Anyone who rents GPU by the hour in the cloud knows the routine: you spin up the machine, download the models, set up the environment, get to work — and when you shut it down, you lose everything. Next session, you start from scratch again. That is paid GPU time spent just rebuilding what was already done. The persistent disk fixes this. This guide explains what it is, how it works, and why it cuts your bill.
⚡ Summary
A persistent disk is storage that survives the machine. You store models, datasets and checkpoints on it once and reuse them on any new instance — which lets you power the GPU off when you are not using it (and stop paying for it) without losing anything.
The problem: an hourly machine is ephemeral
A GPU instance rented by the hour is, by nature, temporary. The disk that ships with it — where the system, the packages you installed and the files you downloaded live — is discarded when the machine is terminated. That is great for cost (you only pay while you use it), but it creates friction: every new use means rebuilding the environment.
In practice, that means re-downloading tens of gigabytes of model weights, reinstalling dependencies and reconfiguring everything. On a card that costs a few reais per hour, 20 to 40 minutes of "prep" every session becomes wasted money — and the temptation becomes leaving the machine on just to keep the setup, which costs even more.
What a persistent disk is
The persistent disk is a storage volume that is independent of the machine. It exists on its own: you create it once, at whatever size you want, and it stays there whether the machine is on, off, or already deleted.
Think of it as an external drive in the cloud: the GPU is the computer you use for a few hours; the persistent disk is where your files live, and you plug it into any computer you rent later.
How it works in practice
- Create the disk once. You pick the size (say 50 GB, 500 GB or several TB) based on the volume of your models and data.
- Attach it to any machine. When you spin up a new instance, you associate the disk with it. The data is ready the moment the machine boots.
- Work normally. Save models, datasets and checkpoints to the disk. What you write there stays stored.
- Shut down without worry. When you terminate the machine, the disk stays intact. Next time, just attach it again and resume exactly where you left off.
How much it saves
The gain comes from two places: you stop paying for an idle GPU just to preserve the environment, and you eliminate the "download it all again" time every session. A simple example, with a hypothetical GPU at R$ 6/h and a 40 GB model that takes ~20 min to download:
| Scenario | Without a persistent disk | With a persistent disk |
|---|---|---|
| Prep the environment each session | ~20 min of paid GPU (~R$ 2) | seconds — it is already on the disk |
| Keep the setup between uses | machine on 24/7, or rebuild everything | machine off, data preserved |
| 20 sessions a month | ~R$ 40 in "prep" alone | ≈ R$ 0 of prep |
The numbers vary with your case, but the logic is always the same: a persistent disk trades expensive GPU hours for cheap storage. And storage costs a fraction of keeping a card running.
What to keep on the disk
- Model weights (the files you download from AI repositories) — the item that hurts most to download again.
- Datasets for training or evaluation, especially the large ones.
- Checkpoints from training, so you can pause and continue a fine-tune later.
- Environments and caches that take time to assemble.
- Results — generated outputs, logs, artifacts you want to keep.
A good practice is to point your framework's model cache into the disk. That way the download happens once and, on the next machines, the model is already there.
A note on performance
Network storage is excellent for keeping and reusing files, but it has its own characteristics. For sequential reads of large files (like loading a model's weights) it works very well, especially with a local cache on the machine. For workloads with millions of tiny files read at once, packing the data (into a single file, say) performs much better. It is the same common sense as any data pipeline.
How to use it on GPUBrazil
On GPUBrazil the feature is called Persistent Disk and lives in your account, under My Disks. You create the disk at whatever size you need, from 50 GB to 10 TB, and attach it when creating an instance — GPU or CPU. Billing is in reais, straight from your balance, for the contracted size only, with no contract or lock-in.
Stop re-downloading everything each session
On GPUBrazil, the Persistent Disk keeps your models, datasets and checkpoints off the machine. Power the GPU off when you are not using it and pick up in seconds next time — in reais, no contract.
Create my Persistent DiskFrequently asked questions
What happens to the disk when I delete the machine?
Nothing. The persistent disk is independent of the machine: when you terminate the instance, the disk and all its contents stay stored, ready for the next one.
Do I need a disk to use a GPU?
No. The disk is optional. The machine already comes with its own temporary storage; the persistent disk is an extra for those who want to reuse data across machines.
Can I use the same disk on different machines?
Yes — that is the whole idea. You reuse the same disk on any new instance you create, without downloading everything again.
Read next: where to store models, datasets and checkpoints · how much it costs to run AI in Brazil.