Guaranteed capacity
Once running, the instance is yours until you stop it. On-demand workloads are never preempted — that is what interruptible is for, at half price.
Products · on-demand
On-demand cloud GPUs.
Fixed price, guaranteed resources
The default way to rent a GPU: deploy in ~30 seconds, keep it as long as you want, pay per second at a price that was fixed ≥30% under the market median before you arrived. No auctions to win, no evictions to survive.
Price fixed at deploy
Weekly market re-checks only change prices for new deploys. A training run started today finishes at today's rate.
Per-second, no minimums
A 90-second experiment costs 90 seconds. No hourly rounding, no daily minimum, no idle fees when stopped (storage only).
1× to 8× per machine
Scale within a chassis at the same per-GPU price — NVLink on SXM hardware, no multi-GPU premium anywhere.
Today's on-demand rates, spot-checked
Full sheet on the pricing page — here is the shape of it.
| GPU | VRAM | Market median | On-demand | Est. / month | Action |
|---|---|---|---|---|---|
| H100 SXM | 80 GB | $2.27 | $1.587 −30% | $1,159 | Deploy |
| H200 | 141 GB | $4.37 | $3.058 −30% | $2,232 | Deploy |
| A100 SXM4 | 80 GB | $0.83 | $0.583 −30% | $426 | Deploy |
| L40S | 48 GB | $0.67 | $0.466 −30% | $340 | Deploy |
| RTX 5090 | 32 GB | $0.45 | $0.318 −30% | $232 | Deploy |
| RTX 4090 | 24 GB | $0.37 | $0.262 −30% | $191 | Deploy |
Built for stateful, interactive work
On-demand is the right mode when losing the machine mid-job costs more than the discount you would get on interruptible:
- Serving & APIs — vLLM endpoints, ComfyUI backends, anything with users on it.
- Interactive sessions — Jupyter, SSH development, dataset exploration.
- Un-checkpointed jobs — long renders, one-shot ETL, anything you cannot resume.
- Deadline work — when "it re-queued overnight" is not an acceptable answer.
Fault-tolerant batch jobs should look at interruptible (−50%); permanent workloads at reserved (−35%).
$ powergpu launch --gpu h100-sxm --template vllm \
--env MODEL=meta-llama/Llama-3.1-8B-Instruct
✓ capacity reserved on m-1d29c04a (eu-central-1)
✓ instance i-52ab77c1 running (24.1s)
# endpoint: https://i-52ab77c1.powergpu.io:8000/v1
# billing: $1.587/hr · per second · stop anytimeOn-demand GPUs: FAQ
More detail in the instance lifecycle docs.
What does "on-demand" guarantee exactly?
The GPU is yours from start to stop: no preemption, no eviction, no bidding. The price you deploy at is contractually your price for the life of the instance, and capacity is reserved to your account the moment the deploy succeeds.
How fast is a deploy really?
Container templates cold-start in about 30 seconds — image pull is the variable (popular templates are pre-cached on hosts). Full VMs take 2–4 minutes. You can watch the state machine live in the console.
Can I stop an instance and keep my data?
Yes. Stop ends GPU billing that second; the disk stays, billed at $0.08/GB/mo, and the instance restarts on the same data. Destroy ends everything, including storage billing.
Is on-demand available on every GPU model?
All 76 models in the catalogue, from flagship B200s to GTX-class cards — same rules, same per-second billing, 1× to 8× per machine.