A cloud bill is not a list of what you use. It is a list of what you are charged for, organised for billing rather than for understanding, and the gap between those two things is where the money goes. Teams that set out to reduce spend usually start with instance sizes, because that is the line item they recognise. It is rarely where the surplus is.
The recurring pattern is that overspend concentrates in resources nobody owns. A volume left behind after an instance was terminated. A load balancer serving an environment decommissioned last year. Snapshots on an automated schedule with no expiry. A NAT gateway routing traffic that could go via an endpoint. Log retention set to forever because that was the default and no one chose otherwise. None of these were decisions; each was the absence of one, and none will be noticed by a team looking at compute.
Data transfer is the second concentration and the least understood. Traffic between availability zones, out to the internet, and across regions is priced differently in each direction, and an architecture that spreads chatty services across zones for resilience can pay for that chatter continuously. The bill shows it as a modest line with an opaque name. Tracing it back to a specific service calling another one is real work, which is why it is usually left alone.
Non-production is the third. Development and staging environments running at production scale, twenty-four hours a day, for a team that works eight, is one of the most reliable savings available in any organisation, and one of the least popular to raise because it is nobody's fault and everybody's habit. Scheduling them down outside working hours is straightforward and the saving is arithmetic.
Before optimising anything, make ownership visible. A tagging standard that attributes every resource to a team and an environment, enforced at creation rather than audited afterwards, converts the bill from a single number into a set of numbers each of which somebody recognises as theirs. Almost every meaningful reduction we have seen followed that step rather than preceded it, because unowned resources do not get cleaned up by people who cannot tell whether they are needed.
Then be careful about the order of operations. Buying reserved capacity or savings plans before rightsizing locks in the waste for one to three years — a commitment against the wrong baseline is worse than no commitment. Clean up what is unowned, switch off what is idle, rightsize what remains, and only then commit to the level you have settled at. Doing it in that order is the difference between a saving and a longer-term liability that looks like one.
