What are spot instances, and how do they differ between clouds?
Spot instances are virtual machines that a cloud provider sells from its unused capacity at a large discount, on the condition that it can take them back when it needs the capacity. AWS advertises discounts of up to 90% and gives a two-minute interruption notice, Google Cloud up to 91% with a shutdown period of up to 30 seconds, and Azure evicts with about 30 seconds of notice. Spot capacity suits work that can stop and restart, such as batch jobs, CI and stateless services with several replicas.
How spot instances work
A provider keeps spare capacity in every region for growth and failover. Spot instances let customers use it until on-demand customers need it back. When that happens, the provider sends a notice and stops or deletes the instance. Nothing is guaranteed: a spot instance may run for weeks or for minutes, and a request may not be fulfilled at all when a region is short of a machine type.
Spot instances are the same hardware as on-demand instances. Only the purchase terms differ: a lower price in exchange for no guarantee that the instance keeps running.
Comparison across AWS, Google Cloud and Azure
| AWS Spot Instances | Google Cloud Spot VMs | Azure Spot Virtual Machines | |
|---|---|---|---|
| Advertised discount | Up to 90% off on-demand | Up to 91% | Variable by region and size |
| How the price is set | Set by AWS, adjusts gradually with long-term supply and demand | Set by Google, changes at most once a day | Variable price; you can set a maximum price or -1 |
| Notice before interruption | 2 minutes, plus an optional earlier rebalance recommendation | Configurable preemption notice (120 s recommended), then a best-effort shutdown period of up to 30 s | Up to 30 s through Scheduled Events, best effort |
| What happens on interruption | Terminate, stop or hibernate | Stop (default) or delete | Deallocate (default) or delete |
| Maximum runtime | None | None, unless you set one | None |
Sources: AWS Spot pricing, interruption notices and rebalance recommendations; Google Cloud Spot VMs; Microsoft Azure Spot Virtual Machines. Checked on 14 November 2026.
AWS
AWS publishes the interruption notice two minutes before it stops or terminates the instance, both in instance metadata and as an EventBridge event, and recommends polling every 5 seconds. A separate rebalance recommendation signals that an instance is at elevated risk of interruption. It can arrive earlier than the two-minute notice, but AWS says it sometimes arrives together with it. Auto Scaling groups can use it to launch a replacement before the interruption. You do not bid: you pay the spot price in effect while the instance runs. Fargate Spot applies the same idea to serverless containers, covered in our Fargate pricing breakdown.
Google Cloud
Google calls Spot VMs the latest version of preemptible VMs and recommends them instead. The key difference is runtime: Compute Engine always stops preemptible instances after 24 hours, while Spot VMs have no maximum runtime unless you set one. A Spot VM’s preemption starts with a metadata signal, followed after the configured notice duration by an ACPI soft-off signal and a best-effort shutdown period of up to 30 seconds for shutdown scripts.
Azure
Azure evicts Spot VMs when it needs the capacity or when the current price rises above the maximum price you set. Setting the maximum price to -1 removes price-based eviction and caps the price at the standard rate. The default eviction policy, Deallocate, keeps the disks, which you keep paying for. Delete removes the VM and its disks. B-series and promo sizes are not available as Spot VMs, Spot VMs have their own quota, and Microsoft gives them no SLA.
Spot GPUs
All three hyperscalers sell GPU machine types as spot capacity, and Google Cloud documents that Spot GPUs cost less than standard GPUs and follow the same preemption process. Spot GPU capacity depends on how much on-demand demand leaves free, so training jobs that use it need frequent checkpoints. The reasons GPU capacity is scarce are covered in why it is hard to get GPU capacity in the cloud.
Providers without a spot market
Hetzner does not sell spot or preemptible cloud servers. Its prices are fixed per hour with a monthly cap, and on-demand prices are already low; the article on why Hetzner is cheap compares them with AWS. The closest Hetzner equivalent is the Server Auction, which sells refurbished dedicated servers at prices that fall over time. An auction server is a regular rented dedicated server that Hetzner does not take back for capacity, so it behaves like a normal dedicated server rather than like spot capacity.
Which workloads fit spot capacity
| Good fit | Poor fit |
|---|---|
| Batch processing and data pipelines that checkpoint | Single-instance databases |
| CI runners and build jobs | Stateful services without replication |
| Stateless web and API replicas behind a load balancer | Long jobs that cannot resume |
| Dev and test environments | Anything that must not lose in-memory state |
The common pattern is to mix: keep a baseline of on-demand capacity for the parts that must stay up and add spot capacity for the rest, spread across several instance types and zones so one capacity shortage does not take out every node.
Spot instances and Kubernetes
Kubernetes handles spot interruptions well when workloads have several replicas, PodDisruptionBudgets and graceful shutdown. The details of draining nodes on notice and choosing which Pods may land on spot nodes are in which Kubernetes workloads can run on spot instances.
On Cloudfleet, node auto-provisioning can launch spot capacity on AWS and Google Cloud when a Pod’s node affinity allows it, and replaces nodes that the provider reclaims.