A cloud compute instance offered at 50-90% discount vs on-demand pricing, but the provider can reclaim it with little notice (typically 2 minutes) when capacity is needed elsewhere.
Spot Instances (AWS), Preemptible VMs (GCP), and Spot VMs (Azure) sell otherwise-unused cloud capacity at deep discount. The catch: the cloud can take the instance back any time, terminating workloads. Spot is right for: batch jobs (rerun cheaply), stateless workers behind a queue (other workers pick up the slack), CI/CD runners (rebuild on terminate), distributed training (checkpoint and resume). Wrong for: stateful databases, long-running stateful services, anything that can't tolerate interruption.
Running a 200-node ML training cluster on Spot Instances at 75% lower cost than On-Demand, with periodic checkpointing to handle reclamations.
Spot Instances are the single largest cloud cost lever for workloads that tolerate interruption, and savings often dwarf any other optimization technique.
Need help implementing this in your business?
Get Started