AWS Spot Instances: How to Save Up to 90% Without Risking Uptime
AWS Spot Instances can cut EC2 costs by up to 90%. Learn which workloads fit, six rules for running Spot safely, and how to prove the savings.
AWS EC2 Spot Instances Cost Optimization

Quick answer: AWS Spot Instances are spare Amazon EC2 capacity sold at discounts of up to 90% compared with On-Demand prices. The catch is that EC2 can take an instance back with a two-minute warning. That makes Spot a strong fit for stateless, fault-tolerant and flexible workloads such as batch jobs, CI runners, containers and data processing.
Most teams know Spot exists. Far fewer run it in production, usually because one bad interruption during a test scared everyone off. That's a shame, because Spot is one of the few levers on an AWS bill that can cut a line item by more than half without changing the instance you run or negotiating anything.
What are AWS Spot Instances?
A Spot Instance is an EC2 instance that runs on spare capacity AWS hasn't sold yet. You launch it like any other instance and it behaves like one too. The only real difference, according to AWS documentation, is that EC2 can interrupt it with two minutes of notice when it needs the capacity back.
Three terms are worth knowing:
- Spot capacity pool: a set of unused instances with the same instance type and Availability Zone. An m5.large in one zone is a different pool from an m5.large in another.
- Spot price: the hourly price of a pool. EC2 sets it and adjusts it gradually based on long-term supply and demand. You don't bid for it.
- Interruption notice: the two-minute warning issued before EC2 stops, hibernates or terminates the instance.
The EC2 console shows three months of Spot price history, which is the quickest way to see how steady a pool has been.
How much can you really save with Spot Instances?
AWS says up to 90% off On-Demand. Treat that as a ceiling, not a forecast. The real discount depends on instance family, region, zone and timing, and it differs for every pool.
Here's a quick way to estimate your own number. Say a fleet costs $10,000 a month On-Demand. You move 60% of its hours to Spot at an average 65% discount:
$10,000 × 60% × 65% = $3,900 saved per month
Those figures are illustrative. The logic is what matters: your saving is the share of hours you can safely move multiplied by the discount you actually get, and the share of hours is usually the harder number to raise.
Which workloads fit Spot, and which don't?
Ask one question: if this instance vanished in two minutes, what would break? If the answer is “nothing, the work gets retried”, Spot fits.
| Good fit | Poor fit |
|---|---|
| Batch and data processing | Single-node databases with no replica |
| CI/CD build and test runners | Apps that keep their only copy of data on local disk |
| Stateless web and API tiers behind a load balancer | Long jobs that can't checkpoint |
| Containers and Kubernetes worker nodes | Software licensed to a fixed host |
| Rendering, simulation and ML training with checkpoints | Anything with a hard, no-retry SLA |
| Dev, test and staging environments | Single points of failure |
Spot vs On-Demand vs Savings Plans
These three aren't rivals. They solve different problems.
| On-Demand | Savings Plans | Spot | |
|---|---|---|---|
| Price | Full list price | Discount for a 1- or 3-year spend commitment | Up to 90% off On-Demand |
| Commitment | None | Yes | None |
| Interruption risk | None | None | Yes, two-minute notice |
| Best for | Unpredictable or short-lived needs | Steady baseline usage | Flexible, interruptible capacity |
A sensible pattern is to cover the always-on baseline with commitments, run the elastic and interruptible layer on Spot, and keep a thin On-Demand layer for anything that can't be interrupted. Our Savings Plans and Reserved Instances guide covers sizing the commitment side.
One trap: Savings Plans discount On-Demand usage, not Spot. If you size a plan on today's On-Demand hours and then move half the fleet to Spot, your commitment can end up underused. Move to Spot first, then size the plan on what's left, and watch Savings Plan utilization afterwards.

Six rules for running AWS Spot Instances safely
1. Diversify across instance types and zones
The most common cause of Spot trouble is asking for one instance type in one zone. That's a single pool, and when it runs dry you get nothing. Spread across many pools: several sizes, several generations and families, and every zone you can use. Auto Scaling groups and EC2 Fleet support attribute-based instance type selection, where you describe the vCPU and memory you need and EC2 picks from every matching type.
2. Use the price-capacity-optimized strategy
AWS recommends this allocation strategy for Auto Scaling groups and EC2 Fleet. It draws from the pools with the best capacity availability that also carry the lowest price, and AWS says it usually gives a better chance of getting capacity and lower interruption rates than the older lowest-price strategy. If you're still on lowest-price, switching is a single parameter.
3. Handle the two-minute notice
An instance can read the notice from the instance metadata service, and EC2 also publishes it as an EventBridge event called “EC2 Spot Instance Interruption Warning”. Use those 120 seconds to deregister from the load balancer, hand back in-flight work, write checkpoints to S3 and flush logs. AWS Fault Injection Service can simulate interruptions, so you find the gaps on a quiet afternoon rather than during a traffic peak.
4. Turn on Capacity Rebalancing
EC2 can send an instance rebalance recommendation when a Spot Instance is at elevated risk of interruption, often ahead of the two-minute notice. With Capacity Rebalancing enabled in Auto Scaling groups and EC2 Fleet, replacement capacity is launched proactively, so the new instance is warm before the old one goes.
5. Keep On-Demand as a floor
Use a mixed instances policy: a small On-Demand base, with everything above it split between On-Demand and Spot by percentage. A floor of On-Demand capacity means a bad hour in the Spot market slows you down instead of taking you down.
6. Check pool health before you commit
The Spot Instance Advisor shows the average interruption frequency for each instance type, and Spot placement scores indicate how likely a capacity request is to succeed in a given region or zone. Check both before choosing types for a new workload.
Running Spot Instances on Kubernetes
Kubernetes is where many teams start, because pods are already disposable. On Amazon EKS you can use managed node groups with the Spot capacity type, or an autoscaler such as Karpenter that can provision Spot nodes on demand.
A few habits make it work: run at least two replicas spread across zones, set pod disruption budgets, handle SIGTERM so pods exit cleanly, and keep system components and stateful databases on On-Demand nodes. Our EKS cost optimization guide goes deeper, and the same logic suits CI/CD runners, among the easiest workloads to move.
How to prove your Spot savings
In billing data, Spot hours appear with usage types that include “SpotUsage”, so you can separate them from On-Demand and compare cost against what the same hours would have cost at the On-Demand rate.
Track four numbers each month:
- Spot share of compute hours: how much of the fleet runs on Spot.
- Effective discount: what you paid against the On-Demand equivalent.
- Interruptions per week: a health signal, not a failure count.
- Cost per unit of work: per build, per job or per transaction. Our post on AWS cost per transaction shows how to build that metric, and the cost allocation guide explains how to give every dollar an owner.
Frequently asked questions
Are AWS Spot Instances safe for production?
Yes, for the right workloads. Stateless services behind a load balancer, container fleets and batch pipelines run on Spot in production at many companies. Databases with a single node and anything that can't be retried should stay on On-Demand or committed capacity.
What happens when a Spot Instance is interrupted?
EC2 issues a two-minute notice, then stops, hibernates or terminates the instance, depending on what you configured. Auto Scaling groups can launch a replacement automatically.
Can Savings Plans cover Spot Instances?
No. Savings Plans discount On-Demand usage. Spot is a separate purchase option, which is why commitments and Spot work best when each covers a different layer of the fleet.
Final thoughts
Spot rewards teams that treat interruptions as routine. Diversify the pools, use the recommended allocation strategy, give workloads a graceful exit, and measure the result in cost per unit of work.
If you'd like to see where Spot, commitments and On-Demand sit in your own bill, book a demo with TruCost.Cloud and we'll walk through it on your AWS data.
Related reading

AWS Graviton Cost Savings: How to Migrate to Arm in 6 Steps
AWS Graviton can lower compute costs with better price performance. Follow a six-step migration plan, avoid common mistakes and prove the savings.

AWS Cost and Usage Report: Setup, CUR 2.0 and Athena Queries
The AWS Cost and Usage Report is the most detailed billing dataset AWS offers. Learn CUR 2.0 setup, Athena queries and the pitfalls to avoid.

AWS Cost Per Transaction: How to Measure Cloud Unit Economics and Prove FinOps Value
AWS cost per transaction links cloud spend to business activity. Learn how to calculate it, which data you need, and how unit economics proves FinOps value.
See your own bill allocated to 100%
Thirty minutes with a FinOps engineer, on your AWS data if you like. No slideware.