Hourly GPU rental: how to calculate the cost
An hour of a GPU server costs from about $0.73 for an L4 or RTX 4090 to $3.30 for an H100, if you divide the monthly price by 730 hours. Hourly billing pays off for training, experiments and one-off jobs lasting a few days. For a server that runs all month it is almost always dearer. Below: the formula, four worked examples and the charges people forget.
With hourly billing you top up a balance and the host deducts the cost of the server for every hour it runs; clouds sometimes charge for every second. Amazon EC2, for example, bills per second with a one-minute minimum for each start. Delete the server and the charges stop. How this works for ordinary VPS is covered in the article when hourly billing pays off. With GPUs the logic is the same, only the sums are dozens of times larger and a mistake in the calculation costs more.
What an hour costs
I took the plans with one whole GPU from the accsly catalog on 13 November 2026, found the median monthly price and divided it by 730, the number of hours in an average month. That is how most hosts that show both prices calculate the hour. The catalog stores prices in roubles, so the dollar figures are converted and rounded.
| Card | Month, market median | Hour |
|---|---|---|
| L4, 24 GB | $532 | $0.73 |
| T4, 16 GB | $549 | $0.75 |
| RTX 4090, 24 GB | $559 | $0.77 |
| L40S, 48 GB | $929 | $1.27 |
| A100, 40–80 GB | $1,099 | $1.51 |
| RTX PRO 6000, 96 GB | $1,499 | $2.05 |
| H100, 80 GB | $2,400 | $3.29 |
Which card to choose for the job is covered in the article which GPU for which job.
Who offers hourly billing
In the catalog 61 hosts rent out servers with a GPU. Hourly billing is confirmed at 23 of them, and they account for half of all GPU plans. Six hosts definitely sell only by the month; the rest do not publish their terms. Cloud and VPS plans with a card: GPU servers; all hosts with hourly billing are on the page VPS with hourly billing.
On expensive cards you pay extra for flexibility. The median monthly price at hosts with hourly billing is higher than across the market: L40S $1,146 against $929, RTX PRO 6000 $1,725 against $1,499, H100 $2,755 against $2,400. The difference is 15 to 23%. On inexpensive cards there is none: the T4 and RTX 4090 cost the same or even slightly less at such hosts.
The formula
The cost of a job has three parts, and the first is almost always underestimated.
cost = price per hour × (preparation + work + margin) + disk storage between runs
- Preparation. Creating the server, installing the driver and libraries, downloading the model and data. A 70-billion-parameter model in 16-bit format weighs about 140 GB. Over a 1 Gbit/s link that is at least 19 minutes, and in practice longer. The card sits idle all that time, but the meter runs.
- Work. The time of the job itself. If you do not know it, first run a short pass on a small part of the data and multiply.
- Margin. A bug in the code, a restart, running out of memory halfway. Add at least a quarter on top.
To see whether hourly or monthly rental is cheaper, divide the monthly price by the hourly price. You get the number of hours a month after which the hourly server is dearer.
break-even hours = monthly price / price per hour of the hourly plan
Four worked examples
Fine-tuning a small model over a weekend
Fine-tuning an 8-billion-parameter model with LoRA needs about 16 GB of video memory, so a 24 GB card is enough. An RTX 4090 at hosts with hourly billing has a median of $506 a month, which is $0.69 an hour. Two hours of preparation, ten of training, three of margin: 15 hours, about $10. A month of the same card costs almost 50 times more.
A three-day experiment with a large model
You need to see how a 70-billion-parameter model handles your data. That needs an 80 GB card, an H100. At hosts with hourly billing its hour costs $3.77. Three days non-stop are 72 hours, about $270. A month of an H100 at the market median is $2,400. Even if the experiment stretches to a week, hourly rental comes out almost four times cheaper.
A model for a team during working hours
The model answers employees from 9 to 6 on weekdays and runs on an L40S. At hosts with hourly billing an hour costs $1.57. Nine hours on 22 working days are 198 hours, about $310 a month, if the server is deleted at night and recreated in the morning from a disk snapshot. A monthly L40S at the market median costs $929. Break-even: $929 / $1.57 ≈ 590 hours, almost 25 days. As long as you need the server for less than three weeks a month, hourly billing is cheaper.
A service that runs around the clock
Here hourly billing loses. 730 hours of an L40S at $1.57 come to $1,146, while a monthly plan at another host is $929. Paying in advance is cheaper still. In the catalog 164 GPU plans from 18 hosts show a yearly price: the median discount is 15%, a quarter of the plans give 32% or more, and 36 plans give none. With a 15% discount a year of an L40S comes to about $9,500 instead of $11,150. But pay for a year only to a host where the service has already run for at least a month.
What you pay for while the card is idle
- A stopped server. Stopping does not always stop the bill. In the big clouds a stopped server is no longer charged for processor and card, but disks and external IP addresses still are, as both AWS and Google Cloud state. At many hosts a stopped GPU server costs as much as a running one, because the card stays reserved for you. Check this in the terms before the first start.
- Idle time between steps. The card waits while you read logs, fix code or download results. Prepare data and scripts on a cheap server without a GPU and create the GPU server when everything is ready to run.
- Disk and snapshots. Between runs the model and data are kept on a disk or in a snapshot. It is cheap, but it adds up over months.
- Setup fee. 24 GPU plans in the catalog have a one-off setup payment, almost all of them dedicated servers. On a short rental it can cost more than the work itself.
- Minimum term. Some hosts rent dedicated GPU servers for at least a month. Hourly billing usually comes with cloud servers; dedicated ones are here: dedicated servers with GPU.
How not to overpay
- Save training checkpoints to disk every half hour to an hour. If the server crashes or the balance runs out, you lose an hour of work, not a day.
- Set up spending and low-balance notifications in the host's panel. When the money runs out the host stops the server and a few days later may delete it with its data.
- Put the date the server should be deleted into its name. An H100 server forgotten for a month costs about $2,400.
- First run the job on a cheap card with a small part of the data, and move to an expensive one once the code works.