AWS Compute Optimizer vs. Manual Rightsizing: What It Catches and What It Misses

AWS Compute Optimizer is a genuinely good product, and it's free. It watches 14-32 days of CPU, memory, network, and disk metrics for your EC2 instances, Auto Scaling groups, EBS volumes, Lambda functions, and ECS services on Fargate, and recommends a smaller (or larger — it flags under-provisioning too) size based on real utilization, not a generic rule of thumb. If you're not already looking at it, that's the first thing to fix before buying any third-party tool, including this one.

You can see your own recommendations right now:

aws compute-optimizer get-ec2-instance-recommendations \
  --query "instanceRecommendations[?finding=='OVER_PROVISIONED']"

Where it stops

Compute Optimizer answers one question: given a resource that exists, is it the right size? It has no opinion on resources that shouldn't exist at all, because that's a different question entirely. A handful of concrete things it will never flag, because they're structurally outside what it looks at:

• An idle NAT Gateway. Not an EC2 instance, not in scope — but it bills exactly like one, hourly, regardless of size.
• An unattached EBS volume. Compute Optimizer only rightsizes volumes that are attached to something; a volume with nothing attached isn't a sizing problem, it's a cleanup problem.
• A load balancer with no healthy targets. Same story — it's not a compute resource Compute Optimizer watches at all.
• An S3 bucket that never transitions to cheaper storage. Storage-class lifecycle is a completely separate lever from compute rightsizing.
• A Savings Plan or Reserved Instance opportunity for commitment-eligible, steady-state spend. Compute Optimizer doesn't cover this at all — that recommendation comes from Cost Explorer's own Savings Plans API, a separate call most tools (including this one) have to make on top of Compute Optimizer, not instead of it.

The other gap: the enhanced-metrics upsell

Compute Optimizer's free tier uses standard CloudWatch metrics at 14-32 day lookback. There's an "enhanced infrastructure metrics" option that extends the lookback and sharpens the recommendation — and it costs money per instance-month. Whether that's worth it depends entirely on how much is riding on getting a specific instance's sizing exactly right; for most fleets, the free tier's recommendation is already good enough to act on.

So what should actually check for what

The practical split: let Compute Optimizer do rightsizing, because it's free, already running, and backed by actual ML over your real usage — rebuilding that badly from scratch is a waste of engineering time. Then separately check for the things that are structurally invisible to it: orphaned networking resources, storage that never got a lifecycle policy, load balancers nothing is using, and the Savings Plans recommendation Cost Explorer holds separately. None of that is rightsizing. All of it is still money.