GPU VPS by the hour

The GPU VPS you rent on DeployGPU is a virtual machine with 1, 2, 4 or 8 NVIDIA GPUs, its own vCPUs, RAM and disk, and SSH access with sudo, rented by the hour from prepaid credits. The offers come from QuantaCloud, Massed Compute, Verda and Hyperstack, start at 48 GB of memory per GPU, and launch from the marketplace or the API. These are the starting prices right now:

What you get with a GPU VPS

Each offer is one VM with a fixed shape: the GPU model and count, vCPUs, RAM, disk and region, all listed on the offer before you launch. The provider's Ubuntu image comes with NVIDIA drivers and Docker installed, and the deploy page pre-selects an image with CUDA and Docker. You get a shell with sudo, so you install what you need yourself, or you pick an app template on the deploy page: PyTorch + Jupyter, ComfyUI or Open WebUI with Ollama. The GPU VM page covers the plain VM, which the dashboard labels Bare Metal even though it is a VM.

The login user depends on the provider, and the dashboard shows the exact ssh command for each deployment:

A GPU VPS, a GPU dedicated server or a container

People who search for a GPU dedicated server, a GPU cloud server or a VPS with a GPU usually want the same thing: a Linux machine with a GPU that they control. A DeployGPU VM gives you that by the hour, with no monthly term and no minimum. If you came to rent a GPU server for AI work, the difference to plan around is persistence. Stopping a VM deletes it and its disk, so data you want to keep goes to your own storage or, on Verda and Hyperstack, to a volume. The stop guide covers what survives a stop.

A fixed term or more than 8 GPUs goes through the cluster request form, and a person answers it. Containers are the other common option, and the VM vs container comparison shows when a container is enough.

How a GPU VPS is billed

You pay the hourly rate on the offer from prepaid credits. Launching reserves the first hour, each new hour is charged when it begins, and when you stop, the unused seconds of the current hour go back to your balance. A launch that fails, or one you stop before it is Running, is refunded in full. Credits are added by card from $5 with a processing fee of 5% + $0.30 per top-up, or in USDC through Stripe with a fee of 1.5% + $0.30. The pricing page has worked examples.

Prepaid credit doubles as a spending cap. Without auto top-up you cannot spend more than you added, and when the balance runs low the VM switches to per-minute billing until it reaches zero, with emails along the way.

1, 2, 4 or 8 GPUs per VM

Offers come in fixed sizes of 1, 2, 4 or 8 GPUs, and the GPUs inside one VM talk over NVLink or PCIe depending on the card and the board. These are the multi-GPU VMs you can launch right now:

The multi-GPU page explains the interconnects, and the SXM, PCIe and NVL page explains the H100 and H200 variant names. The smallest GPU VPS has one 48 GB card, such as an RTX A6000, RTX 6000 Ada, L40 or L40S, and the 48 GB GPUs page compares them. When the GPU you need is out of stock, Auto Provision deploys it when it comes back.

Where the VMs run

VM offers are in 3 countries: the United States (QuantaCloud, Massed Compute and Hyperstack), Canada (Hyperstack) and Finland (Verda, in Helsinki), and the region is listed on every offer. Stock moves between them, so check the live table rather than a fixed list. There are no Asia-Pacific VM offers.

What a GPU VPS here does not do

• There is no pause that stops billing. Stop deletes the VM and its disk. On Verda and Hyperstack, the API can power a VM off and on, but billing continues while it is off.
• There is no firewall. Hyperstack VMs are created with inbound ports open, so a service bound to 0.0.0.0 is on the internet. Bind services to 127.0.0.1 and reach them through an SSH tunnel.
• There are no spot or reserved prices you can buy yourself, and no SLA.

Launch a GPU VPS in 5 steps

1. Sign up with Google, GitHub, an email link or a password. No card is needed.
2. Add credits from $5 on the Billing page.
3. Pick an offer in the marketplace.
4. On the deploy page, keep the GPU VM or choose an app template, check the SSH key and launch.
5. Copy the ssh command from the dashboard once the VM is Running, which usually takes 1 to 3 minutes.

The rent a GPU guide goes through each step in more detail, the first five minutes guide covers what to check once you are in, and the API page shows how to launch the same VMs from a script.