EC2 Spot vs On-Demand: when each one makes sense

The price gap

On-demand instances have a fixed, published hourly price and no interruption risk. You pay full rate for the guarantee that the instance keeps running until you stop it. Spot instances draw from the same EC2 capacity pools, but AWS sells the spare capacity at a discount that typically lands in the 60-90% range for Spot-compatible workloads. The exact discount depends on the region, the instance type, and current capacity in each pool.

That gap is large enough to change how you budget. A single AutoScaling group of about a dozen c6i.8xlarge instances has saved one AutoSpotting customer roughly $7.5k per month. Across its user base since 2016, AutoSpotting has an estimated $100M+ in delivered savings. The savings are real and repeatable; the reason not everyone runs everything on Spot is the interruption trade-off.

<h2>The trade-off: two-minute reclamation</h2>
<p>When AWS needs Spot capacity back, it sends a Spot interruption notice and reclaims the instance two minutes later. Reclamation is driven by capacity in the specific instance pool, not by price. Setting a higher maximum price does not prevent it. If the pool you are drawing from gets tight, your instance goes away regardless of what you were willing to pay.</p>
<p>So the real question for any workload is: what happens when an instance disappears with two minutes of warning? If the answer is "another instance picks up the work and nothing user-visible breaks," the workload is a good Spot candidate. If the answer is "we lose state or drop a transaction," it belongs on on-demand, or it needs on-demand fallback underneath it.</p>

<h2>Which workloads suit Spot</h2>
<p>Good fits are workloads that are already built to tolerate an instance going away:</p>
<ul>
  <li>Stateless web and API tiers behind a load balancer, where a healthy instance absorbs the traffic.</li>
  <li>Containerized services on ECS or EKS, where the scheduler reschedules tasks and pods onto remaining capacity.</li>
  <li>Batch, data processing, CI runners, and rendering, where work is retried or resumed.</li>
  <li>Fleets that scale up and down daily, which Reserved Instances and Savings Plans cover poorly because the baseline keeps moving.</li>
</ul>
<p>Workloads to keep on on-demand, or to protect with fallback: single-instance stateful services with no replica, licensed software pinned to one host, and anything where a two-minute drain is not enough to finish in-flight work cleanly.</p>
<p>Teams that already practice chaos engineering tend to have the easiest time here. If your systems are built to survive a node failing at any moment, Spot is close to free savings, because the failure mode you are protecting against is the one Spot introduces.</p>

<h2>You do not have to choose one</h2>
<p>The practical answer for most fleets is a mix: run on Spot when capacity is available, fall back to on-demand when it is not. That is what AutoSpotting does inside your existing AutoScaling groups. It diversifies across compatible instance types to spread interruption risk, drains connections before an instance is reclaimed, and when Spot capacity runs out across all compatible pools it launches on-demand instances so the group keeps its capacity. When Spot recovers, it moves back.</p>
<p>This is where AutoSpotting differs from the native Auto Scaling group Spot integration, a caveat few teams know about: the native integration does not fall back to on-demand when Spot runs out across all of your configured instance types. It simply launches fewer instances, so you silently run at reduced capacity until Spot frees up, often exactly when demand is highest, such as the end-of-year holiday season.</p>
<p>Because it works from a tag on your existing groups and runs entirely in your own AWS account, you can turn it on for a group, measure the savings, and remove the tag to revert to pure on-demand at any time. There is no re-architecting and no cross-account access to grant, which is the design AutoSpotting has kept since it was first built in 2015-2016.</p>