaCloud, most single-GPU VMs are running in about 3 minutes (median), and 8-GPU VMs take about 10. App templates, such as ComfyUI or PyTorch + Jupyter, take longer while the app's container image is pulled, so raise the boot time when you launch one.
| Output | How it is worked out |
|---|---|
| Cost per run | (Hours of work + boot minutes / 60) x the VM's hourly price |
| Cost per month | Cost per run x runs per month |
| Credit needed to launch | One hour at the VM's hourly price, the charge made at launch |
| Every other GPU | The same run on each other GPU at the same count, at its live price |
| Break-even with a quote | Monthly quote / hourly price: the monthly hours below which on demand costs less, also as a share of a 730-hour month |
Worked example: a 2-hour job on one H100 PCIe
One H100 PCIe cost $2.59 an hour on 2026-09-27 (today: $2.59/GPU-hr). Take a job with 2 hours of work, a 3-minute boot and 20 runs a month. Each run is billed for 2 hours 3 minutes, and the charges go like this (our calculation):
| Time after launch | What happens | Balance change |
|---|---|---|
| 0:00 | Launch: hour 1 charged | -$2.59 |
| 0:03 | The VM is running and the job starts | $0.00 |
| 1:00 | Hour 2 charged | -$2.59 |
| 2:00 | Hour 3 charged | -$2.59 |
| 2:03 | Job done, stop: the 57 unused minutes of hour 3 refunded (57/60 x $2.59) | +$2.46 |
| Net | 2.05 hours x $2.59 | $5.31 |
One run costs $5.31, exactly 2 hours 3 minutes at $2.59 an hour. Twenty runs cost $106.19 a month (our calculation: 20 x 2.05 x $2.59), and the credit needed to launch is $2.59, the first hour. Boot is 1 of the 41 hours billed each month, $2.59 of the total. Short, frequent runs feel it most: the same 3 minutes add 2.5 percent to this 2-hour run and 15 percent to a 20-minute one.
The comparison with every other GPU keeps the hours the same, so it shows the price gap between GPUs, not the speed gap. A faster GPU finishes the same work in fewer hours, and a GPU with more memory can serve more requests at once, as the KV cache guide explains. Estimating fine-tuning cost shows how to get the hours from tokens per second or a short test run, and the GPU catalog compares the cards themselves.
On demand or a reserved quote
The break-even is one division: a monthly quote divided by the hourly price is the number of hours a month below which on demand costs less. The calculator also shows it as a share of a 730-hour month, which is 8,760 hours a year divided by 12. The job above keeps one H100 PCIe busy for 41 hours a month, 5.6 percent of the month, so a quote for the same GPU would have to come in under $106.19 a month to cost less. In the other direction, one H100 PCIe running all 730 hours costs $1,890.70 a month on demand at the 2026-09-27 price (730 x $2.59), so a quote above that never costs less, however busy the GPU is. Reserved vs on-demand GPUs works the same division through two worked examples and covers what else a commitment changes.
Price is one side of the choice. Reserved capacity is built to order: we order and build the hardware to your spec, confirm the configuration, lead time and terms in writing, and hold it for you for the term. On-demand capacity is whatever the live catalog shows when you launch. The reserved capacity page compares the two side by side. Buying the hardware is the third option: AI server cost prices servers to buy, with power and cooling, and finds the break-even against renting.
What it leaves out#
The calculator prices the time between launch and stop at the live rate, and four things sit outside that.
Storage lasts only while the instance runs. The disk is fixed by the configuration and included in the hourly rate, but stopping an instance terminates it and deletes the disk, and there are no volumes or snapshots. Every run starts clean, so downloading the model and data again belongs in the hours of work, and so does copying results off before you stop.
Time around the job bills like the job. Installs, evaluation, a VM left running after the job ends and a failed first attempt, such as a vLLM out-of-memory error at startup, all count. The auto-stop guide stops an instance through the API once its GPUs go idle.
Prices move with capacity. The calculator works from the prices checked at the time its stamp shows, and the deploy page shows the rate again before you click Deploy.
Launches that never run cost nothing. A launch you stop before it is running, or whose provisioning fails, is refunded in full, and the calculator assumes every launch runs.
FAQ#
Do I pay for a whole hour if I stop early?
No. The first hour is charged at launch and the unused seconds come back when you stop, so a 10-minute test on a VM that took 3 minutes to boot costs 13 minutes at the hourly rate.
How much credit do I need?
One hour of the configuration to launch: $2.59 for one H100 PCIe on 2026-09-27 (today: $2.59/GPU-hr), and the smallest deposit is $5. After that, each hour is charged as the previous one ends, and if your balance cannot cover it, the instance is terminated and its disk deleted. QuantaCloud emails you when the balance falls below $2. My rule for a long run is to fund the whole estimate plus one hour, or to turn on auto top-up, which the billing docs walk through.
Are the disk and data transfer extra?
No. The hourly rate covers the whole VM, with its GPUs, vCPUs, RAM and disk, and there are no egress or ingress charges for data moving in or out. The disk is deleted when you stop, so anything you want to keep has to leave the VM first.
My rule: price the month before you decide how to pay for it. If the hours come to a small share of a 730-hour month, stay on demand, because you pay only from launch to stop. If you would keep the same GPUs busy most hours of most weeks, or your plan depends on the hardware being there on a date, send a capacity brief and set the written quote against the break-even above. To check which GPUs your model fits before you price them, use the LLM VRAM calculator, which sits with this one on the GPU planning tools page, or the tables in how much VRAM you need and the LoRA and QLoRA VRAM guide. Pricing has every billing rule.
Launch an H100 PCIe