Should checkpoint and restart time be in a GPU budget?
Include any time for which the provider bills the machine, including checkpoint writes, data loading and restarts. The training calculator uses the total runtime you enter; it does not assume those operations are free.
In our latest check: On the cheapest eight-GPU H100 node ($27.84 an hour, Scaleway, Paris, France), every 10 minutes spent writing checkpoints costs $4.64. Ten minutes of saving per two hours of training adds one twelfth to the billed time.
For interruptible workloads, enter your own expected restart allowance. No universal failure frequency or checkpoint interval is assumed. Add persistent storage and transfer separately.
Related questions
- How much does it cost to train an LLM?
- Do storage charges stop when a rented GPU stops?
- How much does an eight-H100 node cost?
Numbers on this page come from our verified snapshot of 21 November 2026. Full table on the homepage; method in the methodology.