Overview and motivation

In a data centre, spending is decided by purchase orders and arrives as a few large steps. In the cloud, a developer can commit the organisation to a recurring charge with a few lines of configuration, and a forgotten test environment can cost money for months. Cloud computing offers on-demand access to computing resources, and the pay-as-you-go model that comes with it means the bill follows usage. That is powerful and risky. The same flexibility that lets a team scale in minutes lets cost grow faster than anyone notices.

FinOps, a contraction of finance and operations, is a practice and a culture for managing cloud cost. The FinOps Foundation describes it as an operational framework in which engineering, finance, procurement, and business teams collaborate on data-driven spending decisions. Its central idea is that the people who use cloud resources should see and own their costs, while finance provides the structure, forecasts, and commitments. This matches the agile aims of continuous collaboration among stakeholders and of teams that own their services, and it matches the ITIL concern with service financial management. This topic covers the measures that make that practice concrete. It builds on the general costing ideas in the previous topic (see topic 7.5, Service cost and unit economics).

Cloud cost has features that need their own measures.

  • Elasticity. Autoscaling adjusts resources to demand, which makes cost variable and sometimes surprising. Serverless computing charges per execution, and shifts attention from servers to requests.
  • Many small decisions. Each choice of instance size, storage tier, region, and architecture changes cost. Spend is distributed across hundreds of decisions made by many people, and it is hard to see without tags and good reporting.
  • Pricing models. On-demand prices are flexible and highest. Commitments in exchange for a volume discount, such as reserved capacity or savings plans, lower unit price in return for a promise of use. Spot or preemptible capacity is cheap and can be reclaimed. Each model has a risk, and the right mix depends on how predictable the workload is. The exact terms differ between providers and change often, so check current provider documentation before relying on any detail.
  • Shared and untagged cost. Resources without an owner or purpose tag cannot be attributed. Shared platforms spread cost across teams that consume them differently.
  • Environmental cost. Cloud workloads consume energy and cause emissions, and the carbon footprint of a workload is increasingly a requirement to report. Methods for estimating it are still developing, and numbers should come with their method.

Cost measures in the cloud are easy to gain through shortcuts. Cutting spending by turning off monitoring or by shrinking resources until the service falls over shows a lower bill and a worse service. The guardrail from Goodhart's law applies: every cost measure here sits beside a reliability or quality measure, and no team is ranked by its bill (see topic 1.2, Goodhart's law and the psychology of metrics).

Key principles

  • Cost is an engineering metric. Teams that build and run a service see and own its cost.
  • Allocate to owners. Every dollar should map to a team or a declared shared pool.
  • Pair cost with quality and reliability. Savings that break the service are not savings.
  • Optimise continuously, in small steps. Review waste regularly and act as part of normal work.
  • Forecast with ranges. Variable demand makes single-number forecasts fragile.
  • Collaborate across functions. Engineering, finance, procurement, and the business decide together.
  • Do not rank teams by their bill. Services differ, and growth is not waste.

Metric definitions

Metric Definition Formula Unit Data source ITIL practice Agile counterpart Cadence
Allocation coverage Share of cloud spend attributed to an owning team or a declared shared pool spend with a valid owner and service tag, or in a declared shared pool / total cloud spend percent Cloud billing export, tag policy Service financial management; IT asset management Team ownership of what it runs Weekly
Cloud unit cost Cost of cloud resources per unit of something users value cloud cost for the service in the period / units delivered in the period currency per unit Billing data joined to service volume Service financial management; service level management Cost to serve per outcome Monthly
Waste rate Share of spend on resources that are idle or clearly oversized spend on idle or oversized resources, per agreed rules / total spend on the scoped resources percent Utilisation metrics and billing Capacity and performance management; financial management Waste reduction in the toolchain Weekly
Commitment coverage Share of eligible steady usage covered by discounted commitments eligible usage covered by commitments / eligible usage percent Billing and commitment reports Service financial management; supplier management Planning for predictable demand Monthly
Commitment utilisation Share of purchased commitments that were used commitment hours used / commitment hours purchased percent Commitment reports Service financial management; supplier management Value from planned investment Monthly
Forecast variance Difference between forecast and actual cloud spend, by team and in total (actual minus forecast) / forecast percent Forecast log and billing data Service financial management Forecast accuracy Monthly
Anomaly acknowledgement time Time from a cost anomaly alert to a recorded decision by the owner median and 85th percentile of hours from alert to owner response hours Anomaly alerts and ticket records Monitoring and event management; incident management Responsiveness to a production signal Monthly

Define the scope of resources, the tag rules, and what counts as idle or oversized, and record them. Commitment coverage is computed against steady baseline usage only, not peaks. Unit cost uses a unit agreed with the service owner and fixed for the trend. Cost data should say whether it includes credits, discounts, taxes, and support charges, and which currency and exchange rate apply.

Recommendations

Make cost visible and owned

1. Enforce a small, mandatory tag set. Require every resource to carry an owner, a service, and an environment tag, enforced by policy at creation, not by a clean-up afterwards. Allocation coverage measures how well this works. Resources that fail to tag can be blocked, or flagged within a day to the account owner. Keep the tag set small so that people comply.

2. Give teams their own cost view. Provide each team with a daily or weekly view of its spending by service and environment, with trends and anomalies. People respond to data they can see and understand. A report that arrives monthly and only goes to finance changes little.

3. Define shared cost pools and their drivers. Shared platforms, networking, and support charges need a rule for how they are spread (see topic 7.5, Service cost and unit economics). Publish the rule, review it with the teams, and keep the pool small. A rule that people can follow earns more trust than a fair one that nobody understands.

4. Hold a short regular cost conversation. Bring engineering, finance, and product owners together at a steady cadence, weekly or fortnightly for large spenders and monthly for others, for thirty minutes. Review the trend, the anomalies, and the optimisation backlog. This is continuous collaboration in practice, and it keeps cost on the agenda without a heavy process.

Measure value, not just spend

5. Express cost per unit. Total cloud spend rises with success. What matters is cost per active customer, per transaction, or per build. Choose a unit that is meaningful for the service, track it over time, and annotate changes. A falling unit cost with rising total spend is a healthy pattern, and a rising unit cost with flat usage is a prompt to investigate.

6. Show reliability and quality beside cost. Display cost with availability, latency, and error rate for the same service (see topic 5.2, SLIs, SLOs, and error budgets). Cost savings that raise error rates are a transfer of cost to users. An error budget approach can make the trade-off explicit: spend less when the service is far better than needed, and invest when it is not.

7. Track waste with explicit rules. Define idle (for example, no meaningful use over a period) and oversized (for example, utilisation consistently far below the capacity provisioned), and report the spend on each. Rules need care: a standby system for resilience looks idle by design, and should be excluded and labelled. Revisit the rules when they produce false positives.

Act on what you find

8. Right-size and clean up in normal work. Make waste removal a standing item in the team's backlog, tackled in small steps as part of continuous improvement. Schedule non-production environments to stop outside working hours, delete unattached storage, choose appropriate storage tiers, and review the largest line items first. Small regular changes outperform rare heroic clean-ups, which tend to be forgotten when pressure rises.

9. Match pricing model to workload. Use commitments for steady baseline usage, on-demand for variable load, and interruptible capacity for work that tolerates interruption. Commit incrementally, in small amounts, and track both coverage and utilisation. Over-commitment wastes money and under-commitment pays a premium. Involve finance and procurement, and re-check the terms with the provider since they change.

10. Detect anomalies and respond. Use anomaly detection on spend, which flags sudden changes relative to a baseline. Anomaly detection tools range from simple thresholds to statistical models. Alert the owning team, and measure how long it takes to acknowledge and decide. A misconfigured job can run up a large bill in hours, so speed matters (see topic 5.5, Monitoring and alert quality metrics).

Forecast and plan

11. Forecast with ranges and drivers. Build forecasts from the drivers of demand, such as expected users or transactions, together with unit cost. Use variance analysis to explain differences between forecast and actual, and separate those due to volume, price, and mix. Give a range and state the assumptions. Forecast variance should shrink as the model improves.

12. Include sustainability in the picture. Estimate the emissions associated with major workloads using a documented method, such as provider reports or published methodologies, and state the limits of the estimate. Compare choices of region, instance type, and schedule on both cost and emissions. Treat emissions per unit as a companion to cost per unit, and do not claim precision the method cannot support.

Gaming vectors and guardrails

Metric How it gets gamed Paired guardrail metric or practice
Allocation coverage Tagging everything to a catch-all owner; declaring large amounts as shared Share of spend owned by catch-alls; shared pool size and review; owner confirmation sample
Cloud unit cost Choosing a unit that grows cheaply; cutting reliability to cut cost Fixed unit definition; reliability and quality companions
Waste rate Narrowing the definition of idle; moving waste to an account out of scope Rules in the register; scan all accounts; sample resources the rules exclude
Commitment coverage Committing to inflated baselines to chase a high percentage Commitment utilisation; check the baseline against usage history
Commitment utilisation Adding eligible usage that is not needed to fill commitments Compare with growth in real demand; unit cost
Forecast variance Padding forecasts; revising them late Log the forecast at the start; track range width and hit rate
Anomaly acknowledgement time Acknowledging alerts without looking; closing alerts automatically Record a decision and outcome; sample alerts for quality of response

A healthy reading is nearly all spend owned and visible, unit cost that moves for explained reasons, waste that shrinks without outages, commitments that are used, and forecasts whose error narrows over time.

Trade-offs: pros and cons

Approach Strength Weakness
Central cloud team controls spend Consistency and negotiating power Slow; far from the engineering decisions
Teams own their cloud cost Fast, informed decisions Needs data, training, and guardrails
On-demand only Flexible Highest unit price
Heavy commitments Lower unit price Lock-in and waste if demand falls
Strict budget caps Prevents overruns Can cause outages or block valid work
Alerts and review Preserves speed Depends on timely response
Aggressive right-sizing Saves money Risk to performance if overdone
Measured right-sizing with SLOs Balanced Needs reliability data and care

The prose answer is to give teams data and ownership, set guardrails such as alerts and budgets instead of hard stops where service is critical, commit gradually to the steady baseline, and always pair cost with reliability.

Questions to discuss with your team

1. How much of our cloud bill can we attribute to an owner today, and who cannot be found? Look at the share of spend with valid tags and the largest untagged items. Ask who would own each. The facilitator should note how many items nobody claims. Agree a tag policy step to enforce at creation.

2. What did our cost per unit do over the last six months, and why? Plot unit cost and annotate changes in architecture, price, and volume. Ask which changes were deliberate. If the trend is a surprise, discuss what data is missing. Add a quality measure to the chart.

3. Which of our resources would we turn off tomorrow if we dared? List candidates: idle environments, oversized databases, forgotten storage. For each, ask what the risk is and who decides. Agree a safe way to test removing one, such as scheduling or snapshotting first.

4. How would we know within a day if a job started costing ten times more? Walk through the alerting path from the bill to the owner. Test it with a past example. The conversation should reveal delays and gaps in ownership. Agree a target time to acknowledge and decide, and who is on point.

5. How much of our baseline usage could we safely commit to, and what would make us wrong? Examine the usage history for a steady floor and the risks to it: product changes, migrations, price changes. Decide a first modest commitment and a review date. Include finance in the discussion.

6. What trade-off between cost and reliability are we making by accident? Review recent savings and incidents. Were any related? Discuss how to make the trade-off explicit, for example by linking cost reviews to error budgets. Decide who can approve a cut that affects reliability.

Four perspectives

Product and delivery teams

Delivery teams make the decisions that drive cost and are best placed to reduce it. They should tag at creation, check their cost view regularly, build cost into design discussions, and treat waste reduction as normal work. They should be able to see the effect of a change on cost soon after making it, which supports fast learning.

Service and operations teams

Operations and platform teams run shared services, enforce policy, and operate anomaly alerts. They should keep tag compliance simple, provide self-service reports, and publish clear rules for shared cost. They should watch both the cost and the reliability of the platforms they provide, and help teams act without becoming a gate.

Leadership and governance

Leaders and finance want predictable spend and good value. They should fund the tooling and the time for cost work, accept ranges in forecasts, avoid league tables of teams, and ask for unit cost beside quality. Governance should check that commitments are made with data, and that sustainability claims state their method.

Customers and users

Users feel cloud cost indirectly, through prices, features, or the stability of the service. They benefit when savings fund improvements and suffer when cuts reduce quality. Where users are charged for consumption, they want clear bills and the ability to influence them.

In practice: an agile ITIL example

A regional online retailer runs its shop and order platform on a public cloud, with six engineering teams. The numbers below are illustrative.

Cloud spend rose steadily over a year, and the finance director asks for a cut. The FinOps lead pulls the billing data. Allocation coverage is 68 percent, with one third of the untagged spend in test accounts. Cost per completed order is US$0.31, up from US$0.27 six months earlier, while order volume grew 15 percent. The waste rate under the team's rules is 14 percent of the scoped spend, mostly non-production environments running overnight and at weekends and a few oversized databases. Commitment coverage is 20 percent of the steady baseline, with 97 percent utilisation on what was bought. A forecast variance of plus 18 percent in the last quarter followed a marketing campaign that nobody had told the platform team about. A runaway data job in one account ran for two days before anyone noticed.

The lead agrees four changes with the engineering teams. A tag policy blocks new resources without owner and service tags, with a two-week grace period. Each team receives a weekly cost view and a short slot in the fortnightly cost conversation. Non-production environments are scheduled to stop outside working hours, and the three oversized databases are resized after checking their performance against the service level objectives. A first commitment covers 60 percent of the steady baseline, with a review in three months. An anomaly alert goes to the owning team with an expectation of a recorded decision within four hours, and the planning calendar now includes marketing campaigns.

After a quarter, allocation coverage is 93 percent, the waste rate falls to 6 percent, cost per order falls to US$0.26 with no change in checkout latency or error rate, and the forecast variance is plus 5 percent. Commitment coverage is 55 percent with 96 percent utilisation. The median anomaly acknowledgement time is 2 hours, with the 85th percentile at 6. The finance director asks for a league table showing cost per team. The FinOps lead declines, explaining that the teams run different services with different growth, and offers instead each team's unit cost trend with its reliability measures. The decision is recorded, and the next review date set.

Sector lens

Startup. Credits and low bills can hide bad habits that become expensive at scale. Tag from the first day, schedule non-production resources, and track cost per customer early. Commit only when demand is predictable.

Small business. A single account and a few services make tagging and review simple. Review the bill monthly with the person who builds the system, remove forgotten resources, and be wary of long commitments. A budget alert is the most valuable single control.

Enterprise. Many accounts, teams, and providers call for a central FinOps function that enables teams instead of controlling them. Standardise tags and billing exports across providers, negotiate commitments centrally, and give teams their data. Plan for multi-cloud reporting differences.

Government and public sector. Procurement frameworks, data residency, and fixed annual budgets constrain choices. Document cost and region decisions, model the effect of variable spending on appropriations, and keep audit trails. Public commitments on sustainability may require reported emissions with methods (see topic 8.5, Regulated and government reporting).

Business case: motivations, ROI, and TCO

The costs of a FinOps practice are the time of a few people to run reporting, tagging policy, and the regular conversation; tooling, which may be provided by the cloud platform or bought; and the engineering time to act on findings. A small practice can start with a billing export and a spreadsheet.

The returns come from reduced waste, better prices through commitments, fewer surprises, and better decisions about architecture. Build the case from your own numbers: the waste rate and its cost in the first review, the savings from scheduling non-production environments, the commitment discount on the portion committed, and the avoided cost of anomalies caught early. Avoid quoting industry-wide waste percentages or savings figures without a traceable source and method. Include the cost of the practice and the risk of over-commitment. The financial case for improvements in general is covered in the next topic (see topic 7.7, Return on investment and business case).

Anti-patterns and pitfalls

  • A central team that says no. It slows work and teams work around it. Enable teams with data and guardrails.
  • Cutting cost without watching reliability. The savings return as incidents. Pair cost with service level measures.
  • Chasing commitment coverage. A high percentage with low utilisation is waste. Track both.
  • Monthly reports to finance only. Engineers never see them. Give teams daily or weekly views.
  • Tagging as a clean-up project. It decays. Enforce tags at creation.
  • Ranking teams by spend. Growth looks like waste. Show unit cost trends with context.
  • Forecasting a single number. Variable demand breaks it. Use ranges and drivers.
  • Claiming precise emissions. Estimation methods are uncertain. State the method and limits.

Maturity model

Dimension 1 Initiate 2 Develop 3 Standardise 4 Manage 5 Orchestrate
Visibility and ownership Monthly bill reviewed by finance Tags used by some teams Mandatory tags; teams see weekly cost views Allocation coverage managed; shared pools reviewed Cost data flows to product decisions in near real time across providers
Optimisation Occasional clean-ups Schedules and right-sizing by some teams Waste rules and a standing backlog of optimisations Waste, commitments, and unit cost managed with reliability companions Continuous, automated optimisation with human oversight and shared learning
Planning Annual budget Rough forecasts Driver-based forecasts with ranges Variance analysed and fed back; commitments managed incrementally Joint forecasting with the business, suppliers, and sustainability goals

Checklist

  • A small mandatory tag set (owner, service, environment) is enforced at resource creation.
  • Allocation coverage is reported weekly, with the largest untagged items listed.
  • Each team has a weekly cost view with trends and anomalies.
  • A regular engineering and finance cost conversation is on the calendar.
  • Cloud unit cost is tracked beside availability, latency, and error rate.
  • Waste rules are written down and applied, with exclusions labelled.
  • Commitments are made incrementally against steady baseline usage, with coverage and utilisation tracked.
  • Forecasts use drivers and ranges, and variance is analysed.
  • Anomaly alerts reach owners with an expected time to decide.
  • Emissions estimates, if reported, state their method and limits.

AI prompts

Analyse a cost trend.

Here is a summary of our cloud cost and usage for [service] over [period]:
[paste monthly cost, volume, and any known changes]. Break the change in cost
into volume, price, and mix effects as far as the data allows, suggest what to
check next, and propose a unit cost with a quality companion. Do not invent
prices or discounts. List your assumptions and the data you would need to be
more certain.

Draft a tag policy.

Draft a tag policy for [type of organisation] using [cloud provider(s)]. Include
a minimal set of mandatory tags, allowed values, how they are enforced at
creation, how violations are reported to owners, and a grace period. Keep it
short enough for engineers to follow. List your assumptions and anything that
depends on provider features I should check in current documentation.

Define waste rules.

Propose rules for classifying cloud resources as idle or oversized in [type of
environment], including the metrics used, time windows, exclusions (for example
resilience standbys), and a review process for false positives. Describe how
to avoid harming reliability. Do not set numeric thresholds without saying how
we should calibrate them from our own data. List your assumptions.

Plan a first commitment.

Our steady baseline usage over the last [period] is [summary] and demand risks
are [list]. Outline a cautious plan for a first commitment: how to estimate
the baseline, how much to commit, how to review utilisation, and what would
make us stop. Do not assume specific provider discounts or terms. List your
assumptions and questions for finance and procurement.

AI skills

Key takeaways

  • Make cost visible and owned: enforce tags at creation and give each team its own weekly view.
  • Measure cost per unit beside reliability and quality, never alone.
  • Remove waste as part of normal work, guided by written rules.
  • Commit gradually to the steady baseline and track both coverage and utilisation.
  • Forecast with drivers and ranges, and watch for anomalies with a fast response.
  • Do not rank teams by spend, and state the method and limits of any emissions estimate.

References and further reading

  1. Storment, J. R., and Fuller, M., Cloud FinOps, 2nd ed., O'Reilly, 2023.
  2. FinOps Foundation, FinOps Framework, https://www.finops.org/framework/.
  3. Beyer, B., Jones, C., Petoff, J., and Murphy, N. R. (eds.), Site Reliability Engineering: How Google Runs Production Systems, O'Reilly, 2016.
  4. Forsgren, N., Humble, J., and Kim, G., Accelerate: The Science of Lean Software and DevOps, IT Revolution, 2018.
  5. Beck, K., et al., Principles behind the Agile Manifesto, 2001, https://agilemanifesto.org/principles.html.
  6. Wikipedia, Cloud computing.
  7. Wikipedia, Pay as you go.
  8. Wikipedia, Autoscaling.
  9. Wikipedia, Serverless computing.
  10. Wikipedia, Volume discount.
  11. Wikipedia, Carbon footprint.
  12. Wikipedia, Goodhart's law.
  13. Wikipedia, Anomaly detection.
  14. Wikipedia, Variance analysis.