Cloud Cost Optimization Vibe Code Cleanup

What AI Gets Wrong in Cloud Costs

AI provisions infrastructure based on documentation and tutorials, not your actual workload. The result: an RDS db.r5.xlarge ($500/month) for an application that could run on db.t3.micro ($15/month). An ElastiCache cluster for an application that doesn’t need caching. A multi-AZ NAT gateway setup ($100+/month) for a development environment. Each recommendation is reasonable in isolation. Together, they create a $2,000/month bill for an application serving 50 users.

AI doesn’t clean up. Terraform files from experimentation create resources that persist. S3 buckets from features that were abandoned. Lambda functions from approaches that were discarded. Each costs a few dollars per month but the accumulation is significant.

Platform over-engineering is common. AI recommends AWS because it’s the most documented cloud provider. A simple web application gets ECS, RDS, ElastiCache, SQS, and CloudFront - enterprise infrastructure for an MVP that would run perfectly on Railway for $20/month.

Our Cost Cleanup Process

We start with utilization analysis. Every compute resource is checked against actual CPU, memory, and I/O usage. Databases are assessed against actual query load and storage needs. The gap between provisioned and actual usage is quantified in dollars.

Platform fit assessment questions the fundamental hosting choice. Does this application need AWS, or would a simpler platform serve it better at lower cost? We’ve migrated applications from $2,000/month AWS setups to $40/month Railway deployments with identical functionality and better developer experience.

For applications that belong on AWS, we right-size everything. Instance types match actual workloads. Auto-scaling replaces static over-provisioning. Reserved instances lock in savings for stable workloads. Unused resources are decommissioned. Cost monitoring ensures the optimization persists.

Specific Waste Patterns We Find Repeatedly

AI-generated infrastructure tends to produce the same categories of waste across projects. Knowing these patterns lets us move quickly during an audit.

First, networking costs. AI loves setting up NAT gateways in every availability zone because AWS documentation recommends it for high availability. Each NAT gateway costs roughly $32/month plus data processing charges. For an early-stage application that could run entirely in a public subnet behind an ALB, or use VPC endpoints for AWS service access, removing unnecessary NAT gateways alone can save $100-200/month.

Second, storage over-allocation. AI provisions gp3 EBS volumes at 500GB when the application uses 8GB. It enables S3 versioning and cross-region replication on buckets that store temporary uploads. RDS instances get provisioned IOPS storage when baseline gp3 performance would be more than sufficient. We audit actual storage consumption and I/O patterns, then resize volumes and adjust storage tiers to match.

Third, compute redundancy. AI scaffolds separate ECS services for tasks that could be a single process. A web server, a background worker, and a scheduler each running on their own Fargate task with 1 vCPU and 2GB RAM when the combined workload would fit comfortably in a single $5/month container. We consolidate where consolidation makes sense and separate where true isolation is warranted.

Finally, there is the issue of data transfer costs, which AI rarely accounts for. Services chattering across availability zones, CloudFront distributions configured to pass every request to the origin, and Lambda functions making redundant API calls all generate data transfer charges that show up as a mysterious line item on the bill. We trace data flows and eliminate unnecessary cross-AZ and cross-region traffic.

Before and After

Before: An AI-provisioned AWS setup costing $2,400/month for an application with 200 monthly active users. Over-provisioned database, unnecessary cache layer, and three unused services from abandoned features.

After: The same application on right-sized infrastructure - either Railway at $40/month or optimized AWS at $400/month depending on requirements. Auto-scaling handles traffic variation. Cost monitoring tracks spending. The savings fund months of additional runway.