Been working on a read-only GCP cost scanner for a while. Figured the rule list might be worth sharing, both as a reference and to hear what I'm missing for people running real GCP workloads.

Hygiene rules (5):

Stopped/terminated Compute Engine VMs (attached disk charges continue regardless)

Unattached Persistent Disks

Disk snapshots older than 90 days

Reserved static external IPs in RESERVED state ($0.01/hr whether used or not)

Cloud SQL instances with zero connections for 14+ days

AI/ML rules — opt-in with --category ai (5):

Vertex AI Online Prediction endpoints with an always-deployed replica floor and zero

observed requests — these stay billed even with no traffic

Vertex AI Workbench instances with no activity (still tightening this one)

Vertex AI training jobs running well beyond a normal threshold (still tightening this one)

Cloud TPU nodes in READY state with near-zero accelerator utilization

Vertex AI Feature Stores (legacy and Bigtable-backed) with zero serving requests for 30+…