Configure an automated policy that prefers Spot but falls back to On-Demand when Spot is unavailable, and enable periodic checkpointing is the correct choice because it combines the cost benefits of Spot with the reliability of On Demand and it preserves training progress through checkpoints.

This automated policy uses Spot instances when capacity and price make sense and it automatically shifts to On Demand when Spot capacity is reclaimed. Checkpointing saves model state periodically so training can resume from the last checkpoint rather than restarting from scratch. Together this approach minimizes both spend and disruption for long running foundation model training.

Use only On-Demand Instances to ensure capacity is not ideal because it does give stable capacity but it prevents you from taking advantage of lower cost Spot capacity which raises total training cost for large workloads.

Run exclusively on Spot Instances and restart training whenever there is an interruption is not ideal because Spot alone can produce frequent interruptions and lost progress unless you add automated failover and checkpointing which this option does not include.

Purchase Compute Savings Plans for all training workloads to avoid interruptions is misleading because Savings Plans affect pricing and reduce cost for consistent usage but they do not change Spot interruption behavior or provide capacity guarantees.

Full AWS Practitioner Certification Question