For large production runs we could try `spot-with-fallback`

[norlandrhagen]

https://docs.coiled.io/blog/save-money-with-spot.html

For our multi-task batch jobs (process region), I think we could tolerate worker restarts / retries. We could try switching from on-demand to spot-with-fallback to reduce costs.