Summary
- Amazon's compute and storage agreements each promise a "Monthly Uptime Percentage", but the compute agreement subtracts the share of minutes in the month when EC2 was in the state of Unavailability, while the storage agreement subtracts the average of per-request error rates across five-minute intervals.
- The storage agreement states that intervals with no requests are assumed to have a 0% error rate, so a quiet five minutes scores as a perfect one.
- The only consequence in either document is a service credit banded at 10, 30 or 100 per cent, claimable through a support case with prescribed wording and logs, received by the end of the second billing cycle, and described as the customer's sole and exclusive remedy.
Two documents, one phrase
The Amazon Compute Service Level Agreement, last updated 25 May 2022, opens with a sentence worth reading slowly: "AWS will use commercially reasonable efforts to make Amazon EC2 available for each AWS region with a Monthly Uptime Percentage of at least 99.99%, in each case during any monthly billing cycle". A second commitment covers individual instances at 99.5 per cent.
Then comes the definition. "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. The denominator is the calendar, the numerator is a state, and the state is narrow: loss of external connectivity. At region level it is narrower still, requiring all running instances deployed across two or more Availability Zones to be without external connectivity at the same time.
The Amazon S3 Service Level Agreement, last updated 28 November 2023, uses the same phrase to mean something else. There the Error Rate is the number of internal server errors returned as InternalError or ServiceUnavailable, divided by the total number of requests for that request type during the applicable five-minute interval. Monthly Uptime Percentage is 100% minus the average of those interval error rates across the billing cycle.
Minutes and requests are not exchangeable. An hour in which a region answers nothing is 100 per cent unavailable under the first construction. The same hour, if almost nobody was asking, can read as nearly perfect under the second — because the S3 document adds the sentence that settles it: intervals with no requests are assumed to have a 0% Error Rate. Silence is not merely unpunished. Silence is scored.
The denominator decides what counts
This is where a procurement file goes wrong. Two contracts, one vendor, one metric name, and a reader comparing them side by side would find two different quantities. The compute figure is a claim about elapsed time. The storage figure is a claim about the traffic that happened to arrive, averaged with equal weight per interval regardless of volume.
That averaging has a consequence worth stating plainly: a five-minute interval carrying a single failed request and a five-minute interval carrying a million failed requests enter the mean with the same weight. Total errors over total requests across the month would be a different number. Which one a buyer wants depends on what the buyer is buying — a service that answers whenever called, or a service whose published ratio improves when nobody calls.
The region-level definition compounds this. Unavailability requires a joint failure across Availability Zones, so a failure that takes out one zone, however complete, is not region-level Unavailability. It may be an instance-level event for the customers inside it. The scope is stated, not hidden; it simply is not what an executive means by "the region was down".
What the customer must do, and by when
Neither agreement pays out on its own. A credit requires a case in the AWS Support Center, a prescribed subject line — "Amazon Compute SLA Credit Request – Region-Level Claim" for the compute region figure — the dates, times and region of each incident, resource IDs and request logs documenting the errors. The request must be "received by AWS by the end of the second billing cycle after which the incident occurred." Failure to provide the required information disqualifies the customer. Region-level and instance-level claims cannot be stacked for the same instance.
The bands are printed. Compute: below 99.99 per cent but at or above 99.0 per cent earns 10 per cent; below 99.0 but at or above 95.0 earns 30; below 95 earns 100. Storage: below 99.9 but at or above 99.0 earns 10; below 99.0 but at or above 95.0 earns 25; below 95 earns 100, with different thresholds for the infrequent-access and Intelligent-Tiering classes.
And both documents close the door behind the credit. The compute agreement states that it sets forth the customer's sole and exclusive remedies and AWS's sole and exclusive obligations for unavailability or non-performance. Reading it is not an accusation. It is a description of what the instrument is for: a pre-priced discount schedule, triggered by a reading the buyer has to demand and document, inside a clock the buyer has to watch.
The AWS Service Terms, last updated 17 November 2026, look at first like a broader promise: if you have been charged for a service for a period when it was unavailable, "you may request a Service credit equal to any charged amounts for such period." Then the parenthesis does the work — unavailable "as defined in the applicable Service Level Agreement for each Service." The umbrella clause does not widen the definition. It defers to it.
The one rule that needs no claim
Buried in the compute agreement is a sentence of a different shape: AWS "will not charge for any Single EC2 Instance that is Unavailable for more than six minutes of a clockhour, automatically and without a credit request."
Automatic matters. No subject line, no logs, no second billing cycle. It is the one remedy in the documents read that does not depend on the customer filing anything, and it operates on a window — the clockhour — that has nothing to do with the monthly percentage.
My arithmetic, and it is mine, not the document's: a 30-day month holds 43,200 minutes, so one hundredth of one per cent of it is 4.32 minutes. That is the whole region-level allowance at 99.99 per cent for such a month. Six minutes of a clockhour is already larger than that monthly allowance, and the two are not measuring the same thing. A buyer who reads the six-minute rule as a generous version of the monthly promise has merged two instruments.
What this instrument cannot measure
An availability guarantee begins where delivery begins. It measures whether a service already provisioned answered. Nothing in the claim machinery read here reaches the question an AI-era buyer asks most often: whether the capacity can be obtained at all — whether the instance launches, whether the quota increase is granted, whether the region has room. In the three documents read, the credit path is triggered by Unavailability or Error Rate after the fact. A wait is not an outage, and no clause quoted here converts one into the other.
That distinction is the practical one. Uptime percentages get copied into board decks as evidence of dependability. What they actually record is the seller's own arithmetic about a subset of failures the seller defines, remedied by a credit the buyer must chase. The number is not a lie. It is a contract term wearing a measurement's clothes.
Prior coverage, measured on 18 November 2026: a literal search of this desk's archive for "service level agreement" returned no article. Recent briefings here read a threshold a regulator computes, a statistical office's own fraction, a statutory reporting perimeter and a banned charging line. None read a seller-authored, seller-measured contractual percentage whose only consequence is a claim the buyer must file. That is the delta, and the subject is Amazon Web Services, Inc. as it appears in this directory.
One scoping note, placed where it belongs. Across the three documents read, none names an independent measurer, auditor or third-party verification path, and none specifies the monitoring system that decides when Unavailability began and ended. That is an observation about the passages obtained on 18 November 2026, not a finding that no such mechanism exists. The base to which the credit percentage applies also was not stated in those passages, so no credit amount is computed here.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






