Should I use reserved instances for recurring training jobs?

Maybe—but usually not as the first choice for training jobs.

Use Reserved Instances / Savings Plans if:

• Your training jobs run regularly and predictably
• You have steady baseline usage on the same instance families
• You expect to use the capacity for 1–3 years
• The jobs run long enough that the discount outweighs flexibility loss

Prefer On-Demand or Spot if:

• Training is bursty, experimental, or uncertain
• You often change instance type, GPU count, or architecture
• You want to maximize savings and can handle interruptions
• Jobs are short or sporadic

Important nuance for ML training

For recurring training, a common pattern is:

• Reserve only the always-on baseline compute you know you’ll use
• Use Spot instances for the rest of the training fleet, if your framework can checkpoint and resume

Rule of thumb

• Predictable, continuous workload → consider Reserved Instances / Savings Plans
• Flexible, fault-tolerant training → Spot is often better
• Unsure → start with on-demand, measure usage, then commit later

If you want, I can help you decide between Reserved Instances, Savings Plans, and Spot for your specific training setup.