FinOps is, in the FinOps Foundation's words, "an operational framework and cultural practice" that maximizes the business value of technology through collaboration between engineering, finance and business teams. It works through three phases, Inform, Optimize and Operate, and it is more than a way to save money.
This guide covers the 2026 framework, a tagging policy you can enforce in code, shared costs, FOCUS billing data, Kubernetes allocation, savings levers and a 90-day plan. Tags belong in code, so it pairs with our guide to infrastructure as code. Cloud vendor details are as documented in October 2026.
What is FinOps, and what is it not?
The FinOps Foundation's definition, marked updated in March 2026 and read in October 2026, is: "FinOps is an operational framework and cultural practice which maximizes the business value of technology, enables timely data-driven decision making, and creates financial accountability through collaboration between engineering, finance, and business teams." The name is a portmanteau of "Finance" and "DevOps".
Drop two misreadings early:
- It is not a cost-cutting project. "If it seems that FinOps is about saving money, think again," the Foundation writes. Sometimes a business tightens the belt and sometimes it invests more; the goal is to identify, communicate and optimize the value behind technology decisions.
- It reaches beyond the public cloud. The Foundation extends the practice to SaaS, licensing, data centers and data cloud platforms, but this guide covers public cloud and Kubernetes.
If your bill jumped after a migration, our cloud migration strategy guide covers cost and governance after cutover.
How is the FinOps Framework organized?
The Framework is "flexible and non-prescriptive": you take the parts you need. Its six principles, listed in no particular order, cover collaboration, business value, ownership of usage, accessible and timely data, central enablement and the variable cost model of the cloud.
The phases are the working loop, run iteratively. In Inform, you examine cost, usage and efficiency data. In Optimize, you document options, including usage optimization, "using fewer resources to achieve an acceptable outcome", and rate optimization, "paying an appropriate amount for the resources we must use". In Operate, you implement the changes and loop back.
The Capabilities sit in four domains: Understand Usage & Cost, Quantify Business Value, Optimize Usage & Cost and Manage the FinOps Practice. The Framework names six core personas: FinOps Practitioner, Engineering, Finance, Product, Procurement and Leadership. A FinOps Scope is "a defined segment of spending across technology categories", aligned to business constructs such as a product, cost center or environment.
Maturity is Crawl, Walk or Run. The maturity model lists sample goals from the FinOps community:
| Maturity | Cost allocated | Forecast vs actual variance |
|---|---|---|
| Crawl | At least 70% | Below 20% |
| Walk | At least 85% | Less than 10% |
| Run | More than 90% | Less than 5% |
The same page warns that "an organization's goal should never be simply to achieve a 'Run' maturity in every Capability".
The 2026 revision, published on March 19, 2026, updates the definition and adds Executive Strategy Alignment as a Capability.
How do you build a tagging policy that holds?
Allocation comes first, because anomaly alerts, budgets, unit costs and commitments all assume you can say whose spend it is. The Foundation says allocation needs an allocation strategy, a tagging strategy and a shared cost strategy, and that tagging works better with automation, such as infrastructure as code or tools that manage tags after launch. Count coverage in money, because the maturity goals above are stated in cost.

Start with four keys. This table is an example policy, not a standard:
| Key | Values (example) | The question it answers |
|---|---|---|
costcenter | Codes from your finance chart, such as cc-1001 | Which budget pays for this? |
environment | production, staging, development, test | Can it be switched off outside working hours? |
owner | A team alias, not a person | Who is told when it grows, and who books the saving? |
application | The service name in your catalog | Which product does it belong to, for unit costs? |
Use lowercase for keys and values. AWS tag keys and values are case sensitive, so CostCenter and costcenter are different tags. Google Cloud label keys and values can contain only lowercase letters, numeric characters, underscores and dashes (international characters are allowed), so one lowercase convention works on both.
Enforce at creation, not in cleanup. The Foundation calls this shifting left: ownership and metadata standards are "enforced at the point a resource is created, not reconstructed later through tagging cleanup." These controls combine:
| Control | What it does, and its limit |
|---|---|
Code (AWS): Terraform provider with tag_policy_compliance = "error" | Checks required tags before a resource is created and before its tags change; non-tag updates are always permitted. Needs provider 6.22.0 or later |
Pipeline (AWS): CloudFormation hook AWS::TagPolicies::TaggingComplianceValidator | Warn mode lets deployments proceed; Fail mode blocks those missing required tags. Activate it in every account and Region you want covered |
Guardrail (AWS): a service control policy denying ec2:RunInstances when aws:RequestTag/costcenter is missing | Blocks untagged EC2 launches. AWS described this pattern in 2021, when a tag policy enforced a tag's accepted value, not its presence |
| Azure: Azure Policy "Require a tag on resources" (deny) and "Inherit a tag from the resource group if missing" (modify) | The first does not apply to resource groups; the second leaves an existing tag with a different value unchanged |
Warning
A tag on a resource is not yet a tag in your cost data. On AWS, a tag key reaches billing reports only after you activate it as a cost allocation tag, which can take up to 24 hours to appear for activation and up to 24 more to activate. Only the management account (or a standalone account) can open the cost allocation tags manager. On Azure, tag inheritance copies subscription and resource group tags onto usage records, "and not the resources themselves", within 24 hours for the current month.
On AWS, the organization-level piece is a tag policy. This one follows the AWS example for required tag keys, in lowercase, and defines one of the four keys; the others follow the same pattern.
{
"tags": {
"environment": {
"tag_key": { "@@assign": "environment" },
"tag_value": {
"@@assign": ["production", "staging", "development", "test"]
},
"report_required_tag_for": {
"@@assign": [
"ec2:ALL_SUPPORTED",
"rds:ALL_SUPPORTED",
"s3:ALL_SUPPORTED"
]
}
}
}
}
In Terraform, HashiCorp's guide says the tag policy must be created in the organization's management account and the principal that runs Terraform needs the ListRequiredTags permission. In warning mode, resources built on Terraform Plugin SDK V2, a majority of the provider's resources, cannot emit warning diagnostics for violations and log a WARN message instead, shown when you set TF_LOG=warn.
How do shared costs, showback and chargeback work?
Some spend has no single owner: a shared cluster, networking, support plans. The Foundation says many early practices adopt an "informed ignore" approach, a business decision to budget shared costs centrally. AWS Cost Categories offer split charge rules with Proportional, Fixed and Even split methods, though you see the split only on the cost category's details page in the console. Azure cost allocation does not support purchases such as reservations and savings plans and does not affect your invoice; "all chargeback processes happen in your organization outside of Azure".
The Foundation separates showback from chargeback by formality: chargeback sends expenses to official accounting budgets. Its position: "Showback is always required in any FinOps practice, but chargeback is dependent on organizational accounting policies." Invoices arrive "between 3 and 12 days after the end of the previous month in most cases", so showback can run as usage data arrives, and chargeback can follow the invoice.
What is FOCUS, and which clouds export it?
The FinOps Open Cost and Usage Specification (FOCUS) is "an open specification that defines clear requirements for technology billing data generators to produce consistent datasets". Without it, you normalize each vendor's terminology and metrics before analysis.

As of October 2026 the latest version is FOCUS 1.4, ratified on June 4, 2026, with new Invoice Detail and Billing Period datasets and, the FOCUS site says, zero incompatible changes. Providers lag the specification, so check each export's version before combining them:
| Provider | Export status, October 2026 |
|---|---|
| AWS | Data Exports offers FOCUS 1.0 or 1.2. Version 1.2 has been generally available since November 19, 2025, is recommended for all new implementations and is not backwards compatible with 1.0 |
| Microsoft Azure | The schema index lists "Cost and usage details (FOCUS)" for EA and MCA at 1.2-preview |
| Google Cloud | The BigQuery FOCUS export is a Preview, available "as is" with possibly limited support |
Start with these columns (FOCUS 1.4 names):
| Column | What it holds | Use it for |
|---|---|---|
BilledCost | The cost of a charge as invoiced by the invoice issuer in a billing period | Invoice reconciliation, cash-based forecasting, budgeting |
EffectiveCost | The cost recognized in a charge period, including amortized portions of related purchases such as commitments | Accrual-based reporting, forecasting, cost allocation |
Tags | A JSON object of user-defined and provider-defined tags; may be null, and a key may carry a null value | Your allocation keys |
The FOCUS tag coverage use case gives a query and a matching KPI, the percentage of cloud costs that are tagging policy compliant, which "requires an established organizational tagging policy". This PostgreSQL version, which assumes unquoted column names and Tags stored as jsonb, shows where the untagged dollars sit, by service:
SELECT
ServiceName,
SUM(EffectiveCost) AS total_cost,
COALESCE(SUM(EffectiveCost) FILTER (
WHERE NULLIF(Tags ->> 'costcenter', '') IS NULL), 0) AS untagged_cost
FROM focus_data
WHERE ChargePeriodStart >= '2026-09-01'
AND ChargePeriodEnd <= '2026-10-01'
GROUP BY ServiceName
ORDER BY untagged_cost DESC;
ChargePeriodEnd is an exclusive bound, as Google's FOCUS export reference spells out, so <= keeps September 30. The query counts a missing key, a null value and an empty string as untagged: PostgreSQL's ->> operator returns NULL rather than failing when a key does not exist, and NULLIF turns an empty string into NULL. Google's export has no Tags column: the same information arrives in x_Tags, x_Labels, x_SystemLabels and x_ProjectLabels, so the predicate changes.
How do you allocate Kubernetes costs?
Kubernetes places Pods by their requests: the scheduler uses them to pick a node, and the kubelet reserves at least the requested amount. A generous request therefore reserves capacity whether or not the workload uses it, and someone pays for the gap.

Tools price that gap by different rules, so check the rule before you compare numbers. GKE cost allocation data "is based on resource requests, not resources consumed". AWS split cost allocation data for EKS lets you choose resource requests or, with Amazon Managed Service for Prometheus, "the higher of Kubernetes pod CPU and memory resource requests and actual utilization". The OpenCost specification says workload cost should be understood as "max(request, usage)" for resources with allocation costs, such as CPU and GPU. A Pod that requests 2 vCPU and uses 0.5 carries the cost of 2 under each rule. A Pod that requests 0.5 and bursts to 2 carries the cost of 0.5 under a requests-only rule and of 2 under a higher-of rule.
OpenCost is a CNCF project: accepted on June 17, 2022 and Incubating since October 25, 2024. Its specification defines cluster idle cost as the portion of resource allocation cost not allocated to any workload. Choose one allocation rule, write it down and apply it to every team. Subscription platforms add a licence line to the same problem; our OpenShift vs Kubernetes guide covers what the OpenShift subscription adds.
Which optimization levers save money, and what do they risk?
Usage and rate optimization interact. The Foundation's rate optimization guidance warns that a commitment recommendation can conflict with one to change a resource's type or size, so teams may double count the benefit or pay for a year for resources that no longer exist in their original form. Waiting for fully optimized usage "will almost always lead to paying more than a balanced approach", so keep one backlog and rank it by value.
Rightsizing needs care. The Foundation's usage optimization guidance says to examine "longer-cycle periods of high utilization (e.g. higher utilization at month-end or during quarterly busy periods)". AWS Compute Optimizer analyzes metrics from a period over the last 14 days, or up to 93 days if you activate enhanced infrastructure metrics, so by our reading a month-end peak can fall outside the shorter window.

| Lever | How it saves | What can go wrong |
|---|---|---|
| Rightsizing | Fits a resource to measured use | A short look-back window can miss peaks, as above, and resizing can disrupt service |
| Scheduling non-production | Stops environments nobody is using | The Foundation says pre-production resources "should always require scheduled start/stop at launch when feasible" |
| Storage tiers and lifecycle | Moves or expires data nobody reads | Colder tiers have minimum durations: Azure cool 30 days, cold 90, archive 180, and Google Cloud Nearline 30, Coldline 90, Archive 365. S3 skips objects under 128 KB by default and charges a request per transition |
| Commitments | A lower rate in return for committed usage | Paid whether or not a matching resource runs; see the products below |
| Spot and preemptible | Uses spare capacity at a lower price | Capacity can be taken back: AWS warns two minutes before a stop or termination, Azure offers best-effort notice up to 30 seconds ahead and no SLA, and Google Cloud can preempt "at any time" |
| Data transfer | Removes avoidable per-GB charges | AWS charges for cross-AZ and cross-Region transfer and per GB through a NAT gateway. Same-AZ transfer is free, so keeping chatty services in one AZ saves money but gives up multi-AZ availability |
| Architecture | Changes the design: managed services, caching | "Value and cost efficiency is best achieved by architecting it into a system's design", per the Foundation. Infracost shows estimates before changes are deployed |
Table formats keep old data until you remove it: Iceberg snapshots accumulate until they are expired, as our Apache Iceberg guide shows. Serverless can cost more than a container at a steady load, which our serverless computing guide works through. Resilience is a cost decision too: our RTO vs RPO guide matches standby capacity to each system's recovery targets.
A commitment discount is not a resource. The Foundation compares it to a coupon: "If no resource is running which matches the type of the commitment, the organization still pays for the resource." The products differ in how much you can undo:
- AWS Savings Plans commit you to an amount of compute, measured per hour, for one or three years. A plan with an hourly commitment of $100 or less can be returned if it was bought in the last seven days and in the same calendar month, within your return limit.
- Azure reservations can be exchanged for another reservation of the same type, or refunded up to $50,000 USD in a 12-month rolling window. An Azure savings plan is a fixed dollar amount per hour, unused commitment for an hour expires, and purchases cannot be cancelled or refunded.
- Google Cloud committed use discounts are spend-based or resource-based commitments for one or three years.

Buy in stages. The Foundation's KPIs include commitments that return more than 10% on the investment, break even within 9 months, and are spread over the year so that some are always expiring.
How do unit economics, budgets and anomaly alerts fit in?
Unit economics ties technology cost to business outcomes. The Foundation's examples include cost per transaction, per customer, per case resolved, per GB stored and per token.
A worked example with made-up numbers: an orders service has an allocated cost, shared costs included, of $18,400 in March for 230,000 orders, or $0.08 per order. In April it costs $19,500 for 310,000 orders, or about $0.063 per order. Spend rose 6% while the cost per order fell 21%, which the total alone would hide.
AWS and Azure alerts run on delayed data. AWS Budgets is "updated up to three times a day", typically 8 to 12 hours after the previous update. Azure cost data is typically available within 8 to 24 hours and budgets are evaluated every 24 hours. An anomaly, in the Foundation's definition, is spending "different (usually higher) than normal historical or expected spend". AWS Cost Anomaly Detection runs approximately three times a day and can take up to 24 hours to detect an anomaly after usage occurs; a new service subscription needs 10 days of history first.

Warning
Budgets report; they stop nothing unless you set them up to. Azure says "Resources aren't affected, and your consumption isn't stopped", and Google says an alerts-only budget "doesn't automatically cap" usage or spending. AWS Budgets can run an action if you configure one, such as applying a custom IAM policy that denies new provisioning, and Google Cloud offers project-level spend cap budgets for supported services. Use a hard limit where one mistake could burn a month of budget.
Which FinOps mistakes should you avoid?
| Mistake | Why it hurts | Instead |
|---|---|---|
| Tag debt | Ownership is "reconstructed later through tagging cleanup" | Enforce required keys at creation and report coverage by cost |
| Optimizing before allocating | Savings have no owner, and rightsizing and commitment advice can double count | Reach the Crawl goal of 70% allocated, then keep one ranked backlog |
| Commitments bought too early or all at once | They are paid whether or not a matching resource runs, and Azure savings plans cannot be cancelled or refunded | Buy in stages against steady usage and watch utilization |
| Savings nobody books | A saving that never reaches the forecast or budget is absorbed by growth | Record expected savings per item, name an owner, update the forecast |
What does a 90-day FinOps plan look like?
This sequence is ours, not part of the Framework. Each step names an owner and a check.
- Days 1 to 15: owners and data. Name an owner and a cost center for every account, subscription and project. Turn on billing exports (FOCUS where offered), activate cost allocation tags on AWS and enable tag inheritance on Azure. Owner: FinOps lead. Check: last month's export loads into a warehouse and ties to the invoice.
- Days 16 to 30: policy and first showback. Publish the four-key tagging policy and run the coverage query. Owner: FinOps lead and engineering leads. Check: at least 70% of cost allocated to a known owner, the Foundation's sample Crawl goal.
- Days 31 to 45: switch off and expire. Schedule non-production environments and add storage lifecycle rules. Owner: engineering owners. Check: non-production resources start and stop on a schedule, and expired data is deleted.
- Days 46 to 60: backlog and commitments. Rank rightsizing items by savings against effort and risk, then buy a first commitment on steady usage. Owner: engineering, finance and procurement. Check: each item has an owner, an estimate and a date, and utilization is tracked daily.
- Days 61 to 75: enforce and alert. Turn on tag policy checks in code, then budgets and anomaly alerts for each owner. Owner: platform engineering. Check: a plan with a missing required key fails, and a test alert reaches the owner.
- Days 76 to 90: one unit cost and the first review. Pick a business measure such as orders or tenants and report cost per unit as a trend. Owner: product, finance and the FinOps lead. Check: the review happens, and booked savings appear in the forecast.
Computese's cloud transformation service puts tagging, budgets and guardrails in place from the first day and runs FinOps reviews that tie spend to value, with tools such as AWS Cost Explorer, Azure Cost Management, Kubecost, OpenCost and Infracost.


