A GPU virtual machine with SSH access

The template labelled Bare Metal in the dashboard is, on every provider DeployGPU launches on today, a virtual machine: the provider's own Ubuntu 22.04 image with NVIDIA drivers and Docker already installed, reached over SSH instead of a private link.

Why the dashboard says Bare Metal

The name is a holdover, not a description of the hardware underneath. QuantaCloud, Massed Compute, Verda and Hyperstack all provision this template as a virtual machine, the same way they provision an app template, just without an application layered on top. This page calls it a GPU VM and leaves the dashboard's own label alone, since renaming it here would not change what you get. Every other entry in the templates section is a VM underneath as well. This one is simply the version with nothing preinstalled on top of it.

What is on the machine at first boot

Every provider starts you on its own Ubuntu image with the NVIDIA driver installed. The dashboard lists the images that also come with Docker and picks one for you, so on most offers a GPU workload can run in a container from the first SSH session. The API uses each provider's default image unless you pass os_image_id. The login user and the disk differ by provider:

The dashboard shows the exact ssh command for your deployment. Copy it rather than retype it, since the username is the detail people get wrong most often.

The template does not care which GPU sits underneath it. Pick the RTX A6000 above for a general-purpose start, or open the marketplace and choose any offer instead: a plain VM runs the same way on all of them.

Launch it

1. Click the button above, or open the marketplace, pick any VM offer and click Launch.
2. On the deploy page, choose Bare Metal under Template, the default for a VM offer.
3. Check the cost summary and the SSH key, then deploy. The first hour is reserved from your credit balance.
4. Wait on the Deployments page until the status reads Running, typically 1 to 3 minutes, faster than an app template because nothing has to answer a health check first.
5. Copy the SSH command from the deployment and connect. There is no Open Application button on this template, since there is no app and no private link, only SSH.

The first commands to run

nvidia-smi first, to confirm the driver loaded and the card matches what the offer described:

nvidia-smi

Then a smoke test that Docker can actually reach the GPU, with an official CUDA image rather than anything built by hand:

sudo docker run --rm --gpus all nvidia/cuda:12.2.2-base-ubuntu22.04 nvidia-smi

That tag is listed on Docker Hub's nvidia/cuda repository as of 28 September 2026. It targets CUDA 12.2, which the driver on Hyperstack's default image supports and every newer driver does too. If you pick a newer tag, keep its CUDA version at or below the CUDA Version the host's nvidia-smi reports, because the container runtime refuses to start an image that asks for more than the driver supports.

The firewall caveat that catches people

Bind services to 127.0.0.1 and reach them over an SSH tunnel instead:

ssh -N -L 8000:127.0.0.1:8000 ubuntu@YOUR_VM_IP -p 22

Or run ufw and allow only what is meant to be public:

sudo ufw allow OpenSSH
sudo ufw enable

Allow SSH before enabling ufw, not after. Enabling it without that rule first locks out the same connection needed to fix it, with no console access documented to undo it remotely, so the only way back in is a fresh instance.

When something goes wrong

For anything else, the troubleshooting section and the contact page are the next stops.

What it costs

You pay the hourly rate of the GPU offer you launch on, shown live for every family on the GPUs section, and nothing extra for the template itself, since it is only the provider's own image. A job that runs 5 hours and 5 minutes bills five full hours plus the started sixth, then refunds the unused part of that sixth hour when you stop. Credits are bought by card, with a 5% + $0.30 processing fee per top-up. The pricing page has the rest of the mechanics, and the API covers launching and stopping VMs like this one from a script instead of the dashboard.

Where this leads next

A plain VM is the base for anything that is not one of the three app templates: a training job with no UI, a different notebook server, or an inference server such as vLLM. The SSH guide covers keys, ~/.ssh/config and VS Code Remote-SSH in more depth than fits here, and the vLLM guide walks through the next step for serving a model.