Renting a graphics card on gpu.nz means a server with an NVIDIA GPU that you start from the browser and pay for by the minute. Pick a card in the catalog, connect over SSH or open a ready AI app a few minutes later, and stop or delete the server when you are done, so the charges stop.
In short
Renting a graphics card on gpu.nz: a server with an NVIDIA GPU from $0.055 per hour, billed per minute from your balance. It usually starts in 2–5 minutes; connect over SSH or open a ready AI app in the browser.
You rent a graphics card when you have none or yours has too little memory: to run a language model, generate images and video in Stable Diffusion or ComfyUI, train a LoRA, render a scene or run an experiment in Jupyter. You get the whole card: all of its memory and compute are yours while the server runs.
Video memory comes first: the model or scene has to fit on the card. Speed comes second: a card without enough memory runs many times slower or does not start the job at all. The VRAM calculator works out how much memory a given model needs.
| Job | Memory | Cards to look at |
|---|---|---|
| Stable Diffusion images, small LLMs up to 8B | 12–16 GB | RTX 3060, RTX 4060 Ti, RTX 4080 |
| FLUX, ComfyUI, 14–32B LLMs at int4, LoRA training | 24–32 GB | RTX 3090, RTX 4090, RTX 5090 |
| Video, 70B LLMs at int4, large batches | 48–96 GB | RTX A6000, L40S, RTX PRO 6000 |
| Training and serving large models | 80–141 GB | A100, H100, H200 |
If unsure, start with a 24 GB RTX 4090 or RTX 3090: most image models and mid-size language models run on them. Every card's price is on the pricing page.
The price is per hour of one card and is charged per minute while the server runs. It is fixed at launch and does not change while the server is on. A stopped server is not charged for the GPU but keeps its disk with your files, and the disk is billed until you delete the server. How the money is counted is explained on the per-minute rental page.
How to save
Interruptible servers cost noticeably less than on-demand ones: the host may stop such a server if another customer bids more. They pay off for experiments and jobs that save their progress; the guide to interruptible servers explains how they work.
These are headless servers, so a rented card is no good for gaming: there is no Windows and no display output. Cryptocurrency mining is forbidden by the terms of service. Everything that runs on Linux and CUDA — AI models, training, rendering, compute — runs without limits.
One hour on the cheapest card: about $0.055
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