Cloud cost problems are rarely caused by expensive resources. They are caused by an invoice nobody can decompose, which means no team can see the effect of their own decisions and optimization only happens when finance escalates. The engineering fix is attribution: make spend a signal that reaches the people who create it, on the same cadence as the rest of their telemetry.
Attribution before optimization
It is tempting to start with rightsizing, because it produces a visible number quickly. It is usually the wrong order. Without attribution you cannot tell whether a saving persisted, whether it was offset by growth elsewhere, or which team's change caused an increase. Build the allocation model first; the optimization backlog then writes itself.
- Agree a tagging standard covering owner, environment, product and cost center — four tags applied consistently beat twelve applied sometimes.
- Enforce tags in infrastructure modules and in policy checks, not in a wiki. Untagged resources should fail review, not generate a monthly chase.
- Decide up front how shared and untaggable costs are allocated. Data transfer, shared clusters and support plans are where allocation models usually stall.
- Publish the allocation rules. A number a team does not understand is a number they will dispute rather than act on.
Showback usually beats chargeback
Chargeback — actually moving budget between teams — creates strong incentives and equally strong political overhead. Showback, where teams see their spend without it hitting their budget, captures most of the behavioral benefit at a fraction of the organizational cost. Start there, and only move to chargeback if showback demonstrably fails to change behavior.
What a useful team dashboard shows
- 1.Current-month spend against the same period last month, so trend is visible without reading a table.
- 2.Cost broken down by environment, so non-production growth is obvious.
- 3.Unit economics where the team has a meaningful denominator — cost per tenant, per order, per thousand requests.
- 4.The top five resources by cost, with a link to who owns them.
- 5.Anomalies flagged when they happen, not at invoice time.
Cost per unit of business value is the metric worth arguing about. Total spend rising while cost per transaction falls is usually a healthy system, and a dashboard that only shows totals will report it as a problem.
Where the recurring savings actually are
The distribution is consistent across estates, even though the amounts are not. Prioritize by recurring monthly spend against implementation effort, and be honest that some of these require product decisions rather than engineering ones.
- Non-production environments running outside working hours, which is often the single largest quick win.
- Idle and orphaned resources — unattached volumes, old snapshots, load balancers pointing at nothing, forgotten test clusters.
- Over-provisioned compute and databases, sized once at launch against a guess and never revisited.
- Storage on the wrong tier, particularly logs and backups retained at hot-tier pricing far beyond their useful life.
- Cross-zone and egress data transfer created by architecture choices nobody costed at design time.
- Commitment coverage for the stable baseline, once usage data is reliable enough to commit against.
Make it a routine, not a project
Optimization done as a quarterly crisis produces a sawtooth: spend drops, then climbs back as new workloads land without the same scrutiny. The durable version is a short recurring review — thirty minutes a month per team — plus guardrails that make the expensive default harder to reach than the reasonable one.
- Budget alerts routed to the owning team's channel, not only to finance.
- Anomaly detection on daily spend with a threshold tuned to avoid alert fatigue.
- Cost implications noted in architecture decision records, so the trade-off is recorded when it is cheapest to change.
- Retention and lifecycle policies set in code at resource creation, so storage growth is bounded by default.
None of this produces a headline percentage saving that can be promised in advance — the number depends entirely on the estate. What it reliably produces is the ability to answer, at any point, what a system costs and whether that is going up or down for a reason.