11 November 2026 · VRAMfarm team

GPU rental billing explained: per minute, 15-minute minimum, traffic and storage

A six-minute test on an RTX 4090 costs €0.15, not six minutes’ worth, because every launch is billed for at least 15 minutes. After that you pay minute by minute, at a price locked when you click Launch. On top of GPU time there are only two other charges: traffic at the machine’s rate, and storage if you save your folder. Preparing the machine before it is ready is not billed at all.

Per minute, with a 15-minute minimum

Every minute a server runs, 1/60 of its hourly price comes off your prepaid balance in euros. Partial minutes round up to the next whole minute. The first 15 minutes of each launch are always billed, even if you stop after one, because the machine has to be rented from its host for that launch either way.

What that looks like on an RTX 4090 at its current price of €0.59 an hour:

Multi-GPU machines cost the per-card price times the number of cards. The minimum applies per launch, so five quick tests in a row cost five minimums. If you are iterating, keep one server running for the session rather than starting a new one for each try.

To start a server, your balance has to cover one hour of everything you would then have running, the new server included. For a single RTX 4090 that is €0.59. This is only a check: the hour is not charged up front, and you pay for the minutes you actually use.

When billing starts: boot, setup and model downloads

Billing does not start when you click Launch. First we find a machine, the host starts the container, and SSH and Jupyter come up. This usually takes 2–5 minutes, and none of it is billed. The clock starts at the moment the server is ready and its connection details appear in the dashboard.

The ready-made apps are different, because they install their models after that moment. Those minutes are billed as normal GPU time, and the download is billed as traffic. The dashboard shows when the app is actually ready to open, so you don’t have to guess:

The setup time depends on the machine’s connection, and the picker estimates it for each machine before you launch. You can use those minutes: upload your LoRAs or input images through Jupyter while the models download.

The price is locked at launch, the catalogue is live

The catalogue price of a card moves during the day. Machines come from many hosts, cheap ones get rented by other people, and others become free. If you leave the page open for a while, the price you see can change before you click.

Once you click Launch, the price per minute shown at that moment is fixed for that server until you stop it. If the catalogue goes up an hour later, you keep your price. If it goes down, you also keep your price; the lower one applies to your next launch. The hourly price of a machine already includes its working disk. If you choose a larger disk than the app needs, the hourly price shown before launch is a little higher.

Traffic and storage

Traffic is every gigabyte the machine downloads or uploads: models from Hugging Face, a dataset you pull in, results you download to your computer. It is billed per GB at the machine’s own rate, shown in the catalogue before you launch. Model downloads count because they reach the machine over the host’s connection like any other download.

Rates differ between machines because each host pays its own data centre for bandwidth. Most are around a cent per GB or less. Some B200 and H200 machines charge more. We favour machines with cheap traffic, and the picker shows what the app’s download will cost on each machine. For a 17 GB Flux download at a cent per GB, that is about €0.17.

Storage is the saved /workspace folder: €0.03 per GB per month, charged daily as a thirtieth of the monthly amount. It is the only thing that can reduce your balance when no server is running. If your balance goes down overnight with nothing running, a saved folder is the reason, and you can see its size and delete it in the dashboard. The details of what is kept are in what happens to your files when you stop a GPU.

When billing stops: stop, idle, zero balance, failures

Billing stops in the minute you click Stop, or stop the server through the API. There is no stopped-but-billed state. With Save and stop, the server keeps running while it uploads /workspace, usually for a few minutes, and billing ends when it stops itself after the upload.

Idle auto-stop. If a server has no SSH or Jupyter connection and no GPU load for about two hours, we email you. If it is still idle an hour later, it saves its folder and stops. In the worst case a forgotten RTX 4090 costs about three idle hours, €1.77, instead of a whole night. You can turn auto-stop off in the settings for jobs with long quiet phases. Servers from your own Docker image are not auto-stopped.

Balance at zero. As the balance gets low, you get warnings in the dashboard and by email, including when less than an hour of runtime is left. When it runs out, the server stops without saving and its disk is deleted. A folder saved earlier stays in storage; if the balance stays at zero, it is kept for 30 days and then deleted, with a notice first. Traffic is counted from the host’s statement shortly after the fact, so a few cents of it can still appear after a stop.

Machine failure. If a machine goes offline at the host, billing stops from the moment it went missing, and you get a notification. You don’t pay for the time it was down.

A full session bill, worked out

You want to try a few Flux LoRAs on an RTX 4090 and keep the results.

The bill:

So the session costs €0.86 in GPU time plus about a fifth of a euro in traffic on a typical machine. The same work on an RTX 3090 at €0.30 an hour would cost €0.44 for the same 87 minutes, but it generates more slowly, so the same hundred images would take longer. Which card is cheaper per image is covered in the GPU guide for Flux and SDXL.

Top-ups and refunds

You top up from €10 by card, including Apple Pay and Google Pay, or by SEPA transfer. There is no subscription and no monthly fee.

You can request a refund of your unspent paid balance in the dashboard once your servers are stopped, and it goes back to the original payment method. Bonus credit is not refundable. Current prices for every card are on the pricing page. Before a long job, check that your balance covers it with some margin, so it doesn’t stop at 3 a.m. with the last hour unsaved.