AWS Cost Teardowns
The patterns we look for in every audit — NAT gateways, storage tiers, commitment coverage, and the rest. Anonymized real teardowns with exact dollar figures follow as audits close.
AWS runs three different meters under its AI services
August 8, 2026Tokens meter payload, endpoints meter existence, "serverless" vector stores meter availability — each fails in a different shape, and each shape needs different alerting.
Dimensions are for values you could list from memory
August 7, 2026Every unique dimension combination is a billed metric. One "observability improvement" with a container-id dimension prices your CloudWatch bill in your traffic's units.
SageMaker endpoints bill for existing, not inferring
August 6, 2026A real-time endpoint holds an instance 24/7 from deploy until delete — the demo that succeeded in April is still billing in August. Where idle endpoints come from and the hosting modes built for exactly this.
The capacity floor under your RAG prototype
August 5, 2026"Serverless" vector search does not scale to zero — OCU minimums bill around the clock, and the collection outlives the experiment it was built for.
Your Bedrock bill is a payload property, not a traffic property
August 3, 2026Request volume can stay flat while inference cost triples — tokens per request and model tier are the dials that move, and neither is on your dashboard.
A failing Lambda doesn't cost less
August 2, 2026Retries are full billed invocations, timeouts bill per attempt, and most event sources redeliver until something succeeds — how an outage becomes a billing event.
Tagging is the optimization that makes every other one possible
July 9, 2026The “shared/unallocated” cost bucket nobody owns, why it only grows, and a three-tag convention that actually survives.
Fargate bills the task definition, not the container
July 8, 2026Reserved vCPU and memory bill continuously regardless of utilization — and non-prod is usually sized identically to prod.
Provisioned vs. on-demand DynamoDB, and the GSI multiplier both modes share
July 7, 2026A bet on stable traffic vs. a premium for not thinking about it — and why every GSI multiplies your write cost.
The same bytes, billed once instead of per viewer
July 6, 2026Data transfer to the internet is billed per GB from wherever it leaves AWS — caching repeat requests changes where that GB gets billed.
Nothing is using it, and the bill doesn’t know that
July 5, 2026Unattached Elastic IPs, load balancers with no healthy targets, and snapshots that outlived their volume.
Every log group defaults to Never Expire, and almost nobody changes it
July 4, 2026Ingestion is the charge that actually costs the most — retention only controls how long the rest keeps billing storage.
Lambda memory is also your CPU dial, and most functions never get tuned
July 3, 2026Memory set once during a tutorial or an incident, provisioned concurrency left on after a launch — the two easiest Lambda costs to find.
The cross-AZ data transfer nobody put on the architecture diagram
July 2, 2026The tax for multi-AZ resilience is real and worth paying — the problem is when it’s paid without anyone deciding to.
Idle EC2 spend rarely looks like an idle instance
July 2, 2026Auto Scaling minimums set at peak, non-prod running 24/7, and the boxes nobody remembers the purpose of.
The three ways S3 storage cost quietly gets away from you
July 2, 2026Wrong storage class for the access pattern, orphaned multipart uploads, and versioning with no cleanup rule.
RDS instances get sized once and never revisited
July 2, 2026Reading utilization correctly, the real cost of Multi-AZ on non-production, and why storage autoscaling needs a ceiling.
gp2 to gp3: the EBS migration most accounts still haven’t done
July 2, 2026Lower per-GB price, better baseline performance, no downtime to convert. The one check to make before migrating everything in bulk.
Savings Plans vs. Reserved Instances: a simpler way to think about it
July 2, 2026Both trade a commitment for a discount. Where they differ, where accounts get burned, and a framework for picking the right one.
The NAT Gateway tax nobody budgets for
July 2, 2026An hourly charge and a per-GB fee that don’t care whether the traffic passing through was necessary. How to find and fix the traffic that shouldn’t be there.