Cost Management
Cost management is the feedback loop that connects architecture choices to spend. It is not just “make it cheaper”: an SLO, a reliability target, a GPU systems memory requirement, and a scalability target can all be valid reasons to pay more. The important mechanism is attribution: every material cost should have an owner, workload, environment, and unit driver.
What drives cloud cost
Cloud cost is usually:
For systems work, the practical unit is more useful than the bill total: dollars per training run, dollars per million requests, dollars per TB retained, dollars per successful document ingestion, or dollars per evaluated model. Managed storage and managed compute need separate tags because storage retention and compute bursts scale differently. Generative systems should also track token, retrieval, reranking, and evaluation traffic separately, as in cost and latency optimization.
Worked cost check
Using AWS-published S3 example rates for inter-region transfer, Multi-Region Access Point routing, and internet egress:
| Driver | Usage | Rate | Cost |
|---|---|---|---|
| Cross-region replication | 2,000 GB | $0.0200/GB | $40.00 |
| Multi-Region Access Point routing | 500 GB | $0.0033/GB | $1.65 |
| Internet egress | 20 GB | $0.0900/GB | $1.80 |
The total is 1.65+43.45. The small routing line is not the point; the 2 TB replication line is. Cost review should follow data movement and retention, not only instance size. A CostCenter=search, Environment=prod, Workload=rag-ingestion tag set is only useful if budgets and Cost Explorer reports group by those labels.
Caveats
Cutting observability, backups, or evaluation can make the bill smaller while increasing recovery time and defect cost. Committed-use discounts and reserved instances reduce unit price but introduce utilization risk. Autoscaling without request caps can turn a dependency failure into a retry-driven bill spike.
References
Nav
Section — Cloud and Distributed Systems