SLA uptime percentages explained: what 99.9 vs 99.99 actually promise
Each nine is ten times less allowed downtime. The monthly arithmetic, what service credits really pay, and how many nines your business needs.
Every hosting contract quotes nines: 99.9%, 99.99%, 99.999%. The differences look cosmetic and are not - each nine is an order of magnitude less allowed downtime, and the price of achieving it climbs just as fast. Understanding what an SLA actually promises, and what it pays when broken, changes how you evaluate hosting quotes.
This guide explains the arithmetic, what service credits really are, and what to ask a provider, using Amazon's published Compute SLA as a concrete example, checked on 19 November 2026.
The arithmetic of nines
A month has about 43,830 minutes. What each availability level permits in that month:
- 99.9% ("three nines") allows roughly 44 minutes of downtime per month - a bad morning, every month, within the terms.
- 99.99% ("four nines") allows roughly 4 minutes and 23 seconds per month.
- 99.999% ("five nines") allows about 26 seconds per month.
Moving from three nines to four is not a 0.09% improvement; it is ten times less permitted failure. Architecturally that jump usually means redundancy across failure domains - and it costs accordingly.
What an SLA actually pays when broken
An SLA is not an insurance policy on your business losses. AWS's Compute SLA is typical: Amazon commits to a region-level percentage (99.99% for instances deployed across two or more Availability Zones in the same region), and if it misses, you become eligible for a service credit - a percentage of your bill, applied only against future payments for that service. Credits are tiered by how badly the target was missed, and they require you to claim them. If an outage costs you a day of sales, the SLA returns a slice of your hosting invoice, not the sales.
The details that decide what the promise is worth
- The measurement unit. AWS's region-level commitment only applies when you deploy across two or more Availability Zones. A single instance carries a weaker instance-level commitment. The architecture you choose decides which promise you get.
- What counts as downtime. Scheduled maintenance, your own configuration errors, and failures of things outside the provider's control are commonly excluded. Read the exclusions before comparing percentages.
- Monthly vs annual measurement. A monthly window resets the clock often; an annual figure can hide several bad months.
- Claim mechanics. Credits usually must be claimed within a set window with evidence. If nobody on your team would notice and file the claim, the credit does not exist for you.
How many nines you actually need
Start from what an hour of downtime costs your business, not from what sounds premium. A brochure site that serves as a contact page tolerates three nines comfortably. Checkout and booking flows that lose money per minute, and anything contractually promised to your own customers, justify four nines and the redundancy work behind it. Five nines belongs to genuinely critical infrastructure - the engineering and operations cost is real and continuous. Whatever you choose, your own monitoring matters more than the contract: you cannot claim credits, or even know you had a bad month, without measuring uptime yourself.
Sources
- Amazon Web Services, Amazon Compute Service Level Agreement: https://aws.amazon.com/ec2/sla/
Drafted with AI assistance from this source. SLA terms differ across providers and products; read the specific SLA document for each service you buy, including its exclusions and claim procedure.