What a pod is
A pod is a GPU plus the config to run it. It holds an Environment (what you installed) and a Volume (what you made). The GPU is rented. The pod is yours.
A pod is a GPU box you build once and keep. Run it on any provider we support, or on hardware you already own, stop it, and start right where you left off.
The machine underneath is rented. The pod is not. That split is the whole product, and a pod is made of exactly three things:
- The machine is a GPU card at a provider, billed by the second, only while it runs. It is disposable. When you stop paying for it, it goes away.
- The Environment is everything outside
/workspace: the base image, the packages you installed, the startup script. What you installed. - The Volume is
/workspace: your code, checkpoints, and datasets. What you made.
Could you reinstall it? Environment. Would losing it hurt? Volume.
That is the whole mental model. A venv inside /workspace is data and travels with the
volume, not with a shared environment. Every environment we ship points its caches
(HF_HOME, TORCH_HOME, pip's cache) into /workspace for you, so a model you pull down
lands on the right side of the line without you having to think about it.
You never create a pod separately
The pod exists from the first launch. There is no second step, no "convert this box into a template" action. You launch, you work, and the thing you built is already a pod.
What a pod carries
Open any pod in the console and the Overview tab lists exactly what it holds:
| Field | What it means |
|---|---|
| Environment | The name it started from, and its version if it has one |
| Volume | The volume attached, and its size |
| Saved | How long ago the volume's contents were last written to storage |
A pod has no version history of its own. Rolling back what you installed means picking an older Environment version; rolling back your data means picking a Volume history point. Neither touches the other. See Environment and Volumes.
The states a pod can be in
| State | What is true |
|---|---|
| Running | A machine is rented and billing by the second. You can connect to it. |
| Stopped | No machine is rented. Nothing is billing hourly. Your Environment and Volume are held. |
| Starting / Stopping | A transient state while a machine is being rented or released. |
A stopped pod is not a deleted pod. We do not make pods ephemeral, because no GPU cloud does and people read "Stop" as "Delete" if it acts like one.
What it costs to keep one
A stopped pod rents no machine, so the hourly charge stops entirely. Saving while you work is free. You pay only to hold your Environment and Volume, billed at $0.02/GiB per month.
That is the reason stopping is the default answer to "I am done for today" rather than tearing the box down and rebuilding tomorrow.
Where to go next
Launch a pod
Pick an Environment, pick or attach a Volume, pick a machine
What travels with a pod
Environments, versions, and rolling one back
Stop, start, and move
Stop the hourly charge without losing anything, and what changes if a provider kills the box instead
Do it from the terminal
The aq CLI drives the same pods