What a commitment actually buys

Why it works that way

A Savings Plan does not reserve anything. It is a promise to spend a fixed number of dollars every hour, for one or three years, against whatever eligible usage you happen to have. AWS defines it as a commitment “measured in $/hour” — the plan names a dollar figure, not an instance. One year means 365 days. Three years means 1,095 days. The smallest commitment AWS will sell you is $0.00001 per hour.

That is why a Savings Plan looks wrong on a bill. There is no plan-shaped object anywhere in your account. No resource points at it, and it points at no resource. It arrives on the invoice as a fee identified by a plan ARN, and the discount it produces lands somewhere else entirely — possibly in a different account.

A Reserved Instance is the older idea and a different one. You name attributes: instance family, size, platform, tenancy, and either a Region or one specific Availability Zone. AWS then discounts usage that matches. Whether you also get capacity depends entirely on that last choice, and it is the thing people buy wrong. A regional RI does not reserve capacity — AWS says so in as many words; it is a price, nothing more. A zonal RI does reserve capacity, in the Availability Zone you named. If you bought RIs so that a launch would never fail, and you bought them regional, you bought a discount and no insurance.

There are four Savings Plan types, not three. Database Savings Plans arrived at re:Invent 2025 and the older pages have not all caught up:

• Compute Savings Plans — EC2 in any family, any size, any Region, any OS, any tenancy, plus Fargate and Lambda. Up to 66% off On-Demand.
• EC2 Instance Savings Plans — one instance family in one Region, with size, OS and tenancy still flexible inside it. Up to 72% off.
• Database Savings Plans — Aurora, RDS, DynamoDB, ElastiCache, DocumentDB, Neptune, Timestream, Keyspaces, DMS and OpenSearch Service. Latest provisioned generations plus serverless usage. Up to 35% off, one-year term, no upfront.
• SageMaker AI Savings Plans — SageMaker AI instance usage across family, size, Region and component. Up to 64% off.

Read that list top to bottom and the trade is obvious: the narrower the plan, the deeper the discount. You are paid for giving up optionality.

Now the mechanic that produces the confusing lines. Every hour, independently, AWS sorts your eligible usage and spends the commitment on it. Free Tier is applied first, then Reserved Instances, then Savings Plans — EC2 Instance plans before Compute plans, because Compute plans are broader and are worth saving for usage the narrow ones cannot touch. Within what is left, AWS calculates the savings percentage of every eligible combination and applies the commitment to the highest percentage first, breaking ties by the lowest plan rate. It continues until the usage runs out or the commitment does. Whatever remains is billed on demand.

Two things follow immediately. The hour is the unit of accounting, so commitment left unspent at 14:00 is not available at 15:00. And nothing in that sort is scoped to an account: a plan can be purchased in any account in the organisation, and with sharing on — the default — it covers the purchasing account’s usage first and then everyone else’s, chasing the largest calculated saving. The fee stays with the buyer. The benefit goes wherever the discount is deepest. That is the whole explanation for a fee sitting alone in the management account while the savings show up in a team’s account three levels down.

What it costs

The commitment is dollars per hour; what you get for it depends on which plan and how long you sign for. One instance — m5.xlarge, Linux, shared tenancy, us-east-1 — priced from AWS’s own Price List API, published 5 August 2026:

• On-Demand — $0.192 per hour.
• Compute Savings Plan, 1 year, no upfront — $0.141. About 27% off, and the rate follows you to any family, size or Region.
• EC2 Instance Savings Plan, 1 year, no upfront, m5 in us-east-1 — $0.121. About 37% off, and it follows you nowhere.
• Compute Savings Plan, 3 years, all upfront — $0.088. About 54% off.
• EC2 Instance Savings Plan, 3 years, all upfront — $0.072. That is 62.5%, and it is the deepest rate this instance can reach in a product advertised as “up to 72%”. The headline percentage is a ceiling across the whole catalogue, not a number you can apply to your instance.

None of that is what confuses people. The line items are. A bill with commitments on it carries charge types that a bill without them never shows, and several of them are designed to cancel each other out:

• SavingsPlanUpfrontFee — the one-time payment for an All Upfront or Partial Upfront plan, charged on the day of purchase.
• SavingsPlanRecurringFee — the hourly charge for a No Upfront or Partial Upfront plan, added every hour of the term. On an All Upfront plan this line still appears, and there it reports the portion of the plan you did not use during the billing period.
• SavingsPlanCoveredUsage — the usage that received the benefit. Its unblended cost is what On-Demand would have been, which is why this row looks like you saved nothing.
• SavingsPlanNegation — the offsetting negative. AWS groups it hourly by plan ARN, operation, usage type and Availability Zone, so one negation row can cancel several covered-usage rows and the counts will not match. The real number is in savingsPlan/SavingsPlanEffectiveCost, not in either row.
• RIFee — the recurring subscription fee for a Reserved Instance. It is monthly, unlike the Savings Plans recurring fee, which is hourly. It appears even for All Upfront RIs where it is $0.00, because AWS uses that row to carry reservation/AmortizedUpfrontFeeForBillingPeriod and reservation/ReservationARN. A $0.00 RIFee is not a bug; it is a container.
• DiscountedUsage — instance usage covered by an RI. AWS documents that the unblended rate on these lines is zero. Those are your $0.00 rows, and they mean covered, not free.

Which brings you to the three cost columns, which disagree with each other on purpose:

• Unblended is cash on the day charged. Upfront fees land on the purchase date, recurring fees on the first of the month. Use it to answer “what did AWS actually charge, and when”.
• Amortised spreads the commitment across the term. AWS’s own example: a $365 All Upfront one-year plan with matching usage shows $1 of amortised cost per day. Use it to answer “what did this workload really cost to run”.
• Blended is an average rate per SKU across the whole organisation, calculated at the management account and used to allocate cost down to members. AWS states plainly that the blended rate shown in the console is informational. Use it for internal chargeback if your organisation has decided to socialise the discount; never use it to reconcile against an invoice, because it will not tie out.

This page carries no meter badges, and that is the point. A commitment does not turn a meter. The instance-hour still turns time whether it is billed at $0.192 or $0.088; egress out of that instance still turns bytes and is not covered by any Savings Plan; a Lambda invocation still turns units. The plan sits beside the meter and edits the price.

The Lambda rate card makes it literal. A Compute Savings Plan covers Lambda, and its us-east-1 rate card lists both meters: duration at $0.0000147 per GB-second, and requests at $0.0000002 each — which is the On-Demand request price. The units meter is inside the plan’s scope and discounted by nothing, and AWS’s own worked example scores Lambda requests at zero percent savings. Same meter, same count, same rate, plan or no plan.

Traps

Unused commitment is spent, not saved. The commitment is a floor on your spend, not a voucher you draw down. Commit to $10 an hour, run $6 of eligible usage in some hour, and the other $4 is gone — AWS is explicit that unused benefit does not carry into the next hour. This is what makes a plan sized on peak usage expensive: you pay for the peak in every trough. On an All Upfront plan you can read the damage directly, because the SavingsPlanRecurringFee line reports exactly the unused portion for the period.

Coverage is not utilisation, and the fix for one is the opposite of the fix for the other. Utilisation is the percentage of your commitment that got used — $9.80 of usage against a $10 commitment is 98%. Coverage is the percentage of your eligible usage that a plan paid for. High utilisation with low coverage means you under-committed and should buy more. Low utilisation means you over-committed and there is nothing to do but wait. A dashboard showing one number without the other will point you the wrong way.

The commitment outlives the workload that justified it. The terms cannot be changed after purchase; you can only buy additional plans on top. The escape hatch is narrow and AWS documents every edge of it: a plan can be returned only if its hourly commitment is $100 or less, only within seven days of purchase, only within the same calendar month, and only up to ten returns per management account per calendar year. Migrate the account into a different organisation and even that closes — AWS states you cannot return the plan from the new organisation regardless of the seven-day window. A three-year plan is three years of a spending decision made by whoever ran the numbers on a workload that may not exist next quarter.