Cloud is a utility. Your invoice disagrees.

Elasticity you never use is just an expensive lease

The commodity promise of cloud is real: compute, storage and bandwidth on demand, priced per unit, no procurement cycle. The failure mode is equally real. Organisations move a static estate into an elastic environment, never change the shape of the workload, and end up renting by the hour something they used to own by the year.

Industry surveys have put self-reported cloud waste around a third of spend for years, and our own reviews land in the same neighbourhood. The waste is rarely one dramatic mistake. It is a hundred small ones: oversized instances chosen 'to be safe', dev environments running through the weekend, snapshots nobody expires, and egress charges from a data flow that should never have crossed a boundary.

We reviewed one client's estate where a single misconfigured cross-region replication job, set up for a disaster-recovery test two years earlier and never torn down, was responsible for 11% of total monthly cloud spend. Nobody had looked at it because it never generated an error, an alert, or a support ticket. It just quietly moved terabytes across a region boundary every night and billed accordingly.

The four line items that decide your bill

• Idle: capacity running when nobody is asking for anything. Schedule it off. This is free money.
• Oversizing: the gap between provisioned and used. Right-size quarterly, automatically.
• Egress: the tax on architectures that move data across boundaries for no good reason.
• Managed premium: the convenience surcharge on services you could run on plain compute, sometimes worth it, often not measured.

Notice that three of those four are design problems, not procurement problems. Negotiating a discount on a wasteful estate just means you waste at a discount. Fix the shape first, then negotiate, you will have real consumption data to negotiate with.

The managed premium line deserves more scrutiny than it usually gets. A managed database service at 2.5x the cost of self-run compute can be an excellent trade if it removes genuine operational burden from a small team. It is a terrible trade if your team already runs the equivalent open-source stack competently elsewhere and is paying for a convenience it does not need. We have seen both decisions made by the same organisation for different services in the same quarter, with no consistent logic connecting them.

Commodity discipline, applied

The organisations that get cloud economics right treat the bill as a product metric. Cost per order. Cost per thousand page views. Cost per model inference. Once a team can see what its feature costs per unit of business output, optimisation stops being a finance campaign and becomes engineering hygiene.

This requires tagging discipline that most organisations claim to have and almost none actually enforce. We have audited 'fully tagged' estates where 40% of monthly spend sat under a generic 'shared-services' tag that had become a dumping ground for anything nobody wanted to categorise properly. A tag that cannot answer 'who pays for this and why' is not a tag, it is decoration.

FinOps is a rate, not a project

Most companies that try to fix cloud cost run a one-off 'optimisation sprint', shave 15% off the bill, declare victory, and watch spend creep back to the old trajectory within two quarters because the underlying incentives never changed. Engineers are still rewarded for shipping features fast, not for shipping features cheap, so cost regresses to the mean the moment attention moves elsewhere.

• Put unit cost (per order, per user, per inference) on the same dashboard as latency and error rate.
• Give every engineering team a monthly cloud budget they own and can see in real time, not a finance report they see quarterly.
• Make cost review part of the same ritual as the sprint retro, five minutes, every two weeks, not a special occasion.
• Reward teams that reduce unit cost the same way you reward teams that reduce latency.

None of this is exotic. It is the same discipline manufacturing applies to cost per unit produced, applied to a factory that happens to be virtual. The companies that treat cloud spend as a recurring engineering signal instead of a recurring finance surprise are the ones whose bill actually tracks their growth instead of outrunning it.

The exit question

A commodity is something you can leave. You do not have to leave, but you should know the number. What would it cost, in weeks and euros, to move the three most expensive workloads to another provider or back on-premise? If nobody can answer, you have discovered your real vendor risk, and it is not priced in your contract.