Popular cards for servers, live prices
Live prices: the cheapest free server in the catalog, updated automatically
GPU server rental on gpu.nz means a Linux server where one or more NVIDIA GPUs (up to eight) are given to you whole, with root access over SSH, a disk and a network connection. You pick the number of cards, the server country and the rental type in the catalog, and pay by the minute while the server runs.
A GPU server fits when a single home graphics card is not enough or you need a machine that runs around the clock: training and fine-tuning models, an API service for a neural network, rendering, batch data processing. It also suits people who set up the environment themselves instead of taking a ready app.
A server is Linux (Ubuntu) with the NVIDIA driver and CUDA. You get root access over SSH, a disk, a network connection and whole GPUs: from 1 to 8 per server, depending on the offer. The country where the server stands is shown in the catalog before you start it.
Partly. The GPUs are dedicated to you: while the server runs you do not share them with other customers. But the physical machine can be larger than your rental, and other customers use its remaining GPUs.
It is also not a full virtual machine: you get a container with no kernel of its own, no Windows and no display output. If you need a whole physical server or Windows, this format is not for you. For training, inference and compute on Linux it is a good fit.
In short
A GPU server on gpu.nz is a Linux server with one or more whole NVIDIA GPUs, root access over SSH and a disk. Servers with 24 GB or more per GPU start from $0.241 per hour per card, billed per minute from your balance.
A full day on the cheapest card of 24 GB or more: about $5.78
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
| Setting | What to look at |
|---|---|
| Video memory | The model has to fit in the card's memory, so start there. The VRAM calculator works out how much a given model needs. |
| GPU count | One card is usually enough for inference and experiments. Several are for training large models and parallel jobs. |
| CPU, RAM, disk | Data preparation and loading weights depend on cores, system memory and disk, not on the GPU. The catalog has a filter for each. |
| Country | It affects latency to you and the speed of downloading data. Every offer shows its country. |
| Rental type | An on-demand server is not interrupted. An interruptible one costs less, but the host may stop it if another customer bids more. |
One card handles most jobs. You need several when the model does not fit in the memory of one card, or when training has to be faster and the data is split between the cards. The guide to GPUs for AI helps with the choice, and training jobs are covered on the fine-tuning page.
How fast the cards talk to each other depends on the machine: servers with SXM cards are usually faster at it than cards on a plain PCIe bus. The command below on a running server shows how the cards are connected.
nvidia-smi topo -mFor a service that must answer all the time (a model API, a bot, a queue worker), take an on-demand server: it is not interrupted. In the catalog you can keep only verified servers and sort them by reliability. For training with checkpoints and for experiments an interruptible server pays off: it costs noticeably less, and the disk stays when it is stopped. How it works is in the guide.
Money is charged per minute, and the hourly price is fixed at launch. A stopped server is not charged for the GPU, but its disk is billed as storage until you delete the server; the console shows what storage costs before you stop.
Watch your balance
When the balance runs out, servers are deleted together with their disks. We warn you by e-mail or in Telegram while it runs low, but keep a reserve for servers that run for days.
Add your public SSH key in the console and connect as root; the steps are in the guide to SSH access. Servers can be started and deleted from code: there is a REST API, a CLI and an MCP server for AI agents.
Live prices: the cheapest free server in the catalog, updated automatically