Azure Spot VMs: Licensing Strategy and Cost Savings Framework

Azure Spot VMs provide 60-90% cost savings compared to PAYG pricing by allowing Microsoft to reclaim the capacity when needed. This is the single largest percentage discount available in the Azure catalogue — deeper than even 3-year reservations (49%). The catch: your VM can be evicted with 30 seconds' notice. For organisations with the right workloads, this is an extraordinary cost lever. For organisations that misapply it to production workloads, it's a reliability catastrophe. This guide gives you the framework to identify suitable workloads, design eviction-tolerant architectures, and extract maximum Spot savings without reliability risk.

Spot VM Pricing Mechanics

Spot VM pricing fluctuates based on supply and demand for Azure capacity in each region and availability zone. You set a maximum price (price cap) — if the Spot price rises above your cap, your VM is evicted. Setting the cap at or near PAYG price means you run whenever Spot is below PAYG (which is the majority of the time) and are evicted only during peak demand spikes.

Spot prices are approximate and fluctuate. East US region, Linux OS. Add Windows Hybrid Benefit where applicable — this reduces the OS component from Spot price too.

The key insight from the table: Spot VMs are typically 56-88% cheaper than even 3-year reserved instances. For workloads that can tolerate eviction, this is the most powerful Azure cost optimisation available — more impactful than any reservation strategy.

Workload Suitability Framework

Eviction Handling Architecture

The 30-second eviction notice is sufficient for graceful shutdown only if applications are pre-designed for it. Three eviction handling patterns:

Pattern 1: Checkpoint and Resume (Batch/ML)

Applications write a checkpoint to persistent storage (Azure Blob, Azure Files, or Managed Disk not attached to the Spot VM) at regular intervals — every 5-15 minutes for most batch jobs. On eviction notice, the application completes the current checkpoint write and exits cleanly. On VM restart, the job resumes from the last checkpoint. Cost of checkpointing: typically 2-5% of job duration in overhead. The 30-second eviction notice is sufficient to write most final checkpoints if the checkpoint write is in progress and is under 30 seconds.

Pattern 2: Stateless Replacement (Web/API)

Stateless application tiers behind a load balancer can use Spot in VMSS with mixed-instance pools. The recommended VMSS configuration: set a minimum of 2 on-demand (reserved) instances as the HA baseline, with Spot instances handling overflow capacity. Azure Load Balancer health probes detect evicted instances within 30-60 seconds and route traffic to remaining instances. For applications that can tolerate a brief (30-60 second) reduction in capacity, this architecture delivers 60-70% savings on overflow capacity.

Pattern 3: Eviction-Aware Kubernetes (AKS)

Azure Kubernetes Service supports Spot node pools. Eviction-tolerant pod workloads (batch jobs, non-critical background processing) are scheduled on Spot node pools via taints/tolerations. When a Spot node is evicted, AKS automatically reschedules pods on remaining nodes or scales up on-demand nodes to absorb the workload. The AKS Spot integration handles the complexity of Spot lifecycle management — it is the most operationally mature pattern for large-scale Spot adoption.

Hybrid Benefit + Spot: Maximum Discount Stacking

Azure Hybrid Benefit for Windows Server can be applied to Spot VMs, reducing the Windows licence component of the already-discounted Spot price. The combined effect for Windows workloads:

• Windows D8_v5 PAYG (with Windows licence): ~$0.576/hour
• Windows D8_v5 Spot (without Hybrid Benefit, 75% off): ~$0.144/hour
• Windows D8_v5 Spot (with Windows Server Hybrid Benefit): ~$0.060-$0.080/hour
• Combined saving vs PAYG Windows: 86-90%

This makes Windows Server Spot VMs with Hybrid Benefit price-competitive with Linux VMs in many scenarios — a significant finding for enterprises with large Windows batch processing estates. The Hybrid Benefit application to Spot VMs uses the same SA rules as standard VMs: Datacenter covers unlimited VMs, Standard covers 2 VMs per licence.

Regional and SKU Strategy for Spot Stability

Not all Azure regions and VM families have equal Spot availability and eviction rates. Strategic considerations:

Use multiple regions: Distribute Spot workloads across East US, West US 2, and North Europe for batch processing. If one region experiences capacity pressure and evicts, the workload continues in another. Azure Batch and AKS both support multi-region Spot pools natively.

Diversify VM families: Configure VMSS or Batch pools to accept multiple compatible VM families (D4_v5, D4s_v4, D4s_v3). When capacity is constrained on D4_v5, Azure falls back to D4s_v4. This multi-family configuration can reduce eviction rates by 50-60% compared to single-SKU configurations.

Avoid GPU Spot during peak demand: GPU Spot VMs (NCsv3, NC A100 v4) are in high demand for AI training workloads. Eviction rates for GPU Spot are consistently higher than general-purpose Spot. Plan for 15-20% eviction rates on GPU Spot versus 3-8% for general-purpose Spot in standard regions.

Common Spot VM Mistakes

1. Deploying Spot without eviction handling. Applications that crash or lose data on eviction — without checkpointing or retry logic — negate the cost benefit. A failed batch job that must rerun from scratch wastes the Spot savings and adds rerun costs.

2. Single-region, single-SKU Spot pools. Concentration in one region and one VM family maximises eviction risk. Always diversify across SKU families and consider multi-region for mission-critical batch workloads.

3. Using Spot for production databases. This seems obvious, but we have seen production SQL Server workloads deployed on Spot VMs "to save money." A 30-second eviction on a primary SQL Server instance causes data inconsistency, corruption risk, and extended recovery time. Do not use Spot for any stateful production database.

4. Not monitoring Spot pricing history before deployment. Azure exposes historical Spot pricing per SKU per region via the portal and API. Review the last 30 days of pricing history before selecting a SKU to confirm the typical discount and eviction rate range. SKUs with "Medium" or "High" eviction rates will produce frequent interruptions for most workloads.

Frequently Asked Questions