Where your cloud bill actually goes
Most over-spending is not the compute everyone stares at. It is data transfer, idle non-production environments, and storage nobody has looked at in two years.
When a client asks us to look at cloud cost, the instinct in the room is always to look at instance sizes. That is occasionally where the money is. More often it is somewhere duller.
Data transfer
Egress is the line item that surprises people, because nothing in the architecture diagram suggests it. Cross-availability-zone chatter between services, a database replica in another region, an asset served from the origin instead of a CDN.
Chatty services placed in different zones can cost more in transfer than in compute. Nothing about that is visible until you read the bill by line rather than by total.
Environments nobody uses
Staging running production-sized instances at 3am on a Sunday. The environment from a project that shipped last year. Review apps that were never torn down.
Non-production is frequently a third of the bill, and most of it is idle. Scheduled shutdown outside working hours is the highest return-per-hour change available in cloud cost work.
Storage that only grows
Snapshots with no retention policy. Logs at full verbosity retained for years because nobody chose a number. Multi-region replication on a bucket that does not need it.
Storage is cheap per gigabyte, which is exactly why it is never reviewed and why it quietly compounds.
Right-size against measurement
Then, and only then, look at instance sizes - against actual utilisation, not against the number someone guessed at launch. Most fleets are provisioned for a peak that was estimated, never observed, and never revisited.
We have taken infrastructure spend down 60% on a system serving identical traffic. Almost none of that came from the thing the team expected.