The disk is what makes an interruption survivable.
Every instance comes with local NVMe scratch for speed. A persistent volume adds the part that matters on spot: data that outlives the instance, the reclaim and the tier change, for $0.08 per GB-month.
- Price
- $0.08/GB-monthProrated, billed while the volume exists
- Size
- 10–4,000GBPer volume, in steps of 10 GB
- Local NVMe
- Included2 TB on a 1× H100 node
- Egress
- FreeNo transfer fees in or out, ever
One is fast and temporary. The other is permanent and cheap.
Use both: scratch for the data being chewed on right now, a volume for anything you would be unhappy to lose.
Local NVMe scratch
Physically inside the node, included in the hourly price, scaled with the node size — 2 TB on a single-GPU H100 node and 16 TB on a 8× node. Wiped when the instance ends. Put datasets being read at full speed here, and the working copy of a model.
Persistent volume
A network volume in one region, attached to one instance at a time, from 10 GB to 4,000 GB. It survives a stop, a reclaim, a tier change and the deletion of the instance itself. Put checkpoints, model weights, results and anything a resume depends on here.
| Local NVMe scratch | Persistent volume | |
|---|---|---|
| Price | Included in the hourly rate | $0.08 per GB-month |
| Survives an instance stop | No | Yes |
| Survives a spot reclaim | No | Yes |
| Moves between instances | No | Yes, one at a time |
| Capacity | Fixed by the node size | 10 GB to 4,000 GB, steps of 10 |
| Billed while nothing runs | Never | Yes, until you delete it |
| Best for | Hot datasets, shard files, temporary output | Checkpoints, weights, results, anything resumable |
One rate, prorated, in every region.
Storage is billed from creation to deletion, whatever the instance is doing. There is no minimum term and no fee for creating or deleting a volume.
| Volume | Per month | Per day | Typical use |
|---|---|---|---|
| 100 GB | $8.00 | $0.263 | Checkpoints of a 7B fine-tune |
| 500 GB | $40.00 | $1.316 | A 70B model plus checkpoints |
| 1,000 GB | $80.00 | $2.632 | A dataset and several model versions |
| 4,000 GB | $320.00 | $10.526 | A working corpus for a team |
To put that in perspective: a 500 GB volume costs $40.00 a month — about 4,363 minutes of a spot H100.
The lifecycle of a volume
- CreatedFrom the console or while configuring an instance. Pick a size and a region; billing starts here.day 0
- Attached and runningMounted in the instance. Checkpoints land on it.running
- The instance is reclaimedThe instance stops; the volume stays attached to it, intact. Compute billing ends, storage continues.reclaim
- ReattachedTo the relaunched spot instance, or to an on-demand one. The job resumes from the last checkpoint.resume
- Deleted, by youDetach it and delete it in the console. Billing stops at that moment; nothing is deleted on our initiative.end
No egress fees, so move what you like.
Instances have a public IPv4 address and full outbound access at 100 Gbps per node. Pull models from a registry, push results to your own bucket, rsync a dataset — none of it is metered.
- SSH and SCPYour key is installed at boot.
scp -r ./data instance:/mnt/vol/works from the first second. - Registries and hubs
huggingface-cli download,docker pull,git lfs— from inside the instance, at node bandwidth. - Your own object storagerclone, s3cmd or the AWS CLI to your bucket. We do not charge for the traffic; your provider may charge for theirs.
# The volume is mounted at /mnt/vol, the scratch disk at /mnt/scratch. $ df -h /mnt/vol /mnt/scratch Filesystem Size Used Avail Mounted on /dev/vdb 500G 12G 488G /mnt/vol /dev/nvme0n1 2 TB 0G 2 TB /mnt/scratch # Datasets on scratch (fast, temporary), checkpoints on the volume (kept). $ export HF_HOME=/mnt/scratch/hf $ python train.py --output_dir /mnt/vol/run-42 --save_steps 200 # After a reclaim, the next instance sees the same volume. $ ls /mnt/vol/run-42 | tail -2 checkpoint-1400 checkpoint-1600
It is the single mistake that turns a 2-minute notice into lost work. Scratch is wiped with the instance; the volume is not.
Storage questions
What is the difference between the disk and the local NVMe?
Local NVMe is scratch space physically inside the node: the fastest storage available, included in the hourly price, and wiped when the instance ends. A persistent disk is a network volume that lives independently of any instance — it survives stops, reclaims and instance type changes, and is billed at $0.08 per GB-month while it exists.
Does my disk survive a spot reclaim?
Yes. That is the point of it. A reclaim stops the instance; the disk stays attached to that stopped instance with all its data, and the next instance — spot or on-demand — can pick it up and carry on.
How is storage billed exactly?
$0.08 per GB-month, prorated, from the moment the volume is created until you delete it. A 500 GB volume costs $40.00 a month, whether it is attached to a running instance, attached to a stopped one, or attached to nothing at all.
What are the size limits?
From 10 GB to 4,000 GB per volume, in steps of 10 GB. You can create several volumes; only one instance can hold a given volume at a time. Need more than 4 TB in one namespace? Tell us — that is a capacity question, not a product limit.
Can I attach one volume to several instances?
No. A volume is attached to a single instance at a time, which keeps the filesystem semantics simple and avoids the failure modes of shared block storage. To share data between instances, copy it or keep the source outside the platform.
Do you charge for moving data in or out?
No. There are no egress fees and no transfer quota: pulling a 200 GB model from a registry and pushing 40 GB of checkpoints back out costs nothing beyond the minutes your instance was running.
Is the data encrypted?
Persistent volumes are encrypted at rest. Access happens through your instance, over SSH or the endpoints you expose yourself. Our staff do not read customer volumes; what we can and cannot access is written in the trust page.
What happens to my disk if my balance runs out?
Running instances stop, and the volume keeps existing — and keeps being billed at the storage rate. If the balance stays at zero, we contact you before anything is deleted. Nothing is removed silently.
Give your next spot job a disk it can come back to.
Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.
Volumes are billed at $0.08 per GB-month from creation to deletion. Nothing else is charged for storage.