Cloud waste is a tagging habit, and that 76% shutdown number needs a second look

The tag policy is the only item on that list that everything else depends on. A scheduler stopping non-production compute needs a way to know what is non-production, and the example in the post targets instances by env=dev ^[5]. The orphan sweep needs an owner to send the ticket to ^[9]. The cost-by-tag report that exposes a test environment running production-sized instances for months cannot exist before the tags do ^[3]. So the order is fixed: mandatory env, project and owner tags first, enforced by something that stops resources lacking them ^[2]. Enforcement is the part teams flinch at, because it converts a forgotten label into a stopped instance.

The 76% figure is arithmetically clean. Forty hours out of 168 is 23.8% of the week, so the saving is 76.2% ^[13]. But the schedule that produces it is not the schedule printed alongside it. Stopping at 7 PM and starting at 7 AM every day leaves twelve hours running seven days a week, which is 84 of 168 hours and a 50% cut ^[14]. Getting to 40 hours needs an eight-hour weekday window and both weekend days dark ^[15]. That in turn means the scheduler has to tell an internal dashboard that can be off on Sunday from a dev cluster someone is actually using on Sunday, and the only thing carrying that information is the tag.

Then the ceiling. The author reports roughly 40% off the total bill ^[19]. If overnight shutdown of non-production compute at 76% were doing all of that work, non-production compute would have to be about 53% of the bill to begin with ^[16]. For most shops it is not, and the post gives no per-lever split, so the 40% is one practitioner's outcome rather than a forecast for yours.

Three of the remaining items are not habits at all. Spot capacity at 60-90% below on-demand requires the application to survive reclaim and shut down gracefully ^[7], which is code someone has to write and test. Swapping a $50 managed database for a $15 self-hosted Postgres saves $35 a month, or $420 a year, a 70% cut on that line ^[10][18], while moving backups and patching onto your own on-call. Keeping traffic inside one region and fronting static assets with a CDN ^[11] is topology. All real money, and all of it the refactor the piece opens by saying you can skip ^[1].

What separates the governance levers from those is who has to remember. Budget alerts firing at 50, 90 and 100% of a per-project budget ^[8] keep firing after the engineer who set them changes teams, and a policy that stops untagged resources keeps stopping them. The weekly cost review ^[12] does not have that property, which is why it is the item most likely to quietly lapse in month four.