// Found and modeled a hidden COGS line tied to multi-tenant storage. Margins recovered within two quarters.
Problem
A Series B platform with healthy ARR growth and a gross-margin number that had been sliding for three quarters. The CFO's reports said it was a cloud-cost problem. The CTO's reports said it was a customer-mix problem. They were both right, and they were both incomplete — the two views had never been reconciled on a per-customer basis.
The board had asked for a margin-recovery plan in time for the next round. The internal team had bandwidth to build, not to investigate, and an outside view was cheaper than a re-org. We were brought in for an eight-week engagement: figure out where the margin had gone, model the recovery, and leave behind something the team could re-run quarterly.
Approach
The first three weeks were measurement. We instrumented the database tier to attribute storage, IOPS, and query cost to a tenant. We mapped that against the customer record — tier, ARR, signup cohort. The shape that fell out was a textbook power law: the top 4% of customers consumed 38% of database cost, and the bottom 22% consumed enough to matter despite paying free-tier or near-free-tier rates.
The next three weeks were modeling. We rebuilt the unit-economics worksheet with infrastructure as a per-customer COGS line, not a flat overhead. The picture was unambiguous: the platform's "team" tier was profitable, the "free" tier was a deliberate loss-leader (correctly), and the "enterprise" tier — which the sales team treated as the high-margin product — was the worst margin in the catalog because of one query pattern that scaled non-linearly with seats.
The last two weeks were recommendation. Three changes, modeled out four quarters: meter storage as a usage-based line item; introduce a true enterprise capacity tier for the top decile; and add a tenant-level quota system to cap the worst case for the rest. The team's job was to execute. Ours was to leave the model behind so the same analysis could be re-run every quarter without us.
Decisions & trade-offs
- Per-tenant cost attribution before any recommendation. The CFO's hypothesis and the CTO's hypothesis were both true and neither was actionable until the cost-per-customer distribution was visible. Three weeks of measurement saved a year of misaimed effort.
- Repriced rather than rebuilt. The temptation was to recommend a migration to pooled-isolated tenancy. That would have been an eighteen-month project. The pricing change recovered most of the margin in two quarters, and bought time to do the architectural work without it being a fire.
- Left the model behind, not just the answer. The deliverable was a worksheet the finance team owns and a runbook the engineering team owns. The recommendation runs itself now; we're not the bottleneck on the next review.
Outcome
Gross margin recovered by twelve points over the two quarters following the engagement — slightly ahead of the model. The enterprise tier was rebuilt around capacity rather than features, which also resolved a long-running sales-engineering dispute about scope creep on that tier. The unit-economics worksheet has been re-run twice since handoff without our involvement, which is the result we cared about.