Azure Reservations vs Savings Plans: the commitment decision.
Reservations save 40 to 72 percent on steady-state compute; Savings Plans save 11 to 65 percent and apply across regions, SKUs and OS. The choice is not discount headline — it is workload predictability. This note is the 2026 per-workload decision matrix, with the exchange, refund and Hybrid Benefit mechanics that decide net cost.
The choice is workload predictability, not headline percentage. Take Reservations for steady-state VMs locked to a known SKU and region — they discount deepest, up to 72% over three years. Take Savings Plans for variable spend where the future family mix is unknown — shallower discount, but no utilisation risk. Reservations strand value when workloads move; Savings Plans carry no exchange, cancellation or refund.
01 Key findings
The decision is predictability, not discount magnitude. A 3-year Reservation on a Standard_D8s_v5 in East US runs ~62% off pay-as-you-go; the Savings Plan on the same instance runs ~34%. The Reservation wins by 28 points — only if that exact SKU runs in that exact region for the full term.
Reservations discount deeper but carry utilisation risk. Best-in-class estates run 95 to 100% reservation utilisation. Below 85%, the commitment is no longer beating pay-as-you-go and should be exchanged or refunded.
Savings Plans cannot be exchanged, cancelled, or refunded. The full hourly commitment is owed for the full term regardless of consumption. This is the single most consequential commercial difference between the two vehicles.
Size Savings Plans against the floor, never the average. Commit against the 95th-percentile compute floor. Any consumption above the commitment simply bills at pay-as-you-go; underutilisation is a sunk cost to Microsoft.
The optimal portfolio blends both, and Hybrid Benefit stacks on either. Reservations cover the stable floor, Savings Plans the drift, pay-as-you-go the burst. AHB layered on Windows workloads pushes the blended discount toward 55%.
02 Commitment scorecard
Relative strength across the dimensions that decide which vehicle fits a workload. Five dots = strongest. Scoring reflects commercial posture, not raw discount depth alone.
03 2026 pricing snapshot
Both vehicles are 1-year or 3-year prepay or monthly commitments applied against compute usage. Neither covers storage, networking, software-included costs, or PaaS consumption — those take service-specific Reservations. List discounts, per instance family and term:
| Commitment | 1-year discount | 3-year discount | Coverage scope |
|---|---|---|---|
| Reserved VM Instance (blended) | 40 to 48 percent | 60 to 72 percent | Specific instance family, region, OS |
| Reserved VM (Standard_D-series) | ~36 percent | ~62 percent | D2-D96 v5, fixed region |
| Reserved VM (Standard_E memory) | ~41 percent | ~64 percent | E2-E96 v5, fixed region |
| Reserved VM (Standard_M memory-opt.) | ~45 percent | ~68 percent | M-series, fixed region |
| Savings Plan for Compute | ~11 percent | 28 to 65 percent | Across families, regions, OS |
| Reserved SQL DB vCore | ~22 percent | ~33 percent | Specific vCore commitment |
| Reserved Cosmos DB throughput | ~20 percent | ~65 percent | Specific RU/sec commitment |
A $10 per hour 3-year Savings Plan commits to $87,600 per year and $262,800 over the term. Reservation size-flexibility matches within a series: a Standard_D8s_v5 reservation fully matches a D8s_v5, partially covers a D16s_v5 (2 units), and covers a D4s_v5 (0.5 units).
04 Vehicle profiles
- 60 to 72% off compute on 3-year terms
- Exchangeable for equal or greater value, no fee
- Cover SQL, Cosmos, Synapse, App Service and more
- Discount applies automatically on matching VMs
- Exact match on instance series and region
- Utilisation risk below 85% erodes the discount
- Refunds capped at $50,000 per 12-month period
- Applies across families, regions, OS and VM size
- No utilisation risk — spend flexes with usage
- Covers ACI, Container Apps, Functions Premium
- Excess consumption simply bills at pay-as-you-go
- No exchange, cancellation, or refund mechanism
- Lower discount (~11% at 1yr, 28 to 65% at 3yr)
- Excludes Spot, software-included, and PaaS reservations
Over-committing to Reservations is the most common mistake: estates that land in the 70 to 85% utilisation band earn an effective discount below what a Savings Plan would have delivered — paying for depth they never realise. The mirror error is under-committing and running stable production at full list price. Model the compute floor first, then buy Reservations to it and let a Savings Plan absorb the drift on top.
05 Exchange & refund mechanics
Flexibility is where the real long-run risk sits. When SKUs drift or workloads migrate, the exit terms decide whether adjustment is manageable or a write-off. The two vehicles are structurally different here.
| Mechanic | Reservations | Savings Plans |
|---|---|---|
| Exchange | Equal or greater value, no fee | Not permitted |
| Cancellation / refund | Capped at $50,000 per 12 months per enrollment | Not permitted |
| Utilisation risk | Real below 85%; exchange or refund | None — flexes across all eligible usage |
| Term commitment | 1 or 3 years; adjustable via exchange | 1 or 3 years; full amount owed regardless |
| Sizing guidance | Size to the stable SKU footprint | Size to the 95th-percentile compute floor |
The 2023 policy change ended free unlimited Reservation refunds — cancellations beyond the $50,000 cap now require Microsoft approval and may incur fees. Savings Plans never had a refund path. Size both conservatively before signing, not after.
06 Per-workload decision matrix
The right vehicle depends on three workload properties: SKU stability, region stability, and term certainty. Match the pattern to the vehicle before committing a dollar.
| Workload pattern | Recommended vehicle | Why |
|---|---|---|
| Production VMs, known SKU and region, stable for 3 years | 3-year Reserved Instance | Highest discount, low utilisation risk |
| Known SKU and region, migration planned in 12-18 months | 1-year Reserved Instance | Captures most of the discount with year-end flexibility |
| Stable spend but uncertain SKU mix | 3-year Savings Plan for Compute | Lower discount, but no stranded reservation cost |
| Dev/test running 8 hours per day | Auto-shutdown + pay-as-you-go | Commitments pay for unused hours; shutdown is cheaper |
| Burst workloads, high variability | Spot VMs or pay-as-you-go | No commitment beats not committing on true variable load |
| SQL Database, Cosmos DB, Synapse | Service-specific Reservation | Compute Savings Plans do not cover PaaS services |
| Mixed estate: predictable floor + variable peak | Reservations for the floor, Savings Plan on top | Best discount on the floor, flexibility on the rest |
07 Hybrid Benefit interaction
Azure Hybrid Benefit (AHB) is a separate optimisation that converts on-premise Windows Server or SQL Server licences with active Software Assurance into Azure compute discounts. Critically, AHB stacks with both Reservations and Savings Plans — it is not an either/or choice.
A Windows VM under AHB pays the Linux compute rate plus a small adjustment, removing the Windows licence component (roughly $0.092 per vCPU per hour). A 3-year Reserved Instance on a Standard_D8s_v5 Windows VM in East US lists at about $1,180 per month; adding AHB drops it to about $580 — a further 51% saving on top of the reservation discount. It is the single largest cost lever available to Windows-heavy estates.
AHB requires Software Assurance entitlements for the underlying cores. The audit risk is real: claiming AHB without sufficient SA triggers a compliance event during Microsoft SAM engagements. See our Azure Hybrid Benefit guide for the entitlement and audit rules.
The optimised mix for a $10M/year Azure compute estate: 50 to 60% on 3-year Reservations for the stable floor, 20 to 25% on a 3-year Savings Plan for SKU drift, 10 to 15% on 1-year vehicles as a planning buffer, and the balance on pay-as-you-go or Spot. Blended effective discount lands at 40 to 55% — the upper end when AHB is layered on Windows workloads.
08 Our recommendation
You run production VMs on a known SKU in a fixed region with a 3-year horizon. Buy to the stable floor, drive utilisation to 95%+, and use no-fee exchanges when families shift. Layer AHB on every eligible Windows core.
You have stable spend but an unpredictable SKU or region mix, and you want to eliminate stranded-commitment risk. Size to the 95th-percentile compute floor — never the average — because there is no exchange, cancellation, or refund.
09 Commercial levers
Both vehicles are list-priced — the discount tables are fixed by Microsoft and not separately negotiable. The leverage sits in the surrounding agreement structure.
MACC burn-down
Reservation and Savings Plan consumption counts toward the Microsoft Azure Consumption Commitment. The MACC tier discount stacks on top, so commitments can partly satisfy a multi-year Azure commitment.
MCA-E level pricing
EA and MCA-E customers receive level-based pricing on the underlying pay-as-you-go rate before the commitment discount, compounding the saving. Level D typically runs 12 to 18% below Level A.
Exchange & refund terms
The standard exchange rights and $50,000 refund cap can be improved for large customers, with bespoke caps written into the MACC or EA amendment. Negotiate flexibility upfront if migration is planned.
Portfolio re-modelling
Independent reviews recover $1.2M to $3.8M per year on $10M of Azure compute by re-modelling the Reservation and Savings Plan mix — without changing workload behaviour.
For the broader Azure commercial framework see Azure MACC vs CTP, Azure MACC negotiation, Microsoft EA Complete Guide, and the Microsoft vendor hub.
Stop mis-allocating Azure commitments
Our Cloud & FinOps practice re-models the Reservation and Savings Plan mix against your real workload floor.
The Licensing Edge
Weekly cloud and licensing intelligence for enterprise IT leaders. 3,000+ subscribers.