Ten Moves That Decide Whether an AWS Savings Plan Saves or Strands Money
A Savings Plan cannot be cancelled, resold, or refunded for its full one or three year term, so the seventy to ninety percent coverage target, not the deepest discount, is what protected a representative 12 million dollar compute estate and cut it 55 percent.
Prepared by Redress Compliance · June 2026 · Representative 12 million dollar annual EC2, Fargate, and SageMaker compute estate (benchmark scenario, not a quote)
Executive Summary
AWS Savings Plans are the primary discount mechanism for compute, and the second most important commercial decision after the Enterprise Discount Program itself. The structure is dollar denominated. You commit to a flat hourly spend for one or three years, and AWS discounts your bill against on demand list rates.
The discount ceilings are public. A three year All Upfront Compute Savings Plan reaches about 66 percent, an EC2 Instance Savings Plan about 72 percent, and a SageMaker Savings Plan about 64 percent. The deepest number is also the longest lock and the largest cash outlay.
The risk lives in one sentence. A Savings Plan cannot be cancelled, resold, or refunded, so an unused commitment is a pure write off with no marketplace to recover it. The buyer who accepts the AWS proposed ninety to ninety five percent coverage pays for compute that the production estate no longer matches.
There is a second trap in the contract. EDP attainment is measured on net spend after the Savings Plan discount, so blanket coverage quietly shrinks the spend that earns your enterprise credit and can return as a true up on the floor.
On a representative 12 million dollar annual compute estate, a disciplined coverage target with the right term per workload class models to a 55 percent reduction, to roughly 5.4 million dollars, before the EDP overlay.
This paper gives the cloud owner, CFO, and CPO ten moves to make before signature: build the baseline, choose the instrument, set the term and payment, fix coverage at seventy to ninety percent, convert Reserved Instances cleanly, hedge by workload class, protect the EDP floor, lock five clauses, and anchor a credible BATNA.
Build the verified baseline coverage analysis first
Move one anchors every other move. Build the baseline from the AWS Cost and Usage Report, not the console summary and not the Cost Explorer recommendation. The CUR is line item truth for every hour of compute.
The baseline classifies every dollar of compute by how stable it is. That classification, not the headline discount, decides which instrument and term each workload earns.
What the baseline must contain
- Coverage map: on demand, Savings Plan, and Reserved Instance hours per family, region, and account.
- Workload class: steady, flexible, seasonal, or spiky, which sets the term and the instrument.
- Utilization: CPU, memory, and headroom, so you right size before you commit a single dollar.
- Commitment expiry: the run off date of every existing Reserved Instance and Savings Plan, so renewals do not stack.
Choose between Compute, EC2 Instance, and SageMaker Savings Plans
Three Savings Plan types exist, and the cost math behind each is a trade of discount depth against flexibility. The AWS Savings Plans pricing page sets the ceilings. The decision is which flexibility you are willing to surrender for the deeper rate.
| Instrument | Max saving | Covers | Flexibility | Best for |
|---|---|---|---|---|
| Compute Savings Plan | 66% | EC2, Fargate, Lambda | Highest. Any family, size, region, OS, tenancy, or service | The moving middle of the estate |
| EC2 Instance Savings Plan | 72% | One instance family in one region | Size and OS flex inside the chosen family and region | A fixed, locked production base |
| SageMaker Savings Plan | 64% | SageMaker ML instances | Instance and component flex across SageMaker | A steady training or inference estate |
Maximum saving versus on demand by Savings Plan type. Numbers match the table above.
Read the gap as a flexibility tax. You give up about six points moving from an EC2 Instance plan to a fully flexible Compute plan. Price that gap against the odds your workload changes family or region inside the term.
Pick the term: one year or three year
Term length moves most of the discount. The deepest numbers all require three years, which is exactly where flexibility risk concentrates. A three year lock on a workload that moves in year one is stranded value with no exit.
How the two terms compare
- One year: roughly two thirds of the maximum discount, with annual re scoping and far lower lock in risk.
- Three year: the deepest rate, but a thirty six month bet on workload shape that no buyer can forecast cleanly.
- The split: three year terms on the proven steady core, one year terms on the flexible and seasonal tail.
Do not put the whole estate on one term. The steady base can carry a three year commitment; the rest should stay on a one year plan or on demand until it proves stable.
Pick the payment option: No, Partial, or All Upfront
Payment option trades cash for a few points of discount. All Upfront gives the deepest rate, No Upfront the shallowest, with Partial in between. The uplift from No Upfront to All Upfront is usually small, and it captures no time value of money benefit.
| Term and payment | Compute SP | EC2 Instance SP | Buyer note |
|---|---|---|---|
| 1 year, No Upfront | up to 40% | up to 42% | Maximum flexibility, lowest saving, no cash outlay |
| 1 year, All Upfront | up to 45% | up to 49% | Cash for a modest uplift over No Upfront |
| 3 year, No Upfront | up to 60% | up to 64% | Long lock without the cash outlay |
| 3 year, All Upfront | up to 66% | up to 72% | Deepest saving, longest lock, largest cash bet |
Compare the All Upfront uplift against your internal cost of capital before you pay it. If a few discount points cost a year of locked cash, No Upfront or Partial often wins on a net present value basis.
Set the coverage target at seventy to ninety percent
Coverage discipline is the single most valuable move in this paper. Cover roughly seventy to ninety percent of steady baseline compute and leave the variable tail on demand or on Spot. The unused part of an over sized plan cannot be recovered.
The coverage target
Captures most of the committed discount while leaving headroom for the workload to shrink without stranding commitment.
The AWS proposed level
Often built on an assumed twenty percent year over year growth the customer never committed to internally.
Where the common advice on Savings Plan coverage is wrong
The standard AWS account team pitch is to cover ninety to ninety five percent of projected compute, because high coverage looks efficient and the projection assumes growth. We disagree. Blanket coverage strands the seasonal and spiky tail, because a Savings Plan cannot be cancelled or resold.
In the estates we have benchmarked, it also erodes EDP credit by cutting net spend below the floor. The buyer side move is to cover the proven steady core at seventy to ninety percent and grow coverage only as the workload proves stable. You give up a few headline points and keep the optionality the deepest discount quietly sells.
Convert existing Reserved Instances without paying twice
Most estates already carry Reserved Instances when the Savings Plan conversation opens. Handle the overlap deliberately, or you commit fresh Savings Plan dollars on top of compute an RI already discounts and pay twice for the same hours.
How to sequence the conversion
- Map the run off: list every RI expiry date, then size new Savings Plans to the gap, not the gross estate.
- Let standard RIs expire: on steady load that stays in one family, an RI and an EC2 Instance plan reach the same ceiling, so there is no rush to convert mid term.
- Sell convertible mismatches: on the RI Marketplace where a Standard RI no longer fits, rather than stacking a plan over it.
Hedge across workload classes
Not every workload belongs on a Savings Plan. Match each tier to the instrument that fits its stability, and keep the uncertain tail off any commitment. The classification from move one drives this directly.
| Workload class | Belongs on | Why |
|---|---|---|
| Steady core | 3yr EC2 Instance SP or Standard RI | Fixed family and region, so the deepest locked rate carries no real flexibility cost |
| Flexible steady | 3yr Compute Savings Plan | Stable spend that moves across families or services, so flexibility is worth the six point gap |
| Seasonal | 1yr Compute Savings Plan | Predictable but time bound, so a short term captures discount without a long lock |
| Spiky and uncertain | On demand and Spot | No commitment. Spot covers fault tolerant batch at the steepest hourly discount with zero term |
The hedge is the point. Committing the uncertain tail to capture a few discount points is the most common over commitment we see, and it is the hardest to unwind because nothing refunds.
Protect EDP attainment and the unused commitment recovery position
The Savings Plan layer interacts with the Enterprise Discount Program, and the interaction is not in your favor by default. EDP attainment is measured on net spend after the Savings Plan and Reserved Instance discount, so your own commitment savings reduce the spend that counts toward the enterprise floor.
Model the representative estate to see the mechanic. The same 12 million dollar on demand equivalent estate produces a clean 55 percent reduction when each tier carries the right instrument and coverage stays disciplined.
| Coverage tier | Share | On demand annual | Instrument | Saving | Net annual |
|---|---|---|---|---|---|
| Steady core | 50% | $6.00M | 3yr EC2 Instance SP or Standard RI | 66% | $2.04M |
| Flexible steady | 25% | $3.00M | 3yr Compute Savings Plan | 55% | $1.35M |
| Seasonal | 15% | $1.80M | 1yr Compute Savings Plan | 45% | $0.99M |
| Spiky and uncertain | 10% | $1.20M | On demand and Spot | 15% | $1.02M |
| Total | 100% | $12.00M | Disciplined coverage | 55% | $5.40M |
On demand annual versus net annual by tier, dollars in millions. Totals: $12.00M to $5.40M, a 55 percent reduction. Numbers match the table.
Lock the five clauses that protect the commitment
The instrument sets the discount. The contract decides whether that discount survives an estate that changes. Five clauses do the work, and AWS rarely offers them unprompted.
| Clause | What it locks | Why it matters |
|---|---|---|
| Gross spend attainment | EDP or PPA credit measured before the SP and RI discount | Stops your own savings dropping you below the commitment floor |
| Rate card lock | On demand and committed rates fixed at signature for the term | Blocks list price drift on the uncovered tail |
| Re scope rights | Right to re scope Savings Plan families and exchange RIs | Preserves flexibility as workloads move across the estate |
| Commitment portability | Plans and RIs apply across accounts, regions, and OUs | Stops an org change stranding a commitment |
| Exit and ramp relief | Shortfall carry forward and a defined wind down | Protects against a true up on a shrinking estate |
Saving range by scenario, percent. Bars span the low to high of each benchmark range below.
Renewal on an optimized estate
Deeper three year coverage on a proven steady core, layered above a freshly classified baseline.
Incremental EDP overlay
An Enterprise Discount Program layers a further band on net spend, on top of the Savings Plan discount.
Benchmark ranges: Redress Compliance advisory engagement file, 2024 to 2025.
Build the BATNA and neutralize AWS tactics
A credible best alternative is what turns a list rate into a negotiated one. For compute the alternatives are real and increasingly portable, and the account team has a small set of repeatable plays against them.
- Microsoft Azure: Reserved VM Instances and the Azure savings plan for compute, with a similar one and three year structure.
- Google Cloud: Committed Use Discounts, including flexible spend based commitments that mirror a Compute Savings Plan.
- Spot and Graviton: Spot for fault tolerant batch and self managed compute on Graviton for the steadiest, most portable load.
| AWS tactic | Buyer counter |
|---|---|
| Cover 90 to 95 percent for three years for the deepest rate | Cover the steady core at 70 to 90 percent, keep the variable tail flexible |
| Take a Compute Savings Plan for everything, it is simplest | Layer instruments to workload class, not to admin convenience |
| Net spend still counts toward your EDP | Demand gross spend attainment in the side letter |
| All Upfront for the best discount | Price the small uplift against your internal cost of capital |
| Use the Cost Explorer recommendation, it is free | Build the baseline from the CUR, the recommendation favors AWS |
Size to the estate, not to the deepest discount. A Savings Plan refunds nothing, so coverage discipline and the right term per workload class are worth more than the last few points of headline rate. The disciplined model cut a representative 12 million dollar compute estate by 55 percent before a single contract clause.
- Cover the proven core, hedge the tail. Three year EC2 Instance plans on the fixed base, Compute plans on the moving middle, one year plans on seasonal load, on demand and Spot on the spiky tail. Hold coverage at 70 to 90 percent.
- Lock the attainment definition. Get EDP spend counted before the SP and RI discount in writing, or your own savings drop you below the floor and return as a true up. Pair it with rate lock, re scope rights, portability, and exit relief.
Redress Compliance runs this as a standing engagement: build the verified baseline, model every instrument and term, sequence the Reserved Instance conversion, and draft the five clauses and side letter. We sit on your side of the table from first conversation through signature. We are glad to tie a meaningful part of the fee to delivered value.