The quickest way to a working app on a rented GPU is a template: an application that launches already installed, instead of an empty Ubuntu machine you have to set up yourself. DeployGPU currently offers three of these, plus a plain GPU virtual machine you reach over SSH instead of a link. Pick one on the deploy page after choosing any VM offer, the same way you would choose a disk size.
What each template installs
| Template | Runs | Reached by |
|---|---|---|
| JupyterLab + PyTorch | quay.io/jupyter/pytorch-notebook:cuda12-latest | A private link |
| ComfyUI | deploygpu/deploygpu:comfyui | A private link |
| Open WebUI + Ollama | ghcr.io/open-webui/open-webui:ollama | A private link |
| GPU VM (labelled Bare Metal in the dashboard) | The provider's own Ubuntu image with drivers and Docker | SSH |
All three app templates run on QuantaCloud, Massed Compute, Verda and Hyperstack, and open on a private HTTPS link at a random subdomain of app.deploygpu.ai that only the DeployGPU account that launched the deployment can reach, signed in, in a browser. None of them is tied to a particular GPU: the dashboard lets you put any template on any offer, so it is on you to match the GPU to the model or workflow you plan to run.
Why an app takes longer than a plain VM
Reaching Running takes about 1 to 3 minutes for a plain GPU VM and about 6 to 8 minutes for an app template, because DeployGPU waits for the application itself to answer before it marks the deployment ready, instead of only checking that the machine booted. That wait is the platform doing the health check for you, rather than leaving you to guess whether ComfyUI or JupyterLab has actually started.
Choosing between an app and a plain VM
An app template is the right pick when the tool you want is exactly one of the three: JupyterLab, ComfyUI or Open WebUI with Ollama. Sign-in, the private link and the container are already wired together. A GPU VM is the right pick for everything else: a different notebook server, a training script with no UI, or an inference server such as vLLM. None of those exist as templates here. The self-hosting guide and the vLLM guide cover running them yourself on a plain VM.
What stopping does to any of them
None of the four survive a stop. Stopping terminates the instance and deletes its disk, including anything a template installed while it ran, so move results off the machine, or onto a volume on Verda or Hyperstack, before you stop. The stop guide has the checklist.