Right-size before you buy another Savings Plan
Feed average and peak utilization — get a conservative downsize recommendation and monthly savings. Commitments on oversized instances lock in waste.
- No signup
- Instant results
- ~2 min
Rates reviewed
Current instance
us-east-1list rates, pricing as of 2026-07-22. Source: AWS public pricing.
Who This Tool Is For
FinOps and platform engineers who see low CPU in CloudWatch but need a defensible downsize recommendation before changing production.
Why We Built This Tool
Buying a Savings Plan on an oversized fleet locks in waste for 1–3 years. This tool only recommends a downsize when average CPU is under 20%, memory under 40%, and peaks stay under the workload threshold — conservative by design. Rates as of 2026-07-22.
What Problem It Solves
- Averages lie. Peak CPU gates the recommendation.
- EC2 and RDS. Same logic, service-specific rate cards.
- Commit later. Explicitly sequenced after rightsizing.
- Risk notes. Every recommendation includes a canary caution.
Next: Savings Plans ROI or Commitment Level Optimizer.
Frequently Asked Questions
When does this recommend a downsize?
Only when average CPU is under 20%, average memory under 40%, and peak CPU under 60% (80% for batch). Otherwise it recommends keep — or burstable for spiky low-average EC2 workloads.
Is this a substitute for Compute Optimizer?
No. Use this for a quick planning estimate. Confirm with AWS Compute Optimizer and a staging soak before production changes.
Why right-size before Savings Plans?
Commitments discount whatever baseline you have. If you commit first, you pay for capacity you are about to remove.
Which instance families are supported?
Common EC2 ladders (t3, m5, m6i, c5, r5) and RDS classes in our rate catalog (db.t3, db.m6i, db.r6i). Other families need a manual rate lookup.
What about memory-optimized workloads?
Memory average is a hard gate — if memory sits above 40%, we will not recommend a downsize even when CPU is low.
How accurate is the dollar savings?
On-Demand us-east-1 list rates as of 2026-07-22 × 730 hours × instance count. Spot, RI, and Multi-AZ are not modeled.