A data platform bill is an engineering decision read backward. Instance sizes, storage classes, replication factors, erasure coding width, retention windows, egress patterns: every line has an owner and a reason. As an independent consultant I trace the bill back to those choices, rank the waste, and keep the headroom worth paying for.
Proof point: GPU infrastructure audits where allocated capacity ran far ahead of measured utilization; object storage workloads where erasure coding quietly multiplied small files. References available under NDA.
What the review examines
Measured utilization against allocated capacity, compute first. Then storage: classes and tiering policy, replication and erasure coding overhead, the small file penalty that multiplies metadata and repair traffic, backup and DR duplicates, cross-zone and egress traffic, license and support lines. Each cost is tied to the engineering choice that produces it, so the fix is a decision, not a discount negotiation.
Deliverables
A ranked waste list: each line with its measured cost, the change that removes it and the failure scenario it does not touch. A capacity model that separates waste from headroom and projects both over 3 and 5 years, so the savings survive growth. Where a capacity planning engagement already covers the ground, its measurements feed this one. References available under NDA.
The review is diagnostic: I sell no implementation, no licenses and no hardware, so nothing in the conclusions depends on what you buy next.
FAQ
How do you tell waste from headroom?
Headroom is utilization a failure scenario will consume: replication, DR copies, burst margin. Waste is capacity nothing in the failure model would ever use. Both are measured against your documented failure scenarios, so cutting the second never weakens the first.
Will cost optimization reduce reliability?
Not if the ranking is honest: redundancy that buys failure tolerance is priced and kept, duplication that buys nothing is removed. Every recommendation names the failure scenario it still covers, and the ones it does not, so the trade-off is explicit before anything is switched off.