Cloud Cost Audit in 10 Practical Steps (AWS, GCP & AI Spend)
A cloud cost audit does not require a dedicated FinOps team. By grouping spend by environment, shutting down idle staging resources, fixing unindexed database queries, and adding model routing to AI endpoints, most early-stage teams can identify 20% to 40% in avoidable monthly spend within two weeks.
Best for: Founders, engineering leads, and Series A teams spending $5,000+ per month on cloud and AI APIs.
Tag and group every dollar by environment and service
Separate production, staging, dev, and data/AI spend so you can see where money actually goes.
- Export the last 90 days of daily cost by service (EC2/Compute, RDS/DB, NAT/Egress, LLM APIs).
- Flag any untagged resources exceeding $100/month.
Shut down idle non-production environments overnight
Staging and QA servers rarely need to run 168 hours per week.
- Schedule staging compute and dev databases to pause outside working hours.
- Delete orphaned preview environments older than 14 days.
Remove unattached volumes, old snapshots, and idle load balancers
Leftover EBS volumes and unused NAT gateways quietly accumulate thousands per year.
- Audit unattached storage volumes and snapshots older than 90 days.
- Consolidate redundant load balancers and NAT gateways.
Rightsize oversized compute and database instances
Inspect 30-day peak CPU and memory utilization—instances sitting under 20% peak are over-provisioned.
- Step down oversized RDS/CloudSQL instances by one tier in staging first, then production.
- Switch general compute workloads to current-generation ARM/Graviton instances where compatible.
Fix the top 5 slowest database queries
Unindexed queries force you to buy larger databases than your traffic requires.
- Enable slow-query logging (pg_stat_statements or equivalent).
- Add missing composite indexes and eliminate N+1 query loops.
Audit data transfer and NAT egress fees
Cross-AZ and external API traffic often hides as 10–20% of an AWS bill.
- Add VPC gateway endpoints for S3 and DynamoDB to avoid NAT per-GB charges.
- Put static assets and public APIs behind a CDN with proper cache headers.
Measure LLM token cost per feature and per customer
Treat AI API spend as a first-class COGS metric.
- Log input/output tokens per endpoint and identify verbose system prompts.
- Cache deterministic or repeated AI lookups.
Introduce tiered model routing for AI tasks
Use smaller, faster models for classification, extraction, and routing; reserve frontier models for complex reasoning.
- Benchmark a 50-example test set on smaller models before switching.
- Set hard max-token and rate-limit caps per tenant.
Apply Savings Plans only after rightsizing
Never lock in 1-year commitments on bloated infrastructure.
- Commit only 60–70% of your steady-state baseline compute after cleanup.
- Avoid 3-year lock-ins at Pre-Seed or Seed stage.
Set automated daily anomaly alerts
Catch runaway loops in hours instead of at the end of the billing cycle.
- Configure daily spend alerts at 120% of rolling 7-day average routed to Slack/email.
- Review cloud unit cost per active customer once a month.
Want a senior CTO to run this with your team?
See how our matched fractional CTOs solve cloud costs rising.