How we cut a client's cloud bill by 40%
No features removed, no performance lost. Just a disciplined audit of what was actually running and why. The playbook, step by step.
The bill had crept from $4k to $19k a month in eighteen months, the way cloud bills do: nobody's fault, everybody's fault. Each service made sense when it was added; nobody owned the total. Our audit brought it back to $11k without touching a single feature.
First: make cost observable. You cannot optimise what you cannot attribute, so before changing anything we tagged every resource by service and team. A third of the spend turned out to belong to environments and experiments nobody remembered starting.
Second: right-size with data, not guesses. Most instances were provisioned for a launch-day peak that never returned. Sizing to the p95 of actual usage, with autoscaling for the real spikes, cut compute by a third on its own.
Third: the quiet money. Storage lifecycle rules for logs nobody reads after thirty days, reserved capacity for the workloads that never sleep, and killing cross-region traffic that existed only because two services were deployed carelessly. None of it is glamorous; all of it recurs monthly.
The lasting fix isn't the audit; it's the habit. Cost now shows up on the same dashboard as latency and errors, reviewed in the same weekly meeting. Cloud spend is an engineering metric; treat it like one and it behaves.