FinOps Budgeting
A cloud budget is a forecast with owners attached. Get the data right, explain the last change before predicting the next one, and make every number traceable to a driver someone controls.
This skill follows the FinOps Foundation Framework's three phases. They are not a sequence you finish; mature teams cycle through them continuously. Use them to check that an answer is complete: a budget without Inform is a guess, without Optimize it bakes in waste, and without Operate nobody notices when it breaks.
| Phase | Question it answers | What you produce here |
|---|---|---|
| Inform | Where is the money going, who owns it, and why did it change? | Clean cost data, allocation, variance bridge, unit costs |
| Optimize | What could we pay less for, at what risk? | Ranked levers with savings estimates |
| Operate | How do we keep it on track? | Forecast worksheet, budgets and alerts, cadence, showback |
Workload design trade-offs (is this the right service at all?) belong to
cloud-architecture-best-practices. Implementing budgets, labels or policies in Terraform belongs
to devops-engineer; describe the configuration, let it write the code.
Inform
Get the data before the opinion
Never set a number from the invoice total alone. Start from line-item billing data:
| Cloud | Source | Notes |
|---|---|---|
| GCP | Cloud Billing export to BigQuery — detailed usage cost (resource-level) plus pricing export | Not retroactive: enable it now if it is off. Net cost = cost + sum of credits.amount. GKE cost allocation adds namespace/label breakdown |
| AWS | Data Exports — CUR 2.0 or the FOCUS export, to S3, queried with Athena | Use amortized cost for commitments; split by linked account and cost-allocation tags (activate them first) |
| Azure | Cost Management exports (FOCUS format available) to a storage account | Scope at billing account or management group; amortized view for reservations |
Prefer the FOCUS schema (FinOps Open Cost and Usage Specification) when the provider offers it
or when more than one cloud is involved, so the same queries work everywhere. The columns that
matter most for budgeting: BilledCost (what the invoice says), EffectiveCost (amortized, after
discounts — use it for trends and unit costs), ListCost (for savings math), ChargeCategory
(usage, purchase, tax, credit, adjustment), ServiceName/ServiceCategory, SubAccountId
(project/account/subscription), ResourceId, Tags, PricingCategory (on-demand, commitment,
dynamic), CommitmentDiscountId, ChargePeriodStart.
Starter query on the GCP detailed export (adapt the dataset, months and label key):
SELECT
invoice.month AS month,
service.description AS service,
project.id AS project,
(SELECT value FROM UNNEST(labels) WHERE key = 'team') AS team,
ROUND(SUM(cost) + SUM(IFNULL((SELECT SUM(c.amount) FROM UNNEST(credits) c), 0)), 2) AS net_cost,
SUM(usage.amount_in_pricing_units) AS usage_units -- meaningful per SKU only
FROM `billing_ds.gcp_billing_export_resource_v1_XXXXXX`
WHERE invoice.month BETWEEN '202607' AND '202609'
GROUP BY month, service, project, team
ORDER BY month, net_cost DESC;
Repeat grouped by sku.description for the top services; the cause usually lives at SKU level
(e.g. "BigQuery Analysis" vs. "Active Logical Storage", "Spot" vs. "Core" compute).
Allocate
Report the share of spend with a complete owner (team, service, env, cost-center labels
or an account/project that maps to one). Below ~80% allocated, fixing allocation is the first
action, because no one manages cost they are not shown. Shared costs (networking hubs, shared
clusters, support plans) get an explicit rule: proportional to direct spend, proportional to a
usage measure (CPU-hours, requests), or a fixed split. Write the rule down; showback disputes are
almost always about undocumented shared-cost rules.
Explain the change: the variance bridge
For each service (and each project/team inside the top services), split the change between two periods into:
| Bucket | How to compute | What it means |
|---|---|---|
| Volume | (usage now − usage before) × old unit rate | More traffic, data or environments |
| Rate | (unit rate now − unit rate before) × usage now | Price changes, lost or new discounts, commitment expiry |
| New / removed | Cost of SKUs or resources absent in one period | New services, new projects, decommissions |
| One-off | Credits ending, refunds, purchases, back-charges | Not run-rate; exclude from the baseline |
The bridge tells you which part of the increase is growth the business wanted, which part is waste, and which part is accounting noise. Present it as a table from the old total to the new total before proposing any budget.
Unit economics
Pick one or two unit metrics the business already tracks: cost per active customer, per 1,000 API requests, per order, per TB processed. Plot unit cost alongside total cost. Total cost rising while unit cost is flat or falling is healthy growth; unit cost rising is an efficiency problem. Budget growth should be expressed as "expected units × target unit cost", which is the language finance and product can both argue with.
Optimize
Separate the two kinds of lever, because they have different owners and risks:
- Usage levers (use less): rightsize, schedule, delete idle, change storage class, fix architecture. Owned by engineering; savings persist; need testing.
- Rate levers (pay less per unit): commitments, Spot/preemptible, edition/tier choice, negotiated discounts. Owned by FinOps/finance; savings depend on usage staying put.
Do usage levers on the stable part of the estate before buying commitments, or you commit to waste. For each lever, record: estimated monthly saving and how it was computed, confidence (high/medium/low), effort, risk to reliability, owner, and reversibility. Rank by saving × confidence ÷ effort.
GCP levers worth checking first
| Service | Lever | Evidence to look at |
|---|---|---|
| GKE | Rightsize requests from VPA recommendations or observed usage; HPA on real load; cluster autoscaler with the optimize-utilization profile; node auto-provisioning |
Requested vs. used CPU/memory per namespace (GKE cost allocation, Cloud Monitoring) |
| GKE | Spot node pools for stateless, fault-tolerant and batch workloads; Autopilot for spiky low-utilization clusters | Workload restart tolerance, PodDisruptionBudgets |
| GKE / Compute | Committed use discounts: resource-based (vCPU/memory, 1 or 3 years) or flexible spend-based CUDs, sized to the stable floor | Minimum hourly usage over the last 30–90 days, after rightsizing |
| BigQuery | Partitioning and clustering; require_partition_filter; maximum_bytes_billed on jobs; stop SELECT * on wide tables; materialized views for repeated aggregates |
INFORMATION_SCHEMA.JOBS: total_bytes_billed, top users and queries |
| BigQuery | On-demand vs. capacity pricing (editions with autoscaling reservations, baseline slots, optional slot commitments) | Compare monthly on-demand spend to total_slot_ms converted to slot-hours at edition rates |
| BigQuery | Custom quotas — query usage per day per project and per user — to cap runaway on-demand scans | Largest daily scan totals per user/service account |
| Cloud Storage | Lifecycle rules to Nearline/Coldline/Archive or delete; Autoclass for unknown access patterns; expire noncurrent versions | Access frequency; mind minimum storage durations (30/90/365 days) and retrieval fees |
| Networking | Review internet egress, inter-region and inter-zone traffic, Cloud NAT processing; co-locate chatty services; CDN for static content | Billing SKUs for egress; VPC Flow Logs / Network Intelligence Center top talkers |
AWS equivalents: Compute Optimizer, Savings Plans (Compute vs. EC2 Instance), Reserved Instances for RDS/ElastiCache, Spot, S3 Intelligent-Tiering and lifecycle, NAT Gateway and inter-AZ data transfer review. Azure: Advisor, savings plans and reservations, Spot VMs, Azure Hybrid Benefit, Blob access tiers and lifecycle management.
Commitment sizing rule
Commit to the floor, not the average: take hourly eligible usage after planned rightsizing, find a low percentile (for example the 10th) over the last 30–90 days, and cover that. Ladder purchases (several smaller commitments over months) rather than one big one, so coverage tracks the estate. Track coverage (share of eligible usage under commitment) and utilization (share of the commitment actually used); low utilization means you bought too much. Commitments are financial contracts: recommend them, never purchase them.
Operate
Forecast worksheet
Build the next-period budget from drivers, one worksheet per top-level owner (project, team or business unit), then sum. Each line has a named source so the number can be challenged line by line.
| Line | What goes in | Source |
|---|---|---|
| A. Run-rate | Last 30 days of effective cost, excluding one-offs, × days in the period ÷ 30 | Billing export (bridge "one-off" bucket removed) |
| B. Growth | A × expected change in the unit metric (users, requests, data volume) | Product/business forecast; recent unit-metric trend |
| C. Known events | Launches, migrations, new environments, decommissions, each with a date and a monthly cost | Roadmap, engineering leads |
| D. Planned savings | − Σ (lever saving × confidence factor × months active in the period) | Optimize list; use 1.0 / 0.6 / 0.3 for high / medium / low confidence |
| E. Rate changes | Commitment expiries, announced price changes, currency effects | Commitment inventory, provider notices |
| F. Forecast | A + B + C + D + E | — |
| G. Contingency | F × buffer: past forecast error if known, otherwise 5–10% | Previous budgets vs. actuals |
| Budget | F + G, phased by month (growth and savings ramp, they do not start on day 1) | — |
Present it as a range, not a point: a low case (savings land fully, growth at the bottom of the range) and a high case (savings slip, growth at the top). Finance gets one budget number plus the range and the assumptions that move it. Record actuals against each line every month; the lines that keep missing tell you which assumption to fix.
Guardrails
Budgets and alerts are the mechanism that prevents the next surprise. Set them per owner, not only on the billing account, so the alert reaches someone who can act.
GCP
- Cloud Billing budgets scoped per project (or per label/service where useful), with thresholds on actual spend at 50%, 90% and 100% and on forecasted spend at 100%.
- Programmatic notifications to a Pub/Sub topic, consumed by a function or chat integration that posts to the owning team's channel and opens a ticket. Email to billing admins alone gets ignored. Automatic billing disablement exists but stops every resource in the project; reserve it for sandboxes.
- BigQuery custom quotas per project and per user, and
maximum_bytes_billeddefaults in shared tooling; reservations with a max-slots cap when on capacity pricing. - Organization policy and IaC review that require the cost labels; a weekly report of unlabelled spend.
AWS: AWS Budgets per account/tag with actual and forecasted thresholds, SNS/Chatbot notifications, budget actions (apply an SCP or IAM policy) for sandboxes; Cost Anomaly Detection monitors per service and per cost-allocation tag; tag policies. Azure: budgets per subscription/resource group with action groups; anomaly alerts in Cost Management; Azure Policy requiring tags.
Anomaly detection complements budgets: budgets catch slow drift against a plan, anomaly detection catches a sudden spike days before a monthly threshold would.
Showback, chargeback and cadence
Start with showback (each team sees its allocated cost and unit cost) and move to chargeback only once allocation is above ~90% and shared-cost rules are agreed; charging teams for numbers they cannot explain destroys trust in the data. Keep a light rhythm: anomalies and alert triage weekly, budget vs. actual by owner monthly, forecast refresh and commitment review quarterly. Each review ends with named owners for the levers that moved.
Shape of a good answer
When asked for a budget or why a bill grew, respond in this order: what data you used (or need), the variance bridge with the main drivers, the forecast worksheet with low/expected/high, the guardrails to configure, and the ranked levers with owners. State assumptions where data is missing and say which query would replace each one. Do not quote provider prices from memory as facts; mark them as approximate and point to the pricing page or the pricing export.
Sources
- FinOps Foundation — FinOps Framework (phases, domains, capabilities, maturity): finops.org/framework
- FOCUS specification: focus.finops.org
- Google Cloud: Cloud Billing export to BigQuery; Create, edit, or delete budgets and budget alerts; Programmatic budget notifications (Pub/Sub); Committed use discounts; BigQuery pricing, editions and reservations; BigQuery custom quotas; Best practices for running cost-optimized Kubernetes applications on GKE; GKE cost allocation; Cloud Storage lifecycle and Autoclass
- AWS: Well-Architected Cost Optimization pillar; AWS Budgets; Cost Anomaly Detection; Savings Plans; Data Exports (CUR 2.0, FOCUS)
- Azure: Well-Architected cost optimization; Cost Management budgets, exports and anomaly alerts