AI Infrastructure & Deployment

Spot Instances

Discounted cloud computing resources that use spare server capacity, often 90% cheaper than regular pricing. They're ideal for flexible tasks like data analysis and machine learning that can handle brief interruptions.

Spot Instances cloud compute AWS Azure GCP
Created: December 18, 2025

What are Spot Instances?

Spot Instances are a purchasing model for cloud compute where users access spare capacity at discounts often reaching up to 90% off on-demand pricing. Major cloud providers offer Spot Instance programs with varying features and pricing mechanisms:

Amazon Web Services (AWS) Spot Instances: Users specify a maximum price and receive instances if the market price is below this threshold. Spot Instances are reclaimed with 2-minute notice when capacity is needed or price exceeds the maximum bid.

Microsoft Azure Spot Virtual Machines (VMs): Users set a maximum price; VMs are evicted when the market price exceeds this limit. Azure provides 30-second eviction notice.

Google Cloud Spot VMs: Offer similar discounts with fixed and substantial savings. Pricing is stable for up to a month. Google provides 30-second preemption notice.

Spot Instances are best suited for flexible, fault-tolerant applications. Providers can reclaim these instances with short notice, so resilience to interruptions is a crucial requirement.

How Spot Instances Work

Pricing Mechanism

Dynamic, Supply/Demand Driven: The spot price for each instance type and region fluctuates based on long-term trends in supply and demand, not real-time bidding.

Max Price: Users can set a maximum price they’re willing to pay. As long as the current Spot price is below this, the instance runs.

No Real-Time Bidding (Modern Model): Early systems involved live bidding, but now most providers use a set-max-price model for simplicity.

Billing Granularity: AWS, Azure, and GCP bill per second, after a 1-minute minimum.

Example: On AWS, an m5.large On-Demand instance costs $0.096/hr, while the Spot price could be as low as $0.019/hr (an 80%+ saving).

Availability & Interruption

Capacity Pool: Each Spot Instance draw comes from a pool defined by instance type and availability zone.

Interruption Policy: Providers may terminate or deallocate instances with short notice—2 minutes on AWS, 30 seconds on Azure and GCP—when they need the capacity or the spot price exceeds your max bid.

Interruption Notification:

  • AWS: 2-minute warning
  • Azure: 30-second notice
  • GCP: 30-second notice

Rebalancing: AWS offers “rebalance recommendations,” giving extra early warning that a Spot Instance is at elevated risk of interruption.

Critical Point: Applications must be architected to handle interruptions—statelessness, checkpointing, and auto-recovery are essential.

Comparison: Spot vs On-Demand vs Reserved

FeatureSpot InstancesOn-Demand InstancesReserved Instances
PricingVariable, up to 90% lessFixed, highest priceDiscounted (up to 72%), fixed
AvailabilityOnly if spare capacity existsAlways availableGuaranteed during reservation
InterruptionsCan be interrupted (2-min or 30-sec notice)User decides when to terminateNot interrupted during term
CommitmentNo commitmentNo commitment1 or 3 years required
Use CasesFlexible, fault-tolerant, non-criticalAll workloads, especially criticalPredictable, steady-state
SLANo SLAStandard SLAStandard SLA
BillingPer second (after 1st minute)Per secondPer second

Cross-Cloud Provider Comparison

FeatureAWS SpotGCP Spot VMsAzure Spot VMs
Pricing ModelVariable, supply/demand drivenFixed discount (up to 91% off)Variable, supply/demand driven
Usage Time LimitNo fixed limitMax 24 hoursNo fixed limit
Interruption Notice2 minutes30 seconds30 seconds
BillingPer second (after 1 min)Per secondPer second
SLANo SLANo SLANo SLA
Integration ToolsSpot Fleet, ASG, KubernetesMIG, GKEVM Scale Sets, AKS
Best Use CasesBatch, CI/CD, ML, HPC, statelessBatch, analytics, dev/testBatch, stateless apps, CI/CD
Unique FeaturesSpot block (fixed-term interruption)Sustained use discountsEviction policy customization

Core Concepts

Spot Capacity Pool

A group of unused virtual machines of the same instance type in the same availability zone. Each pool operates independently, and capacity/price can vary between pools.

Spot Instance Request

A user-initiated request to allocate a Spot Instance:

  • One-time: Provisioned once if capacity is available
  • Persistent: Automatically resubmitted if interrupted (useful for jobs that must eventually finish)

Rebalance Recommendation (AWS)

AWS issues a rebalance recommendation before the standard interruption notice if a Spot Instance is at increased risk of termination, allowing workloads to proactively migrate or checkpoint.

Use Cases and Examples

Ideal Use Cases

Batch Processing: Data analytics, video transcoding, rendering, ETL jobs
Example: Researchers running Monte Carlo simulations for climate data, saving 80%+ on compute.

Big Data Analytics: Hadoop/Spark clusters, log/data analysis at scale
Example: Media businesses processing petabytes of logs nightly on AWS EMR with Spot Instances.

CI/CD Pipelines: Short-lived build and test environments
Example: SaaS providers using Spot Instances as Jenkins build agents for cost-effective, parallelized CI.

Machine Learning Training: Deep learning, hyperparameter tuning, checkpointed training on GPUs
Example: Teams training neural networks on Spot GPU instances with autosave for interruption recovery.

Containers and Microservices: Stateless services orchestrated by Kubernetes or Docker Swarm

Dev/Test Environments: Non-production workloads where interruption is acceptable

High-Performance Computing (HPC): Genomic sequencing, financial modeling, scientific simulations

Non-Ideal Use Cases

Mission-Critical Applications: Where high-availability and minimal downtime are paramount

Stateful Apps Without Resilience: Apps that cannot checkpoint state or recover from sudden termination

Risks and Mitigation Strategies

Key Risks

Interruption Risk: Instances can be terminated with as little as 30 seconds’ notice

Capacity Volatility: Spot capacity can disappear unpredictably

Pricing Fluctuations: Spot prices can spike, especially for popular instance types

No SLA: Providers do not guarantee uptime or availability

Mitigation Strategies

Automation:

  • Use orchestration systems (Kubernetes, AWS Auto Scaling) to reschedule and replace interrupted workloads
  • Blend Spot with On-Demand and Reserved Instances for resilience
  • Employ automation platforms that manage lifecycles and fallback

Application Design:

  • Architect services as stateless and leverage external storage
  • Implement checkpointing to resume jobs after interruption
  • Use loose coupling to tolerate node failures

Proactive Monitoring:

  • Monitor interruption rates by region/type using tools like AWS Spot Instance Advisor
  • Act on rebalance recommendations to migrate workloads early

Diversification:

  • Use mix of instance types and regions to avoid “herd” interruptions
  • Set sensible max bids to reduce risk of sudden eviction

Best Practices

1. Start with Non-Critical Workloads
Validate your interruption-handling strategies before moving critical workloads.

2. Diversify
Use multiple instance types and availability zones for higher reliability.

3. Automate
Employ AWS Spot Fleet, Auto Scaling Groups, or Kubernetes autoscalers.

4. Monitor Trends
Use AWS Spot Price History and Spot Instance Advisor.

5. Set Max Prices
Cap your bid at or below the On-Demand price.

6. Plan for Interruption
Architect for fast recovery and zero data loss.

7. Leverage Third-Party Tools
Try Cast AI or CloudZero for optimization and automation.

8. Use Tags
Track usage and savings by team/project.

9. Regularly Review
Update your mix and strategies based on usage reports.

Billing and Pricing

Spot Price: Set by provider, fluctuates with market demand and supply

Billing Granularity: Per second after first minute (AWS, Azure, GCP)

Termination Billing: Generally, if AWS interrupts, you are not charged for the last partial hour; if you terminate, you pay for seconds used.

Savings Tracking: Use provider dashboards, AWS Spot Fleet savings reports, or third-party tools

No Overlap with Savings Plans: Spot usage does not contribute to AWS Savings Plans commitments

Example Scenario

Research Team Running Simulations

A university research group needs to run thousands of Monte Carlo simulations for climate modeling. By using Spot Instances orchestrated by Kubernetes, they cut compute costs by 85% compared to On-Demand pricing. Checkpoints are saved to persistent storage before interruption, so jobs resume without loss.

Frequently Asked Questions

Q: Can I use Spot Instances for production workloads?
A: Yes, if your application is resilient to interruptions. Many organizations blend Spot with On-Demand or Reserved Instances for critical workloads.

Q: How much can I save with Spot Instances?
A: Typically 70–90% compared to On-Demand, depending on region, type, and demand.

Q: What happens when a Spot Instance is interrupted?
A: You receive a short interruption notice (2 minutes on AWS, 30 seconds on Azure/GCP). The instance is then terminated, stopped, or hibernated.

Q: How do I request a Spot Instance?
A: Use the cloud provider’s console, CLI, or API. Specify instance type, max price, and duration as needed.

Q: Can Spot Instances be used with Kubernetes?
A: Yes, container orchestrators like Kubernetes can reschedule pods when a Spot Instance is interrupted.

Q: Is there any SLA for Spot Instances?
A: No. Providers do not guarantee uptime or availability for Spot Instances.

References

Related Terms

Back to Glossary
Ă—
Contact Us Contact