Instance-independent
Destroy the expensive GPU the second a job ends; the volume keeps your state for the next one. The core per-second cost pattern.
Products · volumes
GPU cloud volumes: storage that outlives the instance
Network NVMe volumes at a flat $0.08/GB/month, billed per second. Keep datasets, checkpoints and model libraries warm while instances come and go — attach to any machine in the region in seconds.
NVMe throughput
Multi-GB/s reads put 70B weights in VRAM in tens of seconds — no S3 download tax at every deploy.
Read-only fan-out
One writable attachment, many readers: a serving fleet mounts the same model library without copies.
One flat rate
$0.08/GB/mo everywhere. No IOPS tiers, no throughput classes, no per-request fees.
What storage really costs here
Per-second billing means you can also think of it hourly — both columns are the same number.
| Scenario | Size | Per month | Per hour | Notes |
|---|---|---|---|---|
| A LoRA workspace | 50 GB | $4.00 | $0.0055 | checkpoints + datasets for a fine-tune |
| A model library | 500 GB | $40.00 | $0.0548 | ~6 quantized 70B models ready to serve |
| A render vault | 2,000 GB | $160.00 | $0.2192 | scene files + frames for a studio pipeline |
| A dataset lake | 10,000 GB | $800.00 | $1.0959 | LAION-scale shards staged next to the GPUs |
The pattern that saves the most money
GPU time is the expensive line; storage is cents. Volumes let you stop paying for the former without losing your work:
- Keep datasets & weights on a volume; deploy GPUs against it only while computing.
- Point interruptible training checkpoints at a volume — interruptions become free restarts.
- Build a team model library once; every new instance mounts it read-only in seconds.
- Snapshot instance disks to volumes before destroying — one command.
$ powergpu volume create --name models --size 500 --region eu-west-1
✓ vol-8c31f2 · 500 GB · $40.00/mo
$ powergpu launch --gpu l40s --template vllm --volume models:/models:ro
✓ instance i-4fd02b11 running · library mounted read-onlyGPU volumes: FAQ
Mount options and snapshots in the storage docs.
How are volumes billed?
Allocated size × $0.08/GB/month, per second, whether attached or not. A 100 GB volume costs about $0.0110 per hour — delete it and billing stops that second.
Can several instances share one volume?
One read-write attachment at a time, plus unlimited read-only attachments in the same region — the standard pattern for serving fleets reading one model library while a trainer writes checkpoints elsewhere.
How fast are they?
NVMe-backed over the datacenter fabric: multi-GB/s sequential reads on datacenter hosts — enough to load a 70B model in well under a minute — with local instance NVMe still there for scratch I/O.
Do volumes survive instance destruction?
Yes — that is the point. Destroy instances freely; volumes persist until you delete them, and reattach to any new instance in their region.