EmberKat

Environment

When the estate grew without a decision, and the bill grew with it

Three conditions usually travel together. Where they hold, the problem is not that anyone was careless — it is that creating something takes a minute and retiring it takes an argument nobody is authorised to win.

Book 30 minutes
An expanding software and cloud estate contains several empty ownership nodes while one cost line continues to rise.
01

Nobody can say what a given machine, service or subscription is for, or who owns it.

02

New things appear faster than old things are retired, so spend grows by default.

03

There is no regulator to answer to. The pressure is the budget and the complexity behind it.

01 // Where this applies

Two conditions, and neither one is cloud versus on-premise

This is not an argument about hosting. It is about whether anything in the estate can be traced to a decision, and whether anyone is accountable for the total.

The environment

Ordered by where the money is actually leaking.

  • A cloud estate that nobody can inventory Four hundred machines, and the safest available action is to leave every one of them running.
  • Subscriptions that outlived the reason they were bought A pilot ended, a team reorganised, a person left. The renewal did not notice any of it.
  • A platform adopted because it was current Kubernetes, a data platform, a model endpoint. Chosen against a trend rather than against a need, and now operated at full cost.

The organisation

This is where the problem turns out not to be technical.

  • No single person sees the whole bill It is split across teams, each of which can only justify its own line.
  • Deleting something is riskier than keeping it Not because it is used, but because nobody can prove it is not. Inaction is the rational choice.
  • The same evaluation has been run before And reached no conclusion, because no one was appointed to make one.

02 // Why it accretes

Two forces, and only one of them has anyone pushing back

Most estates have not separated them. Separating them is the difference between a cost exercise that holds and one that grows back within two quarters.

Direction one — creation is frictionless

Provisioning is a minute of work, needs no capital approval and produces something that visibly helps today. Every incentive in the organisation points that way.

Retirement is the opposite: it needs someone to establish what a thing does, who depends on it, and who carries the risk if that turns out to be wrong. That work is unrewarded and the risk is personal, so it does not happen.

Direction two — nobody owns the total

Each team can defend its own spend, and each defence is reasonable. The sum is indefensible, and the sum has no owner.

This is why cost programmes led from finance rarely hold. The lever is engineering ownership and a record of decisions, not a spreadsheet exercise imposed from outside.

The correction most teams need

Cost is not the problem to solve; it is the symptom that makes an ownership problem visible. An estate where every resource traces to a stated purpose and a named owner gets cheaper as a by-product, and stays cheaper. An estate that is cut without that traceability grows back, because the conditions that produced it are untouched. Which is also why the first deliverable here is a decision record rather than a savings number.

03 // What actually goes wrong

Four failure modes, and none of them is overspending on purpose

01

Four hundred virtual machines and no inventory. Nobody can authorise deleting one, because nobody can prove it is unused.

Cost: The bill is paid indefinitely, and the only safe action is inaction.

02

A pilot ended and the subscription did not. It runs with no user, no owner and a renewal that clears automatically.

Cost: Recurring spend for a capability that has no one left to ask about it.

03

A platform was chosen because it was what serious companies were adopting, not because a stated need required it.

Cost: The licence is the smaller half. The larger half is the people and the time it takes to keep it alive.

04

The estate was cut once, hard, without recording why anything was kept or removed.

Cost: It grows back within a year, and the second attempt is met with the argument that the first one broke something.

04 // What you will be asked for

The four things a decision needs, and who can actually supply them

The fourth column is the part most estates have never separated: the mechanism that should carry each answer, as opposed to the person who gets asked for it.

Resource lifecycle against the answer required, its owner, the mechanism that should carry it and the failure it forecloses.
Phase What you will be asked for Who produces it Where the duty comes from What it forecloses
Creation What this exists for, stated at the moment it is created rather than reconstructed later. Which service or team it belongs to. Whoever provisions it — enforced by the platform, not by goodwill. Tagging at creationNaming and metadata policy Applied when the resource appears, or it is archaeology afterwards. An estate that can only be described by guessing.
Ownership A named person accountable for the resource, and a total that someone above them owns. Service owner, with one accountable owner for the whole bill. Cost allocation to a team Allocation is what turns an unarguable total into arguable lines. Every team defending its own spend while the sum has no defender.
Renewal and commitment What the committed term buys, what happens if usage falls, and the date the decision has to be made by. Whoever signs — informed by whoever operates it. Decision record with criteria and exit The renewal date is the decision point. After it, the same choice costs a full term. Committing again to something nobody has evaluated since the first time.
Retirement Evidence a thing is unused, a reversible way to remove it, and someone empowered to accept the residual risk. Service owner, with the risk accepted in writing rather than carried personally. Staged decommission with a rollback Making removal reversible is what makes it approvable. Inaction being the only career-safe option.

None of this is specific to a cloud provider, and none of it is specific to infrastructure. The same four questions apply to a licence estate, a data platform, an internal tool nobody has opened in a year, and a contractor arrangement that renews on its own. Where the answers exist, cost is a consequence of decisions. Where they do not, cost is weather.

A model endpoint is the newest version of the same problem, and the fastest-moving one. Adopted quickly, priced per call rather than per month, and rarely traced to a workflow anyone can name. The rows do not change: what it is for, who owns it, what the commitment buys, and how it comes out again if the answer is that it was never needed.

References

The mechanisms named above, linked

There is no legislation behind this page, which is precisely the difference between it and the other environments. The references are the providers’ own guidance on the mechanisms, so the practices above are checkable rather than asserted.

  1. 01
    AWS — Best Practices for Tagging AWS Resources

    Why metadata is applied at creation, and what breaks when it is applied retrospectively instead.

  2. 02
    Microsoft — Define your tagging strategy

    The Cloud Adoption Framework guidance on naming and metadata, including enforcing it rather than requesting it.

  3. 03
    FinOps Foundation — Framework

    The operating model for allocating cost to the teams that create it. Useful for the allocation mechanics; it is not a substitute for engineering ownership.

Links resolved 28 July 2026. All three are free to read.

05 // The scene, which you will recognise

Four hundred machines and nobody to ask

It is the same scene in most large organisations, which is why no client is named for it. Something was spun up for a deadline. It worked. The person who understood it moved teams, or moved on. The tag was never applied, the wiki page was never written, and the machine kept running because stopping it required someone to be sure.

Multiply that by a few years of deadlines and you have an estate nobody can describe. Ask who owns a given resource and the honest answer is that the question has no addressee. Ask whether it can be switched off and the honest answer is that finding out costs more than the machine does — which is exactly why it is still there.

The arithmetic is worth running with your own numbers, because it is often not close. Take one month of the bill and set it against buying the hardware outright and paying someone for most of a year to look after it. Sometimes it comes out the other way, and then you have something better than a saving: a documented reason to stay, which stops the same argument being reopened every budget cycle.

06 // How this starts

A decision first, because a cut without one grows back

Most engagements begin with fixed scope, because the work should be judged before a relationship is committed to. The first deliverable is a decision record rather than a savings number — the saving follows from the ownership being fixed, and does not survive without it.

If you are assessing this as a supplier

Eitautas Šimaitis The practitioner who scopes the work, forms the recommendation, and stays through implementation.

The work is carried from scope to closure by the same named person, and nothing is recommended that pays a commission: no reseller margin on a platform, a licence or a migration. That matters more here than elsewhere, because in cost work the party giving the advice is usually also selling the destination.

Next step

Bring the bill and the thing nobody can explain

Thirty minutes, to work out whether I’m the right person for what you’re dealing with. Bring the decision that keeps being postponed, the system nobody will switch off, and the date a commitment renews. No credentials or exports in the booking form — we’ll agree a secure route first.

Book 30 minutes