"Why Gross Margin Is a Decision (Not Just a Calculation)" — CJ Gustafson (Mostly Metrics)
Second issue in CJ's "Mostly Controllers" series (translating Controller-produced financials for CFOs/investors/acquirers; first issue covered deferred revenue). This one argues gross margin is a classification choice, not a fixed number — the gray zone sits in COGS vs. OPEX treatment of support, token spend, and solutions-engineer comp — and walks a $50M-ARR worked example where reclassifying the same $13M of cost moves reported gross margin from 74% to 88% to a "defensible" 80%, without changing EBITDA by a dollar.
Why this is in the vault
A concrete, numbers-first framework for exactly the question RDCO's own agent-cost-attribution thinking has circled without resolving: when AI token spend serves a customer vs. serves internal automation vs. seeds free-trial acquisition, which P&L line does it belong on, and how much does that choice alone move the reported margin story.
Mapping against Ray Data Co
This is the classification layer the two already-filed token-cost pieces assumed but never fully diagrammed — [[2026-09-29-mostlymetrics-token-cost-budgeting-frontier-convergence]] forecasts token spend per head, and [[2026-09-01-mostlymetrics-audit-ai-bill-for-savings]] audits where the bill leaks, but neither asks whether that spend sits in COGS or OPEX once it's measured. CJ's own taxonomy — hosting/support/data are "black and white" COGS; token costs, customer success, freemium usage, and model-development spend are the "grayish" zone where a Controller has to draw a line — is the literal checklist Scribble Works' own cost structure would need to run if RDCO ever reports unit economics on it: the AI Gateway token spend behind the scribble-works engine rail ([[reference_scribble_works_engine_rail]]) is customer-serving (COGS) when it generates a kid's art, but any token spend on internal tooling, content QA, or the headless-Chrome PDF-gate pipeline is R&D/G&A by this same framework — a distinction nobody at RDCO has written down yet. Per [[project_credibility_for_phdata_sales]], the founder wants credibility content that demonstrates agents-in-production domain depth rather than selling services; a worked "where does the agent-cost line go" explainer, grounded in a real product's cost stack, is a more defensible artifact than anything hypothetical, and this issue hands over both the taxonomy and a template worked example to build from.
The core argument
Gross margin is a valuation decision disguised as a calculation: GAAP leaves latitude between "cost to deliver" and "cost to run the company," and that latitude is where Controllers do real judgment work — a concrete, dollars-denominated instance of the [[operational-definitions]] problem (a metric isn't defined until the criterion, test, and threshold are written down, and GM's criterion is exactly what drifts here). CFOs care because GM gates how aggressively they can spend (a dollar in S&M reads better than the same dollar in COGS). Investors care because the percentage — not the dollar — signals scalability and sorts a company into a multiple band; true software runs 80%+ gross margin, usage/infra businesses run 50-75%, and counterintuitively too high a margin on an AI-era company can read as "no AI here," hurting the story. Acquirers run their own quality-of-earnings rebuild of a target's GM, loading in costs the seller excluded, partly as negotiating leverage and partly to under-promise their own investors. The worked example: a $50M-ARR company with $13M in COGS-eligible cost can report 74% margin (fully loaded), 88% margin (hosting-only), or a middle "defensible" 80% by reclassifying dev-environment hosting into R&D, internal-automation token spend into G&A, free-trial token spend into S&M, and half of quota-carrying solutions-engineer comp into S&M — all without changing a single dollar spent or EBITDA. CJ's prescription: write down a one-page COGS definition the whole finance team agrees to, tag token spend by purpose now before volume makes retroactive tagging impossible, and split software margin from services margin if the business carries both.
⚠️ Sponsorship
Numeric (financial-data-platform / "agentic accounting" vendor) is the sole disclosed sponsor this issue — the newsletter's banner tagline ("proudly powered by Numeric") plus a dedicated promo block pitching their "Architecture for Agentic Accounting" white paper (data-centric accounting, journal entries as one view rather than the system of record). Not part of the Quote-to-Cash series and not one of the previously-seen rotating sponsors (Brex, Abacum, Koyfin) for this sender — first time Numeric appears as lead sponsor in vault-filed issues to date. The promo doesn't bend the editorial argument (gross margin classification), but Numeric's pitch (a unified financial data layer) is thematically adjacent to this issue's "where does the line go" argument, which is worth naming as a soft alignment even though the body content reads as independently argued.
Related
- [[2026-09-29-mostlymetrics-token-cost-budgeting-frontier-convergence]]
- [[2026-09-01-mostlymetrics-audit-ai-bill-for-savings]]
- [[2026-08-29-agent-cost-attribution-stack]]
- [[project_credibility_for_phdata_sales]]
- [[reference_scribble_works_engine_rail]]
- [[feedback_api_cost_budget_controlled]]