Volumes & storage

Instance disk

Sized at deploy, local NVMe, fastest possible I/O. It survives stop/start and dies with destroy. Use it for scratch, caches and the OS — anything you can regenerate.

Network volumes

Created independently, attached to instances in the same region, NVMe-backed over the fabric:

powergpu volume create --name datasets --size 500 --region eu-west-1
powergpu launch --gpu a100-sxm4 --template pytorch \
  --volume datasets:/data          # read-write
powergpu launch --gpu l40s --template vllm \
  --volume models:/models:ro       # read-only, unlimited attachments

• Attachment model — one read-write attachment at a time; unlimited read-only attachments. The classic split: one trainer writes checkpoints, a serving fleet reads the model library.
• Region-bound — a volume lives in one region; attach from any machine there. Cross-region moves are a snapshot + restore (bandwidth billed at $0.01/GB).
• Resize — grow online anytime; shrink = create smaller + copy.

Snapshots

Point-in-time copies of an instance disk or a volume:

powergpu snapshot create i-9f2c41ab --name golden-env
powergpu volume create --from-snapshot golden-env --name env2 --size 100

Snapshots bill as allocated GB at the same flat rate while they exist. The common pattern: snapshot a configured environment, destroy the instance, restore next week for pennies of storage.

Cost intuition