Google Cloud pricing explained: sustained use, committed use and per second billing

Google Cloud pricing differs from its two larger rivals in one significant way: a meaningful part of the discount is automatic. You do not sign anything, you do not commit, and the reduction appears because the machine kept running. That is genuinely useful, and it also makes the invoice harder to predict if you do not know the rules.

Per second billing with a one minute floor

Compute Engine bills per second after a one minute minimum. For short lived work, that is materially cheaper than hourly rounding: a fleet of build agents that live for four minutes each is billed for four minutes each.

It also changes how you think about scaling. Aggressive scale down is worth more here than on platforms that round up, because the saving is immediate rather than deferred to the next billing boundary.

Sustained use discounts

On eligible machine families, Google applies an increasing discount as an instance runs for a larger share of the month. Run something continuously and the effective rate drops without any action on your part.

Two details make this more useful than it first appears. The discount is calculated on the aggregate of a machine type in a region, not per instance, so replacing an instance mid month does not reset the clock. And it stacks with the free tier where that applies.

It does not apply everywhere. Newer families and some specialised types are excluded, and committed use discounts replace it rather than adding to it.

Committed use discounts

Two shapes, and the difference matters:

• Resource based commitments. You commit to a quantity of vCPU and memory in a specific region for one or three years. The discount is the deepest available, and the commitment is tied to that region and family group.
• Spend based commitments. You commit to an hourly spend figure and the discount applies flexibly across eligible services. Smaller discount, far more freedom.

The decision is the same as on any cloud: commit to the part of your usage that genuinely runs all year, and let everything above it float on demand or on spot.

Spot VMs

Spot capacity is spare capacity, offered at a large discount and reclaimed when Google needs it back, with a short termination notice. It suits batch pipelines, CI, rendering, and anything that checkpoints and resumes.

Managed instance groups make this practical: mix a small on demand base with a spot majority, and the group replaces reclaimed instances automatically. That pattern gets most of the saving with none of the fragility.

Custom machine types

Compute Engine lets you specify vCPU and memory independently within family limits rather than choosing from a fixed ladder. If your workload needs 6 vCPUs and 40 GB of memory, you can buy exactly that rather than paying for the next size up on both axes.

For memory heavy or CPU heavy workloads that sit awkwardly between standard sizes, this is a real saving and a genuine differentiator against the fixed size catalogues elsewhere.

The Google Cloud pricing line items that surprise people

• Persistent disks bill on provisioned size. A 500 GB disk holding 20 GB costs the same as a full one, and it survives instance deletion unless you set the disk to auto delete.
• External IP addresses. Charged hourly, and an unattached static address costs more than an attached one.
• Network egress. Internet egress, inter region traffic and inter zone traffic all have separate rates. This is the single most common source of unexplained cost, and it is covered across providers in data transfer costs.
• Cloud Logging ingestion. Generous free allowance, then per gigabyte. A chatty application in debug mode can move this to the top of the bill.
• Load balancers. A forwarding rule bills hourly whether or not it serves traffic.
• Snapshots. Incremental and cheap, but they accumulate and nobody deletes them.

A cost routine that works

1. Separate projects per environment, so cost attribution is structural rather than a tagging exercise.
2. Budget alerts on the billing account at several thresholds.
3. Billing export to BigQuery from day one. It cannot be backfilled and it is what you will want when somebody asks why the bill moved.
4. Recommender is worth checking monthly. It flags idle disks, idle addresses and oversized instances reliably.
5. Wait for three months of stable usage before buying any commitment.

How it compares

Against AWS and Azure, the headline differences are the automatic sustained use discount, per second billing and custom machine shapes. Against them, AWS has the deepest service catalogue and Azure the strongest position for organisations already licensed with Microsoft. The specifics are in AWS vs Google Cloud and Azure vs Google Cloud.

For the account structure that all of this sits inside, see Google Cloud account options.