Three precise line items: disks still billing on stopped VMs (€60/mo), seven reserved IP addresses nobody uses (€50/mo), and snapshots accumulating with no upper bound (€70/mo recoverable). One hard cleanup, rollback points kept.

52 test/dev services are configured to stay permanently warm, 24/7. Six of them bill full compute around the clock (≈ €43/mo each); the other 46 hold warm instances at roughly €9.50/mo per lane. Test traffic does not justify either. The fix is a small, per-lane, fully reversible configuration change; the work is deciding which lanes qualify and validating them. Several sibling lanes already wake on demand, proving the pattern in this estate.

Eighteen virtual machines (≈ €1,540/mo of compute) serve environments that are used at most a quarter of the hours in a month: nights, weekends and holidays included. They can be put on an automatic calendar so they run when the team does; data and disks are untouched. The catch: overnight jobs and backup windows have to be mapped first, which is where estates usually get this wrong.

Two managed databases sized like production serve test environments around the clock (≈ €380/mo combined). They can sleep out of hours, and their sizing deserves a second look. Which lever to pull first depends on how the test pipelines use them.

Two premium compute-optimised build machines (≈ €420/mo) wait for jobs most of the day. Two distinct paths: move them to a cheaper machine family (a hard, one-off change), or switch to build capacity created on demand per job at heavily discounted rates, the standard end-state for CI. The second needs a proper setup to keep build start times snappy. (Overlaps with the calendar above; the simulator accounts for it.)

60% of 140 storage buckets have no retention or tiering rules. Each fix below is priced from the measured data; the client’s retention obligations are confirmed before any rule is applied.

Not a single committed-use discount exists anywhere in the estate: the genuinely 24/7 production workloads (≈ €2,900/mo of compute) have been paying full on-demand rates all along. Google discounts committed spend by roughly 28% (1-year) or 46% (3-year). The catch, and where this most often goes wrong: the commitment must be sized on the optimised baseline, in the right order relative to everything above. Get the sequence wrong and you lock today's waste in for years.