Skip to content
FinOps workspace
COST LEDGER / DECISION GUIDE

GCP cost optimization: audit and implementation

Updated: 10 November 2026Get a quote for my project

GCP cost optimization starts with billing evidence that your team can reconcile. A Google Cloud cost optimization review should connect projects to owners and evaluate workload changes before sizing new commitments.

Illustrative single server cabinet in a bright room with large windows
Infrastructure illustration, unrelated to any customer account.

For build and deployment artifacts, the Artifact Registry cost optimization review covers retained versions, download routes and controlled cleanup. For analytics guardrails, use the BigQuery cost-control matrix to distinguish query limits from billing notifications.

On this page
  1. Enable the exports first, because not all of them backfill
  2. GKE: read the allocation data for what it is
  3. Where findings come from
  4. Sequence of a review
  5. Deliverables and what they cost you
  6. Permissions for the review
  7. Official sources
  8. Questions about a Google Cloud review

For machine learning in BigQuery, separate model training and remote inference with the BQML pricing and ML workload cost guide.

For serverless container services, the Cloud Run cost optimization review connects billing mode, warm capacity and concurrency with application acceptance tests.

For a concrete analytics estimate, the 12 GB and 12 GiB BigQuery query calculation distinguishes scan units, free allowance and capacity billing. For recovery storage, the Compute Engine snapshot cost review covers incremental dependencies and retention rules.

GCP cost optimization: separate analytics and capacity

FinOps on Google Cloud needs workload-specific evidence. A BigQuery scan, an idle VM and GPU capacity are different cost decisions. Use the BigQuery pricing review for analytics and the cloud GPU cost comparison for accelerator workloads. GCP cost optimization consulting should preserve those separate boundaries in the recommendation register.

Enable the exports first, because not all of them backfill

Cloud Billing export to BigQuery is configured per export type, and Google recommends enabling it when the billing account is created. The documentation lists these types:

  • Standard usage cost: account-level costs, usage, credits and adjustments, suited to broad trends.
  • Detailed usage cost: the standard content plus resource-level data, needed to find which resource drives a line.
  • Pricing data: prices by service and SKU.
  • CUD metadata and FOCUS usage cost, both listed as Preview in the documentation reviewed.

For standard, detailed and FOCUS usage-cost exports, initial backfill depends on dataset location. On first enablement, an EU or US multi-region dataset can receive data from the start of the previous month; the initial load can take up to five days. A supported single-region dataset begins at enablement. Pricing exports have no retroactive data. Re-enabling an export or changing its destination does not backfill previously exported data into the new dataset. If the export was enabled late, the provider's analysis window is shorter than the invoice history, and the proposal should say so. The dataset location is also the place where your billing detail is stored, so choose it with your data protection lead.

GKE: read the allocation data for what it is

GKE cost allocation writes cluster and workload attributes into the labels of the detailed export, among them the namespace, the workload type and name, and Kubernetes Pod labels. If a Pod has more than 50 labels, none of its Pod labels appear in billing cost allocation. Four limits matter when you read it:

  • Allocation is based on resource requests, not on resources consumed.
  • Data can take up to three days to appear and is not backfilled before the feature was enabled.
  • Use the detailed usage export for the allocation labels. Cloud Billing also documents GKE breakdown in its FOCUS export after cost allocation is enabled; the standard export does not provide this breakdown.
  • Two special namespaces identify system overhead and capacity not requested by workloads: kube:system-overhead for capacity reserved for Kubernetes itself and kube:unallocated for capacity requested by no workload.

A large kube:unallocated share is a finding in itself: it usually points to node pools larger than the requests placed on them.

Where findings come from

Google's Recommender publishes cost-related recommenders that a review should run through, each with its own limits:

  • VM machine type recommender: analyses CPU and memory over an eight-day observation window using 60-second averages, so short spikes can be missed. Installing the Ops Agent improves the memory signal. It gives no recommendation for instances with GPUs, TPUs or ephemeral disks, or in the memory-optimized family.
  • Idle VM, idle persistent disk, idle IP address and idle custom image recommenders: candidates for removal once an owner confirms. Back up a disk before deleting it.
  • Cloud SQL idle and overprovisioned instance recommenders.
  • Cloud Billing cost anomalies and budgets for spend that moves unexpectedly between reviews.

Applying a machine type change needs stop and start permissions on the instance, which is a production change for your team or the provider under your change process.

Sequence of a review

  1. Reconcile: tie export totals to invoices for at least one closed month, and list credits separately.
  2. Allocate: map projects, folders and labels to owners. Unlabelled spend gets a named holder, not a blank.
  3. Rank: sort the top services and SKUs by spend, then by month-on-month change.
  4. Investigate: for each top item, confirm with the owner whether the demand is intended.
  5. Act on usage: idle resources, machine types, storage classes and scheduling, each with an owner and rollback plan.
  6. Commit last: size commitments on the remaining floor, as described in the guide on committed use discounts.

Carry the selected migration scope into the GCP TCO calculator scenario checklist. Reconcile asset coverage and estimated sizing before adding the operating and transition costs required for the ownership comparison.

Deliverables and what they cost you

Ask for a findings register (item, evidence, owner, action, rollback, expected billing effect), the BigQuery queries used, and the reconciliation to the invoice. Note that BigQuery usage charges apply when you store and analyse billing data in it, so ask the provider to state how much querying it expects to run in your project or whether it will work from a copy elsewhere.

Plan internal time as well: one engineer per major workload to confirm intent, and one finance contact for reconciliation.

Want this sequence run on your own Google Cloud billing data? Scope my project.

Illustrative server rack in the foreground with a row of cabinets behind it
Illustration only: findings come from billing exports and recommenders.

Permissions for the review

Separate analysis from administration. Google’s billing role reference gives Billing Account Viewer read access to costs and pricing; Billing Account Costs Manager can manage budgets and cost exports, but not pricing exports. For SQL analysis, grant data read access only to the required dataset and permission to run query jobs on the designated analysis project, for example BigQuery Job User. Billing roles do not automatically grant resource, metrics or recommender access. Define those permissions separately and set a revocation date.

Official sources

Official documentation used for this page.

Questions about a Google Cloud review

Should billing exports be sent in the quote form?

No. State the cloud and the approximate scope only. Agree a secure transfer and the minimum dataset with the specialist after scoping.

Does an old CUD article describe the current model?

Not necessarily. Compare the article's date with the Compute Engine release notes and the purchase options available in your billing account.

Can we recover billing detail from before we enabled the export?

Only partly. On first enablement, standard, detailed and FOCUS usage-cost exports in EU or US multi-region datasets can backfill the current and previous month. Single-region and pricing exports do not recover earlier data. Re-enablement and destination changes have separate limits; retain earlier datasets and invoices.

Questions before a cost review

What is cloud cost optimization?

It connects billed usage to required business outcomes, then reduces unnecessary consumption or changes the purchase model with service constraints preserved.

Is GCP cheaper than AWS?

No cloud is universally cheaper. Compare an equivalent workload and location, contract rates, availability and operating requirements. Include migration and engineering costs in the decision.

How to calculate GCP cost?

Use the official Google calculator for the planned configuration and current price assumptions, then apply your contract terms and include operational and implementation costs. Compare the estimate with the billing export for an existing workload. Open the Google Cloud calculator.

Workload guides for this review

Continue the review with bigquery pricing. Each guide addresses a separate decision within the same cloud cost boundary.

Estimate BigQuery costs for query processing and storage as separate lines. Record whether query compute is billed on demand or through capacity before comparing workload changes.

Continue your review

Get support on my project

A specialist in the subject reviews your request and sends you a quote.

Get support on my project

Further decisions in this review

Continue the workload review with gcp storage pricing, gke cost optimization. Keep the selected decisions in the scope and ask for the evidence needed to approve each change.

Get my FinOps quote