Strategy / commitment instruments
Cloud Discount Mechanisms 2026: Savings Plans vs RIs vs CUDs vs SUDs
Every provider has a different discount model with different trade-offs. This is the decision framework with worked examples: thirteen mechanisms in one ranked table, a flowchart for choosing, total annual cost at $100K monthly spend, the common mistakes, and when enterprise agreements pay off.
Discount stack / 13 mechanisms across providers
The full discount landscape
| Mechanism | Provider | Max discount | Commitment | Flexibility |
|---|---|---|---|---|
| Compute Savings Plans | AWS | up to 66% | 1 or 3 years, hourly $ | Across families, regions, EC2 / Fargate / Lambda |
| EC2 Instance Savings Plans | AWS | up to 72% | 1 or 3 years, hourly $ | Locked to family + region |
| Standard Reserved Instances | AWS | up to 72% | 1 or 3 years | Locked to instance type |
| Convertible Reserved Instances | AWS | up to 66% | 1 or 3 years | Exchangeable for different family |
| Reserved VM Instances | Azure | up to 72% | 1 or 3 years | Locked to family or instance + region |
| Azure Savings Plan for compute | Azure | up to 65% | 1 or 3 years, hourly $ | Across VM families, regions, OS |
| Azure Hybrid Benefit | Azure | up to 80% with RIs (Windows Server, Microsoft figure) | Existing on-prem licences | Windows, SQL, RHEL, SLES |
| Committed Use Discounts (resource) | GCP | up to 37% (1y), up to 70% (3y) | 1 or 3 years per machine type | Locked to family in region. General-purpose 37% / 55%; memory-optimised up to 70% at 3y |
| Committed Use Discounts (spend) | GCP | 28% (1y), 46% (3y) | 1 or 3 years, hourly $ | Across general-purpose families |
| Sustained Use Discounts | GCP | up to 30% | none, automatic | Apply to most general-purpose VMs |
| Spot Instances | AWS | up to 90% | none | Interruptible, 2-min notice |
| Spot VMs | Azure | up to 90% | none | Eviction-based, 30-sec notice |
| Spot VMs | GCP | up to 91% | none | No 24-hour limit |
Decision framework / pick by workload pattern
Which discount fits which workload
If
Workload stable for 1+ year, single configuration
Goal
Maximum discount, locked
Pick
AWS Standard RI / Azure RI / GCP CUD resource-based
If
Workload stable but composition changes
Goal
Discount with flexibility
Pick
AWS Compute Savings Plan / Azure Savings Plan / GCP CUD spend-based
If
Variable usage, hard to predict commitment
Goal
Automatic discount or no commit
Pick
GCP SUDs (automatic) or stay on-demand and rely on Spot for buffer
If
Fault-tolerant, can handle interruption
Goal
Spot / Preemptible
Pick
Spot Instances (AWS) / Spot VMs (Azure / GCP)
If
Microsoft-heavy estate (Windows, SQL Server)
Goal
Stack Hybrid Benefit on top of RIs
Pick
Azure Hybrid Benefit + Azure RI
If
Large, steady spend you can forecast over several years
Goal
Ask about a negotiated agreement
Pick
AWS private pricing (EDP) / Azure EA or MCA / GCP committed spend contract
Worked example / annual cost at $100K monthly spend
What different strategies actually cost
| Strategy | AWS | Azure | GCP |
|---|---|---|---|
| Baseline: $100K/mo on-demand, no commitments | $1,200,000/yr | $1,200,000/yr | $1,200,000/yr |
| 70% Savings Plan / RI coverage, 3-year all-upfront | approx $700,000/yr (-42%) | approx $700,000/yr (-42%) | approx $720,000/yr (-40%) |
| 70% commit + 20% Spot on tolerant workloads | approx $560,000/yr (-53%) | approx $565,000/yr (-53%) | approx $585,000/yr (-51%) |
Our model, rounded: committed usage is assumed to earn an average 60% discount on AWS and Azure and 57% on GCP, and Spot usage 55%, both below the published ceilings in the table above. Your realised rates depend on instance mix, term, payment option and interruption tolerance.
Six common mistakes
Where teams overpay
Buying too many commitments (stranded capacity)
Coverage above 80% leaves no headroom for workload reduction. Most teams should target 60-75% coverage and accept some on-demand exposure.
Not reviewing utilisation quarterly
Workload composition changes. A Savings Plan bought 2 years ago for x86 m5 may be stranded if you migrated to Graviton. Schedule quarterly reviews.
Ignoring Spot for eligible workloads
CI runners, batch processing, and ML training are all Spot candidates. Most teams undercut savings by leaving these on on-demand.
Committing before rightsizing
Buying RIs for over-provisioned instances locks in the waste. Run rightsizing first, then commit to the right size.
Ignoring Hybrid Benefit on Microsoft estates
The single biggest lever on Azure for Windows-heavy estates. Audit eligibility annually as Software Assurance terms change.
Mixing 1-year and 3-year RIs randomly
Build a coverage curve: 3-year for the bottom of the steady-state, 1-year above that, on-demand at the top. Reconsider whenever workload steps change.
Enterprise agreements / above the public price list
When negotiation pays off
AWS
Private pricing / Enterprise Discount Program (EDP)
Terms not published
Negotiated privately with AWS against a committed spend over a multi-year term. AWS publishes no qualifying spend level and no discount schedule for it.
Azure
Enterprise Agreement (EA) / Microsoft Customer Agreement (MCA)
Terms not published
Negotiated commercial agreement with Microsoft or a partner. Any discount is set in the agreement itself; Microsoft publishes no Azure spend level at which a discount starts.
GCP
Committed spend contract
Terms not published
Negotiated with Google Cloud sales against a committed spend. Pricing is specific to the contract; Google publishes no qualifying threshold.
Common questions
FAQ
Compute Savings Plan or Reserved Instance, which is better in 2026?+
For most teams in 2026, Compute Savings Plans are the right default. They cover EC2, Fargate, and Lambda, work across instance families and regions, and adapt to architecture changes (x86 to Graviton, EC2 to Fargate, region migrations). Standard RIs offer slightly higher ceilings (72% vs 66%) but only for fixed configurations. Use RIs when you have a workload that genuinely will not change for 1-3 years.
How much can I save with reserved instances?+
On AWS and Azure: up to 72% for 3-year all-upfront RIs versus on-demand. On GCP: up to 70% for 3-year resource-based CUDs. Realised savings depend on coverage ratio (target 60-75%) and workload stability. Most teams realise an effective 35-50% blended discount once you account for partial coverage and the on-demand spike portion.
Can I stack discounts across mechanisms?+
Yes, with rules. Savings Plans and RIs are mutually exclusive (Savings Plans apply first, then RIs to remaining usage). Hybrid Benefit stacks on top of Reserved Instances on Azure. Spot is a separate purchasing model and stacks with everything for eligible workloads. Enterprise discounts (EDP, EA) layer on top of the underlying mechanism.
When does it make sense to negotiate an EDP or EA?+
None of the three providers publishes a spend level at which a negotiated agreement becomes available, or what discount it carries; both are set in the contract. It is worth raising with your account team once you can forecast a large, steady spend over several years and have already used the public mechanisms above (Savings Plans, Reserved Instances, CUDs). Compare any offer against what those public mechanisms already give you, and weigh the commitment and its shortfall terms, before signing.
What is the difference between Spot, Preemptible, and Spot VMs?+
Functionally equivalent: interruptible compute at 60-90% off list. AWS Spot has been GA the longest. GCP Preemptible was the legacy GCP option (24-hour max runtime); GCP Spot VMs replace it with no time limit. Azure Spot VMs evict on capacity or price thresholds. All three give 30 seconds to 2 minutes of interruption notice.
Continue reading