The cost of downtime for an online business, calculated
Nines are a format that makes large differences look small. Converted to minutes, and read against what a provider has actually committed to, the reliability decision stops being a matter of taste and becomes arithmetic you can do in an afternoon.
Downtime gets discussed as an engineering failure and budgeted as an act of God. Both framings avoid the only question that decides what you should spend on it: what does an hour off air actually cost this business. Until that number exists, every conversation about redundancy is a conversation about taste.
The number is calculable, and it is usually smaller than the engineering instinct and larger than the finance instinct. What follows is the arithmetic, the published availability terms you are actually buying, and the point where paying for more reliability stops being rational.
What the percentages mean in minutes
Availability is quoted in nines, which is a format designed to make large differences look small. Convert it to time and the comparison becomes legible. A thirty day month holds 43,200 minutes, so the allowance at each level is simple division.
| Availability | Downtime per 30 day month | Downtime per year |
|---|---|---|
| 99% | about 7 hours 12 minutes | about 3 days 15 hours |
| 99.5% | about 3 hours 36 minutes | about 1 day 19 hours |
| 99.9% | about 43 minutes | about 8 hours 46 minutes |
| 99.95% | about 21 minutes | about 4 hours 23 minutes |
| 99.99% | about 4 minutes | about 53 minutes |
Those are arithmetic from the percentages, not measurements of anybody’s service. The useful observation is the shape. Going from 99% to 99.9% buys back six and a half hours a month. Going from 99.9% to 99.99% buys back thirty nine minutes. The first step changes how a business feels to its customers. The second is mostly invisible to them and expensive to you.
What your provider has actually promised
It is worth reading the agreement rather than the marketing page, because the promise is narrower than the number suggests. Amazon’s EC2 service level agreement commits to commercially reasonable efforts to achieve a Monthly Uptime Percentage of at least 99.99% for instances deployed across two or more availability zones in the same region. A single instance gets a separate and much lower commitment, an Instance-Level Uptime Percentage of at least 99.5%.
Read those two figures together, because the gap between them is the entire argument for architecture. The 99.99% headline is conditional on running in more than one zone. If you run one server, the published commitment is 99.5%, which the table above puts at roughly three and a half hours a month. Most small online businesses are buying the second number while quoting the first.
Then look at what a breach is worth. The EC2 credits are 10% of the bill when uptime falls below 99.99% but stays at or above 99.0%, 30% below 99.0%, and 100% below 95.0%. The remedy is a discount on hosting, not compensation for lost trade. An outage that costs you a day of orders is refunded with a fraction of a hosting invoice. That asymmetry is the reason the cost of downtime is yours to model and not your provider’s to cover.
Unavailability is also defined tightly. For a single instance the test is that it has no external connectivity. A server that is reachable but returning errors, or answering in forty seconds, may not count as unavailable under the agreement while being entirely unavailable to a customer.
The arithmetic for your own business
Four inputs, all of which you already have.
- Revenue per trading hour. Annual online revenue divided by the hours you actually trade, not by 8,760. If demand is concentrated in evenings or in one season, use the peak rate, because outages do not politely occur at 4am.
- Recoverable share. Some orders come back an hour later. Subscription renewals almost always do. One-off impulse purchases and first-time visitors mostly do not. Estimate the fraction you lose for good rather than assuming either extreme.
- Staff cost of the incident. People stop doing their jobs during an outage, and the engineering time is the smaller half. Support answering messages, operations contacting customers by hand, management explaining.
- The tail. Refunds, chargebacks, support contacts for the following week, and the advertising you paid for that landed on an error page. Paid traffic during an outage is the line most often forgotten and the easiest to quantify, because the spend is sitting in the platform.
Multiply the first by the second for direct loss per hour, add the third and the fourth, and you have a defensible hourly figure. Now multiply it by the annual downtime row that matches your current setup. That product is your real reliability budget, and in our experience it is the first time the conversation becomes a business decision rather than an argument between caution and cost.
If the resulting number is small, that is a finding and not a failure. Plenty of businesses discover that a few hours a year is genuinely tolerable, and the right answer is to spend the money on something customers can see.
Where the money should go first
The common mistake is to spend on redundancy before spending on recovery. The question that decides the ordering is not how often you go down, it is how long you stay down and whether you find out before your customers do.
In rough order of return: monitoring that alerts a human, because an outage discovered by a customer email is an outage that ran all night. Then backups that have been restored at least once, since an untested backup is a belief rather than a control, which we worked through in a data backup strategy in plain terms. Then a static holding page that can be served when the application cannot, which converts a dead site into a visibly handled incident. Only then multi-zone redundancy, which is the expensive step and the one that changes the hosting bill shape, discussed in what is driving the cloud hosting bill.
Why 100% is the wrong target
This is settled in the reliability literature and worth borrowing. Google’s Site Reliability Engineering chapter on service level objectives states that "although 100% availability is impossible, near-100% availability is often readily achievable", and it distinguishes three things that get conflated in these discussions. An indicator is "a carefully defined quantitative measure of some aspect of the level of service that is provided". An objective is "a target value or range of values for a service level that is measured by an SLI". An agreement is "an explicit or implicit contract with your users that includes consequences of meeting (or missing) the SLOs they contain".
The test offered there is clean. If there is no consequence for missing a target, it is an objective rather than an agreement. Most small businesses have neither, which means nobody has decided what reliability is for.
The same chapter recommends an error budget, "a rate at which the SLOs can be missed", tracked weekly, and it advises against setting targets from current performance and against absolutes. It also gives a reference point that is worth keeping in mind before promising customers more: it cites Google Compute Engine’s own published availability target as 99.95%, which is three and a half nines, not five.
If an infrastructure provider at that scale publishes 99.95%, a promise of anything higher on a small custom system deserves scrutiny. The honest commitment is usually a lower number with a tested recovery process behind it.
Honest limits
The minutes table is arithmetic, not a benchmark. It converts percentages to time on a thirty day month and nothing more. It is not a claim about what any provider delivers, and real months vary in length.
The AWS terms were read on 17 November 2026 and they change. Figures, credit tiers and definitions were taken from the EC2 service level agreement on that date. Other providers differ, service level agreements are revised, and your own contract may carry negotiated terms. Read yours rather than ours.
We have given no cost of downtime figure. Industry averages for this circulate widely and they are dominated by enterprises whose revenue per hour has nothing to do with a regional online business. Publishing one would be worse than useless, which is why the section above is a method rather than a number.
Availability is not the only failure mode, and often not the worst. Silent data corruption, a failed payment integration that accepts orders it cannot charge, and a slow site that nobody abandons loudly all cost money while uptime monitoring stays green. A site that is up and wrong can be more expensive than a site that is down, and it is much harder to notice. Treat the uptime number as the easiest part of reliability to measure rather than the most important part to get right. The broader pattern, and why averages mislead here too, is in why the average is the wrong number, and the running cost side is in what a web app costs per month to run.
Describe it. We build it.
Seven or twelve days, pay on delivery, a year of maintenance included. Bring the problem, not a spec.
Book a meetingRead next
A phased software rollout plan, and the order that holds
A phased software rollout plan borrowed from reliability practice: why one branch goes first, the few signals worth watching, and keeping a way back.
ReadAn employee leave management system, and the rules it has to encode
An employee leave management system is a balance engine, not a form. The entitlement bands, the holiday calendar you cannot hardcode, and what to build first.
Read