Rent a GPU for AI: LLMs, images, video and training

Which card to rent for your task

Start from the task, not the card name: it decides how much video memory you need and which template to launch. The table lists typical AI jobs, the memory they need and a ready app to begin with.

If unsure, take a 24 GB RTX 4090 or RTX 3090: they cover most rows of the table except 70B models and heavy video. For those, look at cards with 48 GB and more — the RTX A6000, L40S and A100. Every card's price is on the pricing page.

Why memory matters more than speed

A model that does not fit in the card's memory either fails to start or runs many times slower, because part of the weights has to live outside the GPU. So for AI work you count gigabytes first and look at speed second. The memory for a language model depends on the parameter count, the weight precision and the context length: work it out in the VRAM calculator, and find the rules and examples in the guide to VRAM for LLMs.

24 GB is a sensible starting point; more headroom is for long context, large batches and video. If you are choosing a card to buy rather than rent, read the guide on how to choose a GPU for AI.

Your own model on your own server

When the model runs on your rented server, your requests go to it and not to somebody else's service: you decide which weights to load, what to keep on the disk and who gets access. Web apps open through a protected link carrying your rental's token, and you can log in over SSH with your own key. The server's country is shown in the catalog.

The server is still rented and stands at a host in a data center, so assess data that you are legally required to store in a particular way separately. When you delete the server, the files on its disk are erased. If you want a model without built-in refusals, some ready chats use such weights; the guide to uncensored AI on your own server covers it.

How fast you can start

A server usually starts in 2–5 minutes. The template downloads the model on first start: a small one is ready within minutes, a 60 GB model in 10–30 minutes. A ready chat opens straight into its model, which then stays on the server's disk and is not downloaded again.

If you need your own environment, give your own Docker image instead of a template. You can also start servers from code through the API, the CLI or the MCP server, so an AI agent can bring up a GPU for a job by itself.

How not to overpay

• Take memory for the task, not the biggest card: a model that fits in 24 GB does not need an H100, and an hour on it costs several times more.
• For experiments and training that saves checkpoints, choose interruptible servers: they cost less, but the host may stop one if another customer bids more; the disk stays. The details are in the guide to interruptible servers.
• Stopping ends the GPU charge, but the disk with your models is billed as storage until the server is deleted; the console shows its cost before you stop. If you no longer need the server, delete it, and nothing is charged.
• The hourly price is fixed at launch and does not change while the server runs. When the balance runs out, servers are deleted together with their disks; while it runs low, you get warnings by e-mail or in Telegram.

More tricks are in the guide on saving money on GPU rental.