Reserved Instances vs Spot vs Savings Plan: Which Azure Pricing Option Fits Your Workload

Azure offers four distinct pricing models for compute: pay-as-you-go, Reserved Instances, Spot VMs, and Savings Plans. Each is optimised for a different workload profile. Choose the right one and you maximise savings without risk. Choose the wrong one and you either pay 40–70% more than necessary (PAYG on stable workloads) or introduce outage risk (Spot on production services).

This is a technical decision guide for architects, DevOps engineers, and cloud infrastructure leads. It covers the mechanics of each pricing model, the workload profiles they suit, the technical implementation requirements, and a decision matrix you can apply to your own environment.

Understanding the Four Pricing Models

Pay-As-You-Go (PAYG)

The default. You pay the full list price for every compute hour, with no commitment and no penalty for stopping. Maximum flexibility, maximum cost.

PAYG is appropriate for:

• Workloads with genuinely unpredictable lifecycles (new projects, proof-of-concepts)
• Short-lived resources (under 30 days)
• Workloads that may be decommissioned within 1 year

For everything else, PAYG is the most expensive option by a significant margin.

Reserved Instances (RIs)

A commitment to use a specific compute resource configuration — VM size, region, and optionally resource group — for 1 or 3 years, in exchange for a pre-negotiated discount off PAYG pricing.

The key technical parameters:

Scope:

• Shared scope: The RI applies to any VM matching the configuration across the entire Azure subscription or all subscriptions in the billing account. Highest flexibility within a commitment.
• Single subscription scope: The RI applies only within a specific subscription. Use when you need to attribute cost to a specific team or project.
• Resource group scope: Applies within a specific resource group. Most restrictive, least flexible.

Term:

• 1-year: 36–45% discount vs. PAYG
• 3-year: 55–72% discount vs. PAYG

Payment options:

• All upfront: Maximum discount, full payment at purchase. Best if capital budget is available and the organisation can absorb the upfront cost.
• Monthly upfront: Slight reduction from maximum discount, but spreads the commitment across monthly payments. Better for OpEx-constrained organisations.
• No upfront: Smallest discount but no capital outlay. Still 20–30% cheaper than PAYG; useful when the full commitment cannot be funded upfront.

Instance size flexibility: Modern RIs offer "Instance Size Flexibility" within the same VM series (e.g., an RI for a D8s_v5 can also apply to two D4s_v5 instances or one D16s_v5 at proportional rates). This flexibility applies automatically within the same series and region — you do not need to exchange the RI to change VM sizes within the series.

RI exchange and cancellation rules:

• RIs can be exchanged for another RI (different size, region, or term) — but the exchange must result in equal or greater total cost. You cannot exchange for a cheaper RI and receive a refund on the difference.
• RIs can be returned (cancelled) with a 12% early termination fee, up to a $50,000 return limit per rolling 12 months.
• Exchanges and returns are processed via the Azure portal → Reservations → Exchange or Return.

Spot VMs

Spot VMs use Azure's excess (unused) compute capacity, sold at heavily discounted prices. The trade-off: Azure can reclaim the capacity at any time with a brief eviction notice.

Eviction mechanics — understanding the 2-minute window:

When Azure needs Spot capacity back, it sends an eviction signal. You have approximately 30 seconds before the actual eviction begins (confirmed via the Instance Metadata Service), with a total window of up to 2 minutes. This is the window in which your application must:

1. Detect the eviction signal
2. Save its current state (checkpoint)
3. Gracefully shut down

To detect eviction signals:

# Poll the Instance Metadata Service for eviction notices curl -H Metadata:true \ "http://169.254.169.254/metadata/scheduledevents?api-version=2020-07-01"

When a `Preempt` event appears in the response, the 2-minute clock starts.

Checkpointing patterns for Spot workloads:

For batch jobs, implement periodic checkpointing — save state to Azure Blob Storage at regular intervals so that on eviction, a replacement Spot VM can resume from the last checkpoint rather than restarting from zero.

# Example: checkpoint pattern for Python batch job import json import time

Spot VM eviction policies:

You can configure what happens when a Spot VM is evicted:

• Stop/Deallocate: The VM stops and deallocates (no compute billing). Azure will restart it when capacity is available. Best for workloads that can wait.
• Delete: The VM is deleted on eviction. Use for ephemeral compute in autoscaling pools where the orchestrator manages replacement.

Spot pricing volatility:

Spot prices vary by VM size, region, and demand. They are generally 60–90% below PAYG, but this is not guaranteed. Check current Spot pricing via the Azure portal or:

az vm list-skus --location uksouth --output table # Then check spot pricing: az vm get-skus --location uksouth --name Standard_D4s_v5

For production planning, budget Spot workloads at 70% discount (conservative) rather than the maximum.

Azure Savings Plans

Savings Plans are a commitment to spend a minimum dollar/pound amount per hour across all eligible Azure compute services, for 1 or 3 years. Unlike RIs, they are not tied to a specific VM size or region — the discount applies to any eligible compute that falls within your committed spend level.

What Savings Plans cover:

• Azure VMs (any series, any region)
• Azure Container Instances
• Azure Functions (dedicated and premium plans)
• Azure App Service (premium plans)
• Azure Dedicated Host

Savings Plans discount levels:

• 1-year Compute Savings Plan: 15–47% off PAYG (varies by service and region)
• 3-year Compute Savings Plan: 33–65% off PAYG

Note: these are lower than Reserved Instance discounts. The trade-off is flexibility.

When Savings Plans are applied: Azure automatically applies your Savings Plan to the most expensive eligible compute charges first, maximising the benefit. Usage above your committed hourly amount is billed at PAYG rates.

The Decision Matrix

Use this matrix to determine the right pricing model for each workload in your environment:

Workload Characteristic Recommended Model Notes

Runs 24/7, same VM size for 12+ months 1-year Reserved Instance Highest ROI for stable workloads

Runs 24/7, same VM size for 36+ months, architecture locked 3-year Reserved Instance Maximise discount, higher commitment risk

Runs continuously but VM size may change Savings Plan Flexibility across VM series

Batch processing, can checkpoint Spot VM 60–90% saving; requires checkpoint engineering

CI/CD build pipeline Spot VM Natural checkpointing via pipeline design

ML training runs Spot VM + checkpointing GPU Spot pricing can save £5k+/month

Dev/QA environment Spot VM + auto-shutdown Combine for maximum non-prod saving

Load varies 10x between peak/off-peak PAYG + autoscaling Savings Plans secondary option

Short-lived project (under 90 days) PAYG No long-term commitment warranted

Mixed compute (VMs + containers + functions) Savings Plan Single commitment covers all eligible compute

Real Workload Scenarios With Cost Calculations

Scenario 1: Production Web Application (8 VMs, Standard_D4s_v5, UK South)

PAYG cost: 8 × £0.24/hour × 8,760 hours = £16,819/year

1-year Reserved Instance (shared scope, all upfront):

• Discount: ~40%
• Cost: £16,819 × 0.60 = £10,091/year
• Annual saving: £6,728

3-year Reserved Instance (all upfront):

• Discount: ~60%
• Cost: £16,819 × 0.40 = £6,728/year equivalent
• Annual saving: £10,091

Recommendation: 1-year RI if architecture review expected in 18–24 months; 3-year RI if infrastructure is locked. Do not use Spot (production traffic cannot tolerate eviction).

Scenario 2: Nightly Data Pipeline (4× Standard_D16s_v5, runs 3 hours/night)

PAYG cost: 4 × £0.96/hour × 3 hours × 365 nights = £4,205/year

Spot VM cost (70% discount conservative estimate): 4 × £0.29/hour × 3 hours × 365 nights = £1,270/year Annual saving: £2,935 (70% reduction)

Recommendation: Spot VMs with checkpointing. A 3-hour pipeline can checkpoint every 20 minutes, limiting rework to 20 minutes maximum on eviction. The engineering cost to add checkpointing to a typical Python/Spark pipeline: 2–4 developer days.

Scenario 3: Mixed Workload Organisation (£200k/year Azure compute, multiple VM types)

An organisation with a diverse compute footprint: multiple VM series (D-series, E-series, F-series), several regions (UK South, West Europe), and evolving architecture (migration to AKS planned for next 18 months).

Reserved Instances: Locking into specific VM sizes and regions carries risk given the planned migration. RI exchanges are possible but introduce complexity.

Savings Plans (1-year Compute, hourly commitment £10/hour):

• Committed spend: £10/hour × 8,760 hours = £87,600/year
• Discount applied: ~35% average
• Saving on committed portion: £87,600 × 0.35 = £30,660/year
• Remaining usage billed at PAYG (above hourly commitment)

Recommendation: 1-year Savings Plan covering base load, PAYG for variable/peak load above commitment. Revisit RI purchasing once AKS migration is complete and stable VM footprint is known.

Scenario 4: Machine Learning Training (GPU Cluster, NC-series)

ML training workloads have very high GPU costs and are a natural fit for Spot pricing.

Standard_NC6s_v3 PAYG: ~£3.53/hour Standard_NC6s_v3 Spot (70% conservative discount): ~£1.06/hour

For a team running 100 GPU hours/month:

• PAYG: £353/month = £4,236/year
• Spot: £106/month = £1,272/year
• Annual saving: £2,964

For larger ML teams (1,000 GPU-hours/month), this scales to £29,640/year in GPU cost savings — before considering the impact of Spot on parallel training runs.

Recommendation: Use Azure ML's built-in Spot support — Azure ML can automatically restart training jobs from the last checkpoint when Spot capacity returns, with no custom engineering required.

Combining Pricing Models: The Optimal Strategy

The highest-performing Azure cost configurations use multiple pricing models simultaneously:

The "Three-Layer" approach:

1. Base load (always-on stable compute) → Reserved Instances — cover your minimum always-on capacity with the deepest discount
2. Variable load (scales with demand) → PAYG or Savings Plans — cover the capacity above your base load at PAYG or with a Savings Plan discount
3. Batch and non-production workloads → Spot VMs — maximise savings on tolerant workloads

For a £500k/year Azure compute environment:

• £200k in stable base load → 3-year RIs → £80k–£144k savings
• £200k in variable load → 1-year Savings Plan → £30k–£94k savings
• £100k in batch/non-prod → Spot VMs → £60k–£90k savings
• Total annual saving: £170k–£328k on £500k spend (34–66%)

Implementation: Where to Start

The pricing analysis above is straightforward in theory. In practice, the challenge is accurately profiling your workloads before making purchasing commitments. Committing the wrong workloads to RIs — or underestimating Spot eviction frequency — erodes the expected savings.

Techseria's Azure Pricing Optimisation engagement includes:

• 90-day workload profiling across your Azure environment
• Workload classification using the decision matrix above
• RI purchasing recommendations with specific SKUs, scopes, and terms
• Spot VM feasibility assessment per workload
• Savings Plan vs. RI analysis with scenario modelling
• Implementation support for Spot VM checkpointing where required

Fixed-fee pricing: from £8,500 for environments up to £250k/year Azure spend; from £15,000 for larger environments. Typical first-year saving: 8–12x the engagement cost.

[Book a Strategy Session](https://techseria.com/contact) to discuss your Azure compute footprint and get a workload-specific pricing recommendation. Or [get a fixed-fee quote](https://techseria.com/contact) if you are ready to move directly to the optimisation engagement.