← Back to blog

Optimizing Google Cloud costs: compute, networking and storage

Lazy Read

Compare your Google Cloud bill with actual usage and service requirements before reducing resources. This article explains what to review across compute, containers, networking and data, and how to turn spending signals into measurable proposals without compromising availability or security.

  • Review capacity, occasional jobs and network costs.
  • Compare storage, queries and commitments with actual usage.
  • Validate each change with metrics and a rollback path.

Related links

A large machine for a ten-minute job. A gateway carrying hardly any traffic. Test data still stored months later. In Google Cloud, unnecessary spending often appears when infrastructure no longer matches the workload it serves.

Optimization starts by understanding that gap. Cutting a bill without checking latency, availability and security can end up costing more. This walkthrough applies that approach to Compute Engine, Cloud Run, GKE, networking and data.

Understand the bill first

In Cloud Billing, break down spending by project, service and SKU, and distinguish usage, credits and discounts. A rise attributed to an API project might come from queries or transfers between regions rather than its machines.

If a billing export to BigQuery already exists, use it to investigate changes and attribute costs. Enabling it changes configuration; storing and querying its data can also incur charges. A read-only audit should use existing access and leave any expansion for approval.

Identify resource owners and purposes. An unknown owner is a reason to investigate, not permission to delete.

Separate permanent capacity from occasional jobs

Review CPU, memory where instrumented, disks, networking and representative peaks. The Ops Agent collects guest-system signals that complement platform metrics. Installing agents is outside a read-only review.

Low average CPU does not prove a VM is unnecessary. It might process a critical import for an hour. Compare usage with business schedules and Compute Engine machine type recommendations.

For short, separable tasks, compare Cloud Run Jobs, Batch or ephemeral machines. Spot VMs are another option when workloads tolerate interruptions and retries. Check startup time, storage and operational costs before moving them.

Choose Cloud Run, Compute Engine or GKE for the workload

In Cloud Run, review concurrency, CPU, request duration and minimum instances. Keeping instances ready can improve latency, but also incurs costs without requests, depending on the billing mode.

A small VM may suit a steady workload. GKE offers controls and orchestration capabilities that can justify its complexity; it also requires reviewing nodes, resource requests, idle capacity and observability.

Compare alternatives against the same availability and security requirements. When considering a VPS or dedicated server, include administration, support, backups and data transfer. The price per vCPU is not the full cost.

Review network hops

Public NAT adds gateway, processed-data, external-IP and applicable outbound-transfer charges. Check which applications use each gateway and how much traffic they generate.

Also review lightly used load balancers, reserved IPs and interregional transfers. A quiet component may provide backup capacity. Simplify only after understanding the topology and trust boundaries; savings do not justify removing necessary isolation.

Data, queries and observability

Cloud Storage classes have different storage, retrieval and operation conditions. Nearline, Coldline and Archive include minimum storage durations. Changing class can save money while data stays untouched and cost more if it is retrieved or deleted early.

Review old object versions, disks without users and snapshots, preserving copies needed for recovery and retention. Autoclass can help with changing access patterns, but its charges should be compared with actual usage.

In BigQuery, investigate bytes processed, filters, partitioning, clustering and temporary tables. Avoiding an unnecessary query may achieve more than moving the whole platform. In Cloud Logging, review ingestion and retention while preserving the security and diagnostic signals you need.

Caching must also demonstrate its value: measure hits, latency and work avoided at the origin. Adding a component to save on another only pays off if total cost falls while behavior remains correct.

Commit after sizing

Committed Use Discounts exchange flexibility for discounts on eligible usage. Estimate steady demand after adjusting capacity, then review coverage, utilization and terms.

Do not commit resources just because they are running today. Temporary or oversized workloads can leave you paying for capacity you no longer need.

Turn signals into decisions

Compare at least several representative weeks of spending and utilization, extending the window for seasonal demand. For each proposal, document:

  • The resource and usage pattern that justify reviewing it.
  • Current and expected costs in comparable currencies and periods, including discounts and secondary costs.
  • Expected impact on errors, latency, availability and security.
  • Owner, dependencies, validation and rollback.

An inventory identifies candidates; it does not establish savings. Without billing data, sufficient metrics and service context, explain what is missing before recommending a size change or deletion.

A useful result is a plan ordered by benefit and risk that can be applied gradually and measured again. Sometimes that means a smaller machine; sometimes it means separating an occasional job, improving queries or avoiding unnecessary data transfers.

A skill to start the review

The GCP audit skill helps organize this review with an agent. It includes a bounded metadata and metrics collector and generates private local reports. It requires a read-only identity and an explicit project; automated inventory does not replace analyzing billing, workload peaks and service requirements.