Flat −50%, always
Not "up to". Not "from". Every model, every region: exactly half the on-demand price, fixed like everything else here.
Products · interruptible
Interruptible (spot) GPUs: half price, zero auctions
Interruptible instances cost a flat 50% of on-demand — H100 SXM at $0.714/hr instead of $1.428. No bids to place, no eviction markets to monitor: your job may pause under capacity pressure, your data never disappears.
Disk survives
Interruption = stop, not destroy. Checkpoints, datasets and outputs stay on the instance disk, billing drops to storage-only.
Auto-requeue
Interrupted instances rejoin the capacity queue automatically and restart from disk — overnight queues drain themselves.
No babysitting
Nothing to outbid at 2am. The console shows interruption events; webhooks tell your pipeline. That is the whole operational surface.
The math on real jobs
Same silicon as on-demand — the only difference is the eviction clause.
| GPU | On-demand | Interruptible | 100 GPU-hrs cost | You save | Action |
|---|---|---|---|---|---|
| H100 SXM | $1.428 | $0.714 | $71.40 | −$71.40 | Deploy |
| A100 SXM4 | $0.560 | $0.280 | $28.00 | −$28.00 | Deploy |
| L40S | $0.514 | $0.257 | $25.70 | −$25.70 | Deploy |
| RTX 5090 | $0.439 | $0.219 | $21.90 | −$22.00 | Deploy |
| RTX 4090 | $0.327 | $0.163 | $16.30 | −$16.40 | Deploy |
| RTX 3090 | $0.108 | $0.054 | $5.40 | −$5.40 | Deploy |
Versus the marketplaces: their spot is an auction — win it, and a higher bid can still evict you an hour later. Ours is a queue.
Made for fault-tolerant work
- Checkpointed training — resume from the last step; pay half for every step.
- Batch queues — image generation, transcription, embedding jobs, render frames.
- Hyperparameter sweeps — dozens of short runs, none of them precious.
- CI & evaluation — nightly benchmark suites that just re-run if paused.
If an interruption would hurt, use on-demand — or mix: serve on-demand, train interruptible, one balance pays both.
$ powergpu launch --gpu a100-sxm4 --type interruptible \
--template axolotl --volume ckpts:50GB
✓ instance i-b3e19d02 running · $0.280/hr
# 03:12 interrupted — checkpoint step 4200 on volume
# 03:57 capacity free — auto-restarted, resumed 4200
✓ run finished · 9.4 GPU-hrs · $2.63Interruptible GPUs: FAQ
Checkpoint patterns live in the instance docs.
How often do interruptions actually happen?
Only when an on-demand deploy needs your slot — there is no bidding war constantly re-pricing you out. Fleet-wide, interruptible instances run a median of many hours between interruptions; deep-supply consumer cards (RTX 30/40 series) are interrupted least.
What happens when my instance is interrupted?
It is stopped, not destroyed: the container/VM state on disk survives and GPU billing ends instantly. You get a console event and an optional webhook, and the instance auto-requeues for capacity unless you disable that. When capacity frees, it restarts from your disk.
Is the discount really a flat 50%?
Yes — interruptible = on-demand × 0.50 on every model, every region, all the time. It is not a fluctuating spot market: the price is as fixed as our on-demand price, just half of it.
How should I design a job for interruptible?
Checkpoint to the mounted volume (every N steps for training, per-item for batch queues), make startup idempotent, and let auto-requeue do the rest. The docs include ready patterns for PyTorch, axolotl and render farms.