A cloud invoice can run to thousands of lines, which makes it feel unanalysable. It usually is not. For most workloads a handful of categories explain the bulk of the total, and the useful skill is knowing which to look at first rather than reading the whole thing.
Compute: priced by time, not by use
Virtual machines are typically billed for the period they exist in a running state, regardless of how busy they are. An idle instance and a saturated one of the same size usually cost the same. That single fact explains a large share of unexpected bills: instances created for a test and never removed, environments that run through nights and weekends when nobody uses them, and oversized instances chosen defensively at launch and never revisited.
Where to look first
Sort compute by cost and check the top entries against what you believe is running. Then look for anything running continuously that only needs to run during working hours.
Storage: priced by what is kept, including what you forgot
Storage is billed by volume retained over time, and the category is wider than most people picture. Beyond the obvious disks and buckets, it includes snapshots, backups, old machine images, log archives and orphaned volumes left behind when an instance was deleted. None of these announce themselves, and all of them accumulate.
Storage classes matter here. Most object storage offers cheaper tiers for data that is accessed rarely, with the trade-off that retrieval costs more or takes longer. Data that is kept for compliance and read almost never is usually in the wrong class by default.
Data transfer: the one people forget
Transfer is the category that most often surprises teams, because it is invisible in the architecture diagram. The general pattern across providers is that traffic inbound is cheap or free, traffic within one zone is cheap, traffic between zones or regions costs more, and traffic leaving the provider’s network to the internet costs most. A design that moves large volumes across those boundaries can spend more on moving data than on processing it.
Managed services: convenience with a line item
Managed databases, queues, load balancers, Kubernetes control planes and similar services carry their own charges, sometimes a fixed hourly fee per instance of the service plus usage. Individually small, collectively significant, and easy to leave running after the project that needed them has ended.
Support and licensing
Paid support plans are often a percentage of spend, so they scale with everything else. Commercial operating systems and some database engines carry licence costs folded into the hourly rate, which is why two instances with identical specifications can differ in price.
A practical monthly routine
- Compare the total against last month and identify the largest absolute change, not the largest percentage.
- Check the top ten resources by cost and confirm each is still needed at its current size.
- Look for resources with no owner or no tag, which are usually either forgotten or undocumented.
- Review storage for snapshots and volumes not attached to anything.
- Check transfer charges against the architecture, and ask whether any large flow crosses a boundary unnecessarily.
Guardrails beat analysis
Budgets with alerts catch problems in days rather than at the end of the month. Scheduled shutdown for non-production environments removes a recurring cost without any ongoing attention. Both are configured once and keep working, which is more than can be said for a monthly review that depends on someone remembering to do it.