Real AWS Bill Autopsy (Anonymised Case Study)

This is a real AWS bill from a Series A AI startup processing customer support tickets with RAG. We have anonymised the company but the numbers are real. Monthly spend: $34,200. After our optimisation engagement: $14,400. Here is where the money was going.

The Breakdown

Bedrock / LLM API Calls — $18,500 (54%) — The largest line item. The company was using Claude Sonnet for every single query — including simple ticket classification, status lookups, and FAQ responses that a much smaller model could handle. No caching was implemented, so identical questions generated fresh API calls every time.

EC2 Compute — $6,800 (20%) — Three m5.4xlarge instances running 24/7 for the application layer, plus two p3.2xlarge GPU instances for embedding generation. The GPU instances were running at 12% average utilisation — they were sized for peak load but peak load happened 2 hours per day.

RDS (PostgreSQL + pgvector) — $4,200 (12%) — A db.r6g.2xlarge instance with 500GB storage. Significantly oversized for the actual data volume. The pgvector indexes were unoptimised, forcing the database to work harder than necessary.

S3 + Data Transfer — $2,800 (8%) — Document storage and inter-service data transfer. The application was downloading full documents from S3 on every request instead of caching them locally.

Other (CloudWatch, NAT Gateway, etc.) — $1,900 (6%) — Verbose logging to CloudWatch at $0.50/GB ingested was adding up. The NAT Gateway charges were inflated by unnecessary cross-AZ traffic.

What We Changed

Model routing — Implemented a classifier that routes simple queries (60% of traffic) to a smaller model. Inference costs dropped 45%.

Semantic caching — Added a Redis-based cache for near-duplicate queries. Hit rate of 28% in the first week.

Right-sized compute — Moved embedding generation to spot instances with auto-scaling. Reduced the always-on GPU cost by 70%.

Database optimisation — Tuned pgvector indexes and downscaled the RDS instance. Cut database costs by 40%.

Caching and transfer — Implemented local document caching and consolidated services to reduce cross-AZ traffic.

The Result

Monthly bill dropped from $34,200 to $14,400 — a 58% reduction. No measurable quality degradation. Response latency actually improved because of caching and the faster small model on simple queries.