What’s the price difference between spot and pay-as-you-go GPU instances?

When not to use spot instances

For online inference, stateful services, or training tasks with tight deadlines, avoid spot/preemptible instances—interruption costs often far exceed savings.

How to minimize interruption loss

Always enable checkpointing (save model every N steps) in your training script and use a scheduler that supports automatic restarts; offload one-time tasks like data preprocessing to pay-as-you-go instances.