This is the last chapter of Part 5, and it answers the question the other eight generate: how often should any of this actually be looked at?
The usual answers are borrowed from elsewhere — a weekly ops review because that is when the meeting already exists, a monthly finance close because that is the accounting calendar. Neither is derived from anything about the data. There is a better first principle available, and the vendors published it.
Cadence is bounded by data latency
You cannot review faster than the numbers refresh. From the documentation gathered in cost in CI and showback and chargeback:
| Constraint | Documented behaviour | What it forbids |
|---|---|---|
| AWS Budgets refresh | Up to three times a day; updates typically 8–12 hours apart | A meaningful intra-day cost review |
| GCP budget notification | May take several hours for the first alert, plus a separate delay between usage and Cloud Billing | Treating a quiet dashboard as evidence of a quiet hour |
| Azure new subscription | Up to 48 hours before all Cost Management features are usable | A first-week readout on new estate |
| Azure tag visibility | A tag applied less than 24 hours ago may not have surfaced | Same-day reconciliation against the console |
| Azure tag retroactivity | Tags "aren't applied to historical data" | Any backdated allocation, at any cadence |
Read together these give a rule that does not depend on anyone's management philosophy:
A daily cost review is below the resolution of the data on at least one major cloud, and a same-day reconciliation is documented as unreliable on another. Weekly is the shortest cadence at which a full-fleet number is stable enough to argue about.
That single conclusion removes the most common failure in cost programmes — the daily dashboard that generates a churn of explanations for movements that are reporting artefacts. The exception is the runtime ceiling, which is not a review at all: it acts inside the blind window described in incident economics, and nothing on this page substitutes for it.
What belongs at each interval
The intervals below follow from the latency table; what to put in each one is judgment, not measurement.
Weekly — movement. The shortest honest cadence. Look at direction, not level: what moved, who owns it, is it expected. This is where a launch shows up, where a retry storm's tail shows up, and where an orphaned resource first becomes visible as a floor that did not drop back. Fifteen minutes, one owner per line, no slides.
Monthly — unit economics and attribution. A full billing period is the first point at which the unit economics denominator is stable and the attribution coverage from attribution at scale is meaningful. Two questions: is cost per unit moving in the right direction, and what share of spend is still unattributed. The second number matters more than most teams expect, because everything else is only as trustworthy as it is.
Quarterly — structure. The decisions with lead times: commitment coverage against the break-even thresholds from capacity planning, instance-family choices, whether the architecture still fits the workload, vendor terms ahead of renewal. Nothing here can be actioned inside a week, which is exactly why it does not belong in the weekly.
The separation is the point. A weekly meeting that keeps discussing commitment strategy has stopped being a weekly meeting, and the structural question will still not be resolved, because the people who can decide it are not in the room.
What to stop looking at
Cost programmes accumulate reports the way codebases accumulate flags. Worth deleting, on the reasoning of this part:
- Any dashboard refreshed faster than the billing pipeline. It is showing you the same number with more chances to over-read it.
- Absolute spend with no denominator. Rising cost on faster-rising usage is a business succeeding. Without the denominator the two are indistinguishable — the whole argument of unit economics.
- Reports with no named owner. Showback and chargeback is explicit: a line nobody owns produces no action at any cadence.
- Backdated allocation. Azure documents that tags are not applied to historical data. A report that appears to allocate a period before tagging was complete is modelling, and must say so.
Why this decays, and the one habit that holds it
Making it stick opened this part with the observation that savings decay. The cadence is the mechanism by which they do not — and cadences decay in a specific, predictable way: the weekly survives, the monthly slips, the quarterly is skipped once and never rescheduled. The structural review is the first casualty because it is the least urgent, and it is also the only one where the large decisions live.
The durable habit is therefore narrow and worth stating on its own: the quarterly review is scheduled a year ahead, with the commitment expiry dates on it. Not because the meeting is sacred, but because a commitment that lapses unexamined silently reverts to on-demand — and at the rates in what an H100 hour costs, that reversion is the single most expensive thing a forgotten calendar entry can do.
That is the end of Part 5, and the honest summary of it is short. Measure in units the business recognises, attribute it to someone who can change it, put the ceilings where they act faster than the billing pipeline reports, and review it no faster than the data actually arrives.