AWS Cost Optimizer

AWS Cost Optimizer

Overview

Keywords: AWS cost, FinOps, right-sizing, idle resources, orphaned resources, Savings Plans, Reserved Instances, RI, Spot, cost allocation tags, Cost Explorer, CUR, unblended cost, amortized cost, NAT gateway, gp2 to gp3, S3 lifecycle, Graviton, Compute Optimizer, anomaly detection.

This skill drives a repeatable cost-reduction workflow across the five highest-leverage levers: eliminate (idle/orphaned), right-size (over-provisioned), commit (Savings Plans / RIs), modernize (newer/cheaper SKUs and storage tiers), and attribute (tagging + showback). Always quantify monthly savings, rank by effort-vs-impact, and never recommend a change without stating its risk and rollback.

Use the bundled assets:

• references/cost-levers.md — the full catalog of cost-reduction levers per service with detection signals and typical savings.
• references/savings-plans-vs-ri.md — decision framework and break-even math for commitment purchases.
• references/tagging-strategy.md — cost-allocation tag taxonomy and enforcement patterns.
• scripts/cost_analysis.py — stdlib-only analyzer that scores resources for waste and ranks recommendations from JSON inventory.
• templates/cost-optimization-report.md — fill-in deliverable for stakeholders.
• examples/right-sizing-walkthrough.md — a concrete end-to-end example.

Workflow

1. Establish the baseline. Pull the last 1–3 months from Cost Explorer or the Cost & Usage Report (CUR). Always reason in amortized cost (commitments spread over their term), not unblended, so commitments aren't double-counted. Break spend down by service, by linked account, and by tag. Record the top 10 cost drivers — typically EC2, RDS, S3, data transfer, and NAT gateways.

2. Eliminate waste first (zero risk, fast). Idle and orphaned resources are pure savings with no performance trade-off. Scan for: unattached EBS volumes, unassociated Elastic IPs, idle/empty load balancers, stopped instances still holding EBS, old EBS/RDS snapshots, idle NAT gateways, dev/test resources running 24/7, and forgotten dev environments. See references/cost-levers.md "Eliminate" section.

3. Right-size over-provisioned resources. Use CloudWatch / Compute Optimizer signals (p95 CPU, memory if the agent is installed, network, IOPS). Flag instances with sustained low utilization (e.g., p95 CPU < 40% and max < 60% over 14 days). Recommend the smallest instance that keeps headroom. Prefer downsizing one family step at a time. See references/cost-levers.md "Right-size".

4. Modernize SKUs and storage. Migrate gp2→gp3 EBS (≈20% cheaper + tunable IOPS), x86→Graviton (≈20% better price/performance), older instance generations→current (m5→m6/m7), and apply S3 lifecycle / Intelligent-Tiering. These are recurring savings with low risk after testing.

5. Commit to what's left (after steps 2–4). Only buy Savings Plans / RIs for the stable baseline that remains after eliminating and right-sizing — never commit to waste. Use references/savings-plans-vs-ri.md to choose Compute SP vs EC2 Instance SP vs RIs, term (1yr vs 3yr), and payment option. Target coverage of the bottom 60–80% of steady-state usage; leave the variable top on on-demand or Spot.

6. Attribute costs with tags. Implement the tag taxonomy in references/tagging-strategy.md, activate cost-allocation tags in Billing, enforce with tag policies / SCPs, and set up showback per team/product. Untagged spend above a threshold is itself a finding.

7. Quantify, rank, and report. Run scripts/cost_analysis.py over an inventory JSON to score and rank findings by monthly savings and implementation effort. Produce the deliverable from templates/cost-optimization-report.md with prioritized recommendations, total addressable savings, and a rollout order.

8. Institutionalize. Turn on AWS Cost Anomaly Detection and budgets with alerts, schedule a monthly review, and set a coverage/utilization target for commitments. Cost optimization is continuous, not a one-time project.

Prioritization framework (effort vs. impact)

Rank every finding into one of these buckets. Do the top-left first.