Summary

  • “Monthly Uptime Percentage” is computed on three incompatible bases across Amazon EC2, Amazon RDS and Amazon S3 — minutes of Unavailability, one-minute intervals, and an average of five-minute request error rates — and two of the three treat unobserved time as perfect availability.
  • The remedy is a percentage of a bill line applied only to future payments, never a refund, and it exists only if the customer opens a Support Center case inside the second billing cycle after the incident, with request logs the customer must already hold.

A service-level agreement is meant to answer one question: when the thing you paid for stops working, what are you owed? Amazon Web Services answers it in separate documents, by the category, and the three agreements read here do not agree on what the question means.

Start with the promise. “AWS commits to offer Service Level Agreements (SLAs) for all paid, generally available services,” the AWS service level agreements index states, and then lists its headings: ANALYTICS, APPLICATION INTEGRATION, ARTIFICIAL INTELLIGENCE, BUSINESS PRODUCTIVITY, COMPUTE, DATABASE, DEVELOPER TOOLS, END USER COMPUTING, GAME TECH, INTERNET OF THINGS, MANAGEMENT TOOLS, MEDIA SERVICES, MIGRATION & MODERNIZATION, NETWORKING & CONTENT DELIVERY, QUANTUM, SATELLITE, SECURITY & IDENTITY, STORAGE. Amazon Elastic Compute Cloud (EC2) sits under COMPUTE, Amazon RDS under DATABASE, Amazon S3 under STORAGE. Eighteen headings, eighteen chances for the same word to mean something else.

In the Amazon Compute Service Level Agreement, last updated 25 May 2022, “Monthly Uptime Percentage” is calculated by subtracting from 100% the percentage of minutes during the month in which Amazon EC2 was in the state of Unavailability, measured during any monthly billing cycle. The commitment is 99.99% at the region level and 99.5% at the instance level. Unavailability is a narrow state: a Single EC2 Instance with no external connectivity, or, at the region level, running instances in two or more Availability Zones — or one Availability Zone and one in another region — concurrently without external connectivity.

The Amazon RDS Service Level Agreement, last updated 22 January 2024, counts differently: 100% minus the percentage of one-minute intervals during the monthly billing cycle in which the resource was Unavailable, against 99.95% for Multi-AZ DB Instances and Multi-AZ DB Clusters and 99.5% for Single-DB Instances. Its Definitions section then changes the base: “If you have been running that Single-DB Instance for only part of the month, your Single-DB Instance is assumed to be 100% available for the portion of the month that it was not running”, with the same assumption extended to Multi-AZ resources.

The Amazon S3 Service Level Agreement, last updated 28 November 2023, abandons minutes altogether. Monthly Uptime Percentage is 100% minus the average of the Error Rates from each five-minute interval in the monthly billing cycle; an Error Rate is the count of internal server errors — InternalError or ServiceUnavailable — divided by the total requests of that type in the interval. And intervals with no requests are assumed to have a 0% error rate.

Two of those three definitions reward silence. A quiet bucket cannot miss a request it never received; a database that ran for nine days of a thirty-day month is scored as though it had been up for the other twenty-one. Take my own illustration of the S3 rule, which is arithmetic and not a measurement: a thirty-day month holds 8,640 five-minute intervals. If 8,000 of them carried no requests at all and the remaining 640 each carried a 10% error rate, the averaged Error Rate is 0.74% and the Monthly Uptime Percentage 99.26% — above the first credit threshold, because 8,000 empty intervals counted as perfect.

The money follows the measurement, and it is not standardised either. EC2 pays 10% of the relevant monthly bill where the monthly figure falls below the commitment but stays at or above 99.0%, 30% where it falls below 99.0% but stays at or above 95.0%, and 100% below 95.0%. S3 prints two tables: 10%, 25% and 100% for Standard, Express One Zone, Glacier Flexible Retrieval, Glacier Deep Archive and requests not otherwise specified, and 10%, 25% and 100% on different thresholds for Intelligent-Tiering, Standard-IA, One Zone-IA and Glacier Instant Retrieval. RDS pays 10%, 25% and 100%. So the band that pays 30% in EC2 pays 25% in S3 and RDS.

On a hypothetical $1,000 monthly charge in that band — again my arithmetic, not a published figure — that is $300 under the Compute table and $250 under the other two.

What the credit is a fraction of matters as much as the fraction. Under EC2 it is the monthly bill for EC2 in the affected region, excluding one-time payments such as upfront Reserved Instance payments, or the bill for the affected Single EC2 Instance. Under S3 it is the total charges paid for the applicable storage class in the affected Region — or Availability Zone for S3 Express One Zone — for the billing cycle in which the figure fell in the band.

None of the three agreements prints a cap beyond its own 100% tier, and each says plainly what the credit is not: it applies only to future payments for the affected service, issues only where greater than $1 USD, is non-transferable, and “will not entitle you to any refund or other payment from AWS.”

The AWS Customer Agreement, last updated 29 October 2026, keeps that boundary and adds two of its own. Service Credits sit outside the liability disclaimer and outside the twelve-month damages cap, so the cap does not swallow them; but the customer is entitled to no credits “for any period of suspension.” The Agreement incorporates the agreements by reference rather than restating them, which is why the operative numbers live in eighteen separate places.

Then come the conditions, and they are where the remedy actually lives. To receive a Compute Service Credit the customer must open a case in the AWS Support Center, received by the end of the second billing cycle after which the incident occurred, and supply a subject line, the dates and times, the affected region and — at the instance level — the Availability Zone, the affected resource IDs, and request logs documenting the errors. Missing information disqualifies the claim, and the same incident cannot be claimed under both the region-level and the instance-level agreement.

RDS and S3 repeat the pattern, S3 asking for the subject line “SLA Credit Request” and the dates and times of every incident of non-zero Error Rates, and both issuing a credit within one billing cycle after AWS confirms the request. All three require the customer to hold logs proving an outage that the customer’s own monitoring had to notice first.

One remedy in the reviewed set needs no claim at all, and it runs on a different clock: “AWS will not charge you for any Single EC2 Instance that is Unavailable for more than six minutes of a clockhour,” the Compute agreement states, applying automatically. Six minutes out of sixty, hour by hour, is not the monthly percentage and does not convert into it. Two clocks, two units.

No document I read obliges AWS to tell a customer that a commitment was missed. Across the three agreements, the Customer Agreement — which I read in part, to Section 4.1 — and Service Terms sections 1.11 to 1.13, the negative is bounded to exactly that material, and the Service Terms portion stops at section 1.17.3. What the Service Terms do carry, at section 1.12, is a fourth route: “If you have been charged for a Service for a period when that Service was unavailable (as defined in the applicable Service Level Agreement for each Service), you may request a Service credit equal to any charged amounts for such period.” That measure is none of the percentages. It is still a request, and no document in the set says which measure governs the same outage.

The perimeter is finished off by exclusion and exclusivity. The four grounds headed “Amazon Compute SLA Exclusions” — factors outside reasonable control, including force majeure or Internet problems beyond the demarcation point; the customer’s own acts or omissions, including failing to acknowledge a recovery volume or respond to resource health concerns; the customer’s equipment, software or other technology; suspension or termination under the Agreement — apply directly or indirectly, and a parallel list sits in the S3 agreement.

Where availability is affected by factors outside the uptime calculation, AWS may issue a credit at its discretion. S3’s remedy sentence is explicit: “Unless otherwise provided in the AWS Agreement, your sole and exclusive remedy for any unavailability, non-performance, or other failure by us to provide Amazon S3 is the receipt of a Service Credit (if eligible) in accordance with the terms of this SLA.” The Compute agreement frames the same idea as its sole and exclusive remedies and AWS’s sole and exclusive obligations.

None of this is a discovery about honesty; it is a description of an instrument, and it is a shape this publication has not read before. Our coverage of this company has taken the regional dependency map and the evacuation option sold separately from it, FBA capacity pricing, group margin and cash flow, and KDP royalty tiers. None of those reads a commercial service-level agreement as the remedy instrument. Our directory entry for Amazon.com, Inc. records a number-resource footprint and says nothing about the cloud business (the entry).

The measure of a remedy is whether an outsider can reconstruct it. Here a customer can reconstruct the percentage, because the tier tables are printed; it can reconstruct the base, because the bill line is its own; and it cannot reconstruct the figure, because the figure comes from the provider’s own accounting of minutes, intervals or error rates and is published nowhere. What would change the answer is unglamorous: a notification duty, a per-incident record the customer can read, one unit of availability across services, and a remedy that is money rather than a discount on next month.

Until one of those appears, the availability number is computed by the seller, and the remedy is a claim the buyer has two billing cycles to file.