June 12, 2026

GCP sustained use vs committed use: pick the discount that matches your uptime

Google Cloud gives you two ways to discount compute without touching Spot VMs or custom machine types. Sustained use discounts apply automatically as an instance runs more hours in the month. Committed use discounts require you to pledge vCPU and memory spend for one or three years. Most teams get SUDs for free and never buy a CUD because the commitment paperwork feels like AWS Reserved Instances all over again. That is half right and half expensive.

SUDs reward uptime, not planning. A n2-standard-8 running every hour of September earns the maximum tier, roughly 30 percent off the on-demand vCPU and memory rate. The same machine running 50 percent of the month earns a smaller tier. You do nothing. The discount appears on the invoice as Sustained Use Discount.

The baseline math

An n2-standard-8 in us-central1 costs about $0.388 per hour on-demand, roughly $283 per month at full uptime. With maximum SUD the effective rate drops toward $198, saving about $85. A one-year committed use discount on the same shape at a typical 37 percent rate brings spend to about $178, saving $105 over on-demand but only $17 more than SUD alone.

Stretch to a three-year CUD at roughly 55 percent and effective monthly cost lands near $127, saving $156 over on-demand and $71 over SUD-only. The incremental win of CUD over SUD is the number finance needs. It is not the headline CUD percentage.

When sustained use is enough

1. Variable uptime between 60 and 85 percent

Staging clusters, CI workers, and services with autoscaling headroom often sit in the middle SUD tiers. A CUD pledge on peak capacity you only hit during deploy week overcommits spend and leaves you paying for vCPU-hours you never consume.

2. Workloads you plan to migrate off Compute Engine

Cloud Run, GKE Autopilot, and BigQuery slot commitments change the spend shape. Buy CUDs after the platform decision is stable, not while every team is still arguing about Kubernetes versus serverless.

3. Short-lived project accounts

Three-year CUDs on a product line with an 18-month roadmap is a balance-sheet bet. SUDs cost nothing to exit.

When committed use wins

1. Baseline compute above 90 percent monthly uptime

Core GKE node pools, always-on data pipelines, and managed instance groups with a fixed minimum are CUD candidates. Measure vCPU-hours for 90 days, take the P10 floor, and commit at that level rather than at peak.

2. Spend already above $10k per month on a single project

The administrative overhead of tracking CUD coverage only pays back at scale. Below that threshold, rightsizing and SUD tiers usually beat the incremental CUD savings.

3. Three-year horizon with stable machine families

CUDs are tied to region and general machine class, not a single instance name, which gives more flexibility than AWS Standard RIs. Still, a generational shift from n2 to n4 breaks the savings model until you renegotiate. Treat three-year CUDs like datacenter leases.

The audit path

In Billing, open Reports and group by SKU with discount type. SUD lines show as Sustained Use Discount. CUD lines show as Committed Use Discount. Export BigQuery billing export and sum compute.googleapis.com vCPU and RAM cost before and after credits for the last 90 days. Compare average monthly vCPU spend to active CUD commitments in the Commitments page. Unused commitment value is the gap to close, either by growing into it or by letting the term expire and recommitting lower.

Use the calculator before the commitment

The free GCP committed use savings calculator puts SUD, one-year CUD, and three-year CUD side by side on the same monthly compute spend and uptime percentage. If the gap between SUD and one-year CUD is under five percent at your actual uptime, skip the purchase and fix rightsizing first. If three-year CUD incremental savings clear $500 per month on a stable floor, open the commitment.

Google documents both programs on the sustained use and committed use discount pages. CUD percentages vary by region and machine family. Override the calculator defaults with your billing export when precision matters.