Spot Instances
Spot Instances are unused cloud compute capacity offered at a steep discount compared to on-demand pricing, which the provider can reclaim with short notice if it needs the capacity back, making them best suited for flexible,…
Definition
Spot Instances are unused cloud compute capacity offered at a steep discount compared to on-demand pricing, which the provider can reclaim with short notice if it needs the capacity back, making them best suited for flexible, fault-tolerant workloads.
Overview
Cloud providers maintain large pools of physical hardware, and at any given moment some fraction of that capacity is idle. Spot Instances, a term popularized by AWS but with equivalents on Azure (Spot VMs) and Google Cloud (Spot VMs/Preemptible VMs), let customers use that idle capacity at discounts that can reach 70-90% off on-demand pricing. The catch is that this capacity isn't guaranteed: if the provider needs it back for on-demand or reserved customers, spot instances can be interrupted with only a short warning — typically around two minutes on AWS. This tradeoff makes spot instances a great fit for workloads that are stateless, resumable, or fault-tolerant by design: batch data processing, video rendering, CI/CD build agents, and large-scale simulations can all checkpoint progress or simply restart on another instance if interrupted, without materially affecting the end result. It's a poor fit for anything that needs guaranteed uptime, like a production database or a service backing real-time user traffic, unless the architecture is specifically designed to tolerate sudden termination — for example, by running only a portion of an Auto Scaling group on spot capacity, with the rest on stable Reserved Instances or on-demand pricing. Many managed services now handle spot interruption automatically: Kubernetes clusters can run a mix of spot and on-demand nodes with workloads scheduled accordingly, and services like AWS Fargate Spot or EC2 Fleet automate replacing interrupted instances. Spot usage is a common lever in FinOps cost optimization for teams with genuinely interruption-tolerant workloads.
Key Concepts
- Discounts of roughly 70-90% off on-demand pricing for unused capacity
- Capacity can be reclaimed by the provider with a short interruption notice
- Best suited for stateless, resumable, or fault-tolerant workloads
- Often combined with on-demand or Reserved Instances in a mixed Auto Scaling group
- Managed services (Fargate Spot, EC2 Fleet, Kubernetes spot node pools) automate handling interruptions
- Pricing fluctuates with real-time supply and demand for spare capacity