Interruptible capacity at a deep discount — model the monthly saving.
Spot saves 60–90% but the instance can vanish in two minutes — perfect for batch and CI, wrong for databases. For steady always-on load compare the reserved instance savings calculator.
The Spot Instance Savings Calculator estimates how much you save by running workloads on spot instances instead of on-demand capacity. You enter the on-demand hourly rate, the spot discount percentage, the hours per month the instances run, and the number of instances. From these it derives the effective spot rate (on-demand minus the discount), multiplies by hours and instance count to get monthly cost under each pricing model, and reports the difference as your monthly and yearly savings. The two biggest drivers are the discount percentage — which sets how far below on-demand your spot price falls — and total usage (hours × instances), since savings scale directly with how much compute you run.
The practical trade-off to watch is that a spot instance's steep discount comes with interruption risk: the provider can reclaim the capacity with little notice, so the headline savings only hold for workloads that tolerate being stopped and restarted. Use the calculator to size the opportunity, then weigh it against how much of your fleet can safely run on spot. A high discount on continuous, hours-heavy workloads produces the largest yearly figure here, but those same always-on workloads are often the least tolerant of interruption — so treat the output as the upper bound of what's possible, not a guaranteed number, and reserve on-demand or reserved capacity for anything that cannot be interrupted.
Commonly 60–90% off on-demand, depending on instance type and current spare capacity. The trade-off is that the provider can reclaim the instance with about two minutes notice, so the savings only apply to workloads that tolerate interruption.
Fault-tolerant, stateless or checkpointable jobs: CI/CD runners, batch processing, data pipelines, video rendering, ML training with checkpoints, and stateless web workers behind a queue. Avoid spot for databases, stateful singletons or anything that cannot resume mid-task.