The bill you can defend if the job cannot die.
Spot GPU calculator
Spot GPU expected cost calculator
The bid is lower. The finished job often is not. Price expected spot hours plus restarts against the on-demand ceiling.
expected ≈ (spot rate × hours / P(finish)) + (restart hours × on-demand rate).
~34.3 GPU-hours of spot · saves $5.49 vs on-demand
Open H100 listings →How to use it
Four steps, then confirm the live SKU
expected ≈ (spot rate × hours / P(finish)) + (restart hours × on-demand rate).
- 01
Fill on-demand and spot rates
Pick a GPU to load catalog from-prices, then edit the spot bid you actually see. Catalog from-price is on-demand. Spot is whatever the provider labeled interruptible.
- 02
Set wanted hours
The runtime if nothing dies. Serving, eval, and un-checkpointable fine-tunes should stay on on-demand. Do not pretend.
- 03
Be honest about P(finish)
Probability the job completes without a reclaim. 70% is optimistic on a marketplace host. 90% is a story. If you cannot estimate it, pay on-demand.
- 04
Add restart hours
Smoke test, lost steps, queue time, billed at on-demand. People skip this. That is how a cheap weekend becomes three weekends.
Spot is a bid, not a price
On-demand is a published hour you can start. Spot is leftover capacity. The provider can take the GPU back. Your step is gone. If you checkpoint cleanly and you can wait, the bid can be rational. If you cannot, the discount is a fiction.
Expected spot hours = wanted hours / P(finish). Then add restart hours at the on-demand rate. Compare that to the on-demand ceiling. If expected is higher, the cheap cell already lost.
When spot can win
Training with frequent checkpoints and a queue you can tolerate. Hyperparameter search with many short runs. A job that is allowed to die.
Serving, eval, debugging, and un-checkpointable fine-tunes: pay on-demand. If a preemption takes you off the internet, you did not save money. You bought an incident.
- Marketplace hosts add a second failure mode: the host can vanish.
- A cheap spot H100 next to an on-demand H200 is not a comparison.
- If you skip P(finish), you are doing vendor math.
What GPUBeacon does with this
The catalog keeps the flags honest. On-demand is the default. Spot stays visible. It does not get to impersonate a price. This calculator is for the people who still want the bid, with the probability written down.
Catalog from-prices are a snapshot. Estimates are envelopes, not quotes. Confirm the live SKU, region and stock with the provider before you rent.
FAQ
Spot GPU expected cost calculator questions
Is spot always cheaper?+
The bid is lower. The finished job often is not. Restarts, queue time, and a dead serving process eat the delta. Compare expected complete cost, not the cell in the table.
What P(finish) should I use?+
Whatever you can defend. If you have no history, start pessimistic: 50–70% on a marketplace host, higher on a managed interruptible product. If you cannot pick a number, you cannot take the bid.
Should I use AWS Spot for H100?+
Only if your scheduler already speaks Spot and you can reclaim. For a one-off fine-tune, specialist on-demand is less theatre. Put both rates in this calculator instead of trusting a roundup.
Is community cloud the same as spot?+
Sometimes. Community often means someone else’s hardware. It can still be on-demand billing. Read the policy. This calculator cares about interruptibility, not the marketing name.