How It Works

Aquanode splits one thing most GPU clouds treat as inseparable: the machine you rent and what you built on it. What you built on it further splits into two: what you installed, and what you made.

The machine is a GPU card at a provider. It is billed by the second, only while it runs, and it is disposable. The pod is not tied to the card it started on, and it is made of an Environment (the base image, your packages, your startup script: what you installed) and a Volume (/workspace: your code, checkpoints, and datasets, what you made).

Everything below follows from that split.

Launching

You pick an Environment and a Volume, and Aquanode rents the machine on your behalf. Providers differ in their APIs, their request formats and their lifecycle rules, so a launch is not a single call to a single vendor: Aquanode holds one request shape and translates it per provider, which is why the same pod can land on very different hardware.

There is no separate "turn this box into something reusable" step. The pod exists from the first launch.

Stopping

A Stop runs two captures, not one: the environment (everything outside /workspace) goes to its own storage location, and the volume (/workspace) goes to its own, separate location. Only after both are confirmed does the machine get released back to the provider, which is what stops the GPU charge. Storage keeps billing, because your data has to live somewhere.

This is why Stop and Start work on every provider we support, including ones with no native suspend: it is two captures and a terminate, then a fresh box restored from those captures.

Saved while the pod runs

Between stops, the volume saves itself in the background so a sudden provider kill loses only the interval since the last save, not everything since the last Stop. The environment is not saved this way: it is only captured at Stop, because installs do not usually change minute to minute the way a training run's checkpoints do. See Stop, start, and move for the exact guarantee, including what a background save can catch mid-write.

Starting, anywhere

Starting rents a new machine and restores both captures onto it. The card, the region, and the provider can all differ from where the pod last ran. That is the same mechanism whether you are coming back from a Stop, moving to cheaper capacity, or leaving a provider that ran dry, which is why there is no separate migration feature to learn. A restore in progress shows on the pod until it is done; the pod is Running only once it is usable, never on the first reachable signal.

Aquanode keeps a small agent on each running box. It reports health and metrics, and it does the work of capturing and restoring your environment and volume. That agent is why a Stop is a real capture of the disk rather than a provider snapshot you cannot move.

Environments become shareable on demand

Sharing or keeping an environment freezes it as a numbered version. Only that one version leaves when you share it; a recipient's new pod gets that environment plus a brand new, empty volume, so sharing a pod never leaks your data. See Sharing an environment.

Serving an endpoint

An endpoint puts a stable, callable address in front of an image or a saved pod version. Calls queue, and the number of running instances follows the queue depth up to a limit you set. When the queue empties, instances scale back down, to zero by default, so an idle endpoint is not holding a GPU.

Because a cold call has to rent a machine rather than wake up an already-running one, the first call after an idle period takes real time, not milliseconds. Endpoints are built for work where that is acceptable, or you can keep one instance warm to avoid it. There is no per-run dollar cap: the bound that always applies is wall-clock (time limit times attempts times machines), and an optional monthly budget can sit on top of that.

See Endpoints for the full flow.

What this buys you

One account, one bill, and one set of keys across every provider we support, with the environment and data you built surviving each machine they run on.

See which providers and GPUs this covers on the Aquanode Marketplace.