Home›Products›Persistent Storage
Weights, datasets, checkpoints. A GPU is an hourly resource; your data has to outlive it. This page states what works today, what is planned, and exactly how the bill works.
Status
The one-sentence summary: your pod's disk is persistent today, but a storage object fully independent of the pod is not there yet.
Billing honesty
That sentence is written the same way on the billing screen and under the line-item table in the console. It is our product stance; we do not take it out of the interface, and we do not soften it here either.
Compute and storage bill together. They stay separate line items; neither disappears inside a total. You can see at any moment how much came from which line.
Compute stops, storage keeps billing. The disk still occupies space on a host. That conflicts with the "I stopped it, so I am paying nothing" expectation — which is why it is spelled out on screen in capital letters, not as a surprise at the end of the month.
If you chose to delete the disk with it, the storage charge stops that second and the data is gone. The delete confirmation states plainly whether the disk will be deleted and asks for two-step confirmation. Data deleted by accident cannot be brought back.
When the budget cap is reached, pods are stopped, not deleted — so your disk is kept and the storage charge keeps running. How long a paused pod's disk is stored free of charge has not been decided yet; the proposal we are currently discussing is a fixed free period, then the normal storage rate, then deletion with notice. When that decision is settled this paragraph will be updated with the number — until then we are not inventing one here.
Price
There is no storage fee hidden inside the compute price. The rows below are examples: like marketplace pricing they depend on the supply side and are not final.
| Line item | Unit | How it is measured | Example price |
|---|---|---|---|
| Pod disk (NVMe) | GB / month | Allocated size, prorated per second | $0.10/GB/month (example) |
| Network traffic (egress) | GB | Data flowing out of the pod | illustrative · live in the console |
| Snapshot | GB / month | Not there yet — planned | — |
The figures in this table are illustrative and are not a final price list; the real price appears on the console's billing screen with a live counter. Storage is prorated per second: you do not pay a full month for a disk you kept half a day.
How to use it today
You state the disk size on the rent screen or in the API. It is mounted inside the pod as a single persistent directory; your training output, weights and dataset should be written there. Anything you write to the container's root filesystem is not persistent.
If you are not going to work for the rest of the day, stop the pod: the compute bill stops, the disk stays in place, and the next day you continue where you left off. This is far cheaper than keeping a pod idling for hours.
The storage card on the pod detail page shows used space and fill ratio, and warns you as it approaches full. A full disk ends a long training run at the most infuriating possible point.
You can download it from the Files tab over a signed link. Our advice for today is plain: do not keep the only copy on a single pod disk. Until snapshots ship, backups are your responsibility — not saying so would not have been honest.
Security
Environment variables, API keys and tokens are kept encrypted at the application layer. The host cannot see inside your container. Architecture detail on the security page.
When you pin your pod to a country, its disk stays inside that border as well. The pin is enforced server-side; a UI filter does not count as a security boundary. Details: sovereignty.
Get started
Compute is per-second billing; storage is a separate, visible line item. You see what each row costs from day one.