A cloud GPU is an NVIDIA graphics card in a data center that you reach over the internet: instead of buying one, you rent a server with the card for the minutes you need. On gpu.nz you pick a model in the catalog, open an app in the browser or connect over SSH a few minutes later, and stop or delete the server when you are done, so the charges stop.
In short
A cloud GPU is an NVIDIA graphics card in a data center that you rent by the minute instead of buying. On gpu.nz cards with 16 GB of memory or more start from $0.055 per hour per GPU, and you work from a browser or over SSH.
You want a cloud GPU when your laptop or home PC has no strong GPU, when its memory is too small for your model, or when you need a card for a project rather than for good. You get the whole card: all of its memory and compute are yours while the server runs.
Owning a graphics card means buying it, finding room in the case, a strong power supply and cooling. A cloud GPU is the same kind of NVIDIA card in a data center: you switch a server on for as long as you need the card and pay for the minutes. Your laptop stays an ordinary laptop, and all the load lands on the server.
The VRAM calculator works out how much memory your model needs, and every card's price is on the pricing page.
Buying wins when the card is busy almost all the time: if it runs around the clock for many months, a single purchase costs less than many hours of rental. If you need a GPU a few hours a week, for the length of a project, or while you are still choosing a model, the cloud is cheaper. The card's price is only part of it, since power, cooling, space and idle time add up, so count the full cost of ownership: the formula and a worked example are in the guide Buy an H100 or rent one.
A sensible order is to rent first, measure how many hours a month the card really works and how much memory is enough, and only then decide whether to buy your own.
You do not need a powerful computer, only a browser and an internet connection. There are three ways to work with the card.
You can also manage servers from code: there is an API, a CLI and an MCP server for AI agents.
Your data lives on the server's disk. Stop the server and the GPU is no longer charged, the files stay, and the disk is billed as storage until the server is deleted; the console shows the storage price before you stop. Delete the server and the charges end, and the files are erased.
When the balance runs out
If the balance runs out, servers are deleted together with their disks. You are warned in advance by e-mail or Telegram while the balance runs low, but it is safer to copy your results to your own computer right after the work.
This is a compute card, not a game console and not a remote desktop. The server is a Linux (Ubuntu) container: it has no Windows, no display output and no kernel of its own. A cloud graphics card for games is a different kind of service.
| Job | A fit? |
|---|---|
| AI models: chat, images, video, speech recognition | Yes |
| Training and fine-tuning models | Yes |
| CUDA compute and rendering, if the program runs on Linux | Yes |
| Games | No: there is no Windows and no display output |
| Windows programs and a remote desktop | No |
| Cryptocurrency mining | Forbidden by the terms of service |
Half an hour on the cheapest cloud card: about $0.028
Estimate at the current per-GPU hourly price. Billed per minute, the price is fixed at launch.
Prices and availability come from the live catalog and update automatically
RTX 309024 GBfrom $0.241/h
RTX 409024 GBfrom $0.308/h
RTX A600048 GBfrom $0.58/h
A100 SXM480 GBfrom $1.55/h
H100 SXM80 GBfrom $4.33/h
H200 SXM141 GBfrom $7.53/hLive prices: the cheapest free server in the catalog, updated automatically