Where a 37% AWS saving actually comes from
Nobody finds a third of a cloud bill in one heroic change. The real breakdown from a twelve-month engagement, line by line.
Every cost-optimisation pitch implies there is a switch somewhere. There is not. What there is, on almost every AWS account I have looked at since 2018, is a dozen small things that each look too minor to bother with and together account for a third of the bill.
Here is the actual breakdown from one twelve-month engagement, in the order I did the work, with what each part returned.
Measurement first, which returned nothing
Three days on cost allocation tags across every resource, a per-service breakdown, and a monthly report the founders could read without me in the room.
Savings: zero. This is the step everyone wants to skip, and skipping it is why most optimisation efforts stall after the obvious wins. You cannot argue for turning something off if nobody can tell you what it does.
Rightsizing against real utilisation — about 14%
Instances chosen during a load test in 2023 and never revisited. The pattern is always the same: someone sized for a peak that never arrived, the peak got designed out of the product six months later, and the instance stayed.
The rule I use is boring: two weeks of CloudWatch data at p95, not p100, and never resize more than one tier at a time.
Storage lifecycle — about 6%
Unattached EBS volumes. Snapshots of instances that no longer exist. Logs in S3 Standard from 2022 that nobody has queried since the week they were written.
Lifecycle policies are unglamorous and they compound quietly.
Cross-AZ data transfer — about 5%
This one is invisible until you tag for it. Services chatting across availability zones because the scheduler placed them wherever there was room, at a cent or two per gigabyte, all day, forever.
Committed use — about 12%
Savings Plans, last. Deliberately last: committing to a baseline you have not yet stabilised means committing to your own waste for one to three years.
What did not work
I spent a week on a Graviton migration that returned less than 2% after the compatibility work, and one service could not move at all. I have also never once found meaningful savings in the place clients expect it, which is their compute instance type catalogue.
The part that matters
Twelve months in, spend was down 37% and it is still falling, because the monthly review never stopped. Cost optimisation is not a project with an end date — the moment it becomes one, the bill starts climbing again. That is precisely why it is included in every tier rather than sold as an engagement.
Figures are rounded and drawn from more than one account, so no client is identifiable. The proportions are typical.