Compute 2 min read updated 2026-09-03
Instance lifecycle
An instance is a GPU allocation + a disk + a template. This page is the complete state machine — what changes state, and what each state costs.
States
| State | Meaning | GPU bills | Disk bills |
|---|---|---|---|
| pending | capacity reserved, image pulling / VM booting | no | yes |
| running | workload live, ports reachable | yes | yes |
| stopped | processes halted, disk intact, GPU released | no | yes |
| destroyed | gone — disk erased (volumes survive) | no | no |
Transitions are per-second precise: the billing row for a stop written at 14:02:37 ends at 14:02:37.
Actions
- Deploy — needs ≥ $0.10 balance; price is locked to the sheet at that moment.
- Stop — graceful signal, then halt. Use it liberally: stopped instances cost only their disk.
- Start — re-acquires a GPU on the same machine class and resumes from disk. If the exact machine is busy, the start queues until capacity frees (you can destroy instead at any time).
- Destroy — immediate, irreversible for the instance disk. Attached volumes detach unharmed.
# the same four verbs everywhere
powergpu launch --gpu rtx-4090 --template pytorch --disk 50
powergpu stop i-9f2c41ab
powergpu start i-9f2c41ab
powergpu destroy i-9f2c41abInterruptible instances
Interruptible capacity (−50%) can be paused when an on-demand deploy needs the slot:
- The instance receives SIGTERM and a 30-second grace window;
- it moves to stopped — disk intact, GPU billing over;
- with auto-requeue on (default), it rejoins the queue and restarts when capacity frees;
- a console event (and webhook, if configured) records both edges.
Design rule: write checkpoints to a volume and make your entrypoint resume from the latest one. Then interruptions cost seconds, not work.
Access
Every instance gets a hostname (i-xxxxxxxx.powergpu.io) and its template's ports mapped to high ports on it. SSH is available on templates that ship it (a key you provide at deploy, or the in-console web terminal); VMs add full console access. See networking for the port model.
Events & monitoring
The instance page streams state changes, GPU/RAM utilisation and the live cost counter. The same data is available from GET /v1/instances/{id} for dashboards and schedulers.