"How to talk about ARR for Consumption Based Businesses" — CJ Gustafson

Why this is in the vault

A CFO Mailbag where two operating CFOs (Mercury, Notion) field reader questions on consumption-based ARR, AI-cost gross margin, and CAC payback — the exact usage-billing terrain RDCO's founder works in via Snowflake.

Issue contents

This is a "CFO Mailbag" format: CJ relays five reader questions to two guest CFOs, Daniel Kang (CFO @ Mercury) and Rama Katkar (CFO @ Notion). The five questions:

1. What belongs in gross margin once AI costs are large — reader's COGS-vs-opex line is contentious internally (hosting is clear, but CS headcount, free-trial infra, solutions engineering, and "a massive Anthropic bill" each move the number 8-10 points depending on definition).
2. How to talk about consumption-based "ARR" — reader is usage-based; their "ARR" is really a 30-day run-rate annualized, so it bounces with seasonality and ramp, and SaaS-minded investors keep pattern-matching it to subscription benchmarks. Is there a cleaner metric for consumption businesses?
3. Dissecting a rising CAC payback period.
4. What metrics belong on an exec dashboard.
5. Board communication during a hard pivot to AI.

Paywall note: the free email + public web version both cut off after the Q2 question — the CFO answers to Q2 through Q5 are premium-subscriber only. Only Q1's answers render in full.

The core argument

On the one fully-readable answer (Q1, gross-margin definition), both CFOs converge on the same discipline:

• Dan (Mercury): Define COGS to be defensible as a public company under GAAP, do the technical-accounting work to support COGS-vs-opex calls, and get auditor sign-off so you never have to restate. Crucially: when computing LTV, fully burden variable costs regardless of whether they sit in COGS or opex — don't let the classification fool you about true product economics. And diagnose why the debate exists (academic argument vs. investor-optics vs. people whose bonuses key off gross profit). Sophisticated investors see through inflated gross margins anyway.
• Rama (Notion): Align on a strawman definition with auditors, report it consistently, and refuse to re-litigate when the number moves the "wrong" way. A one-time pro-forma can be shown internally for genuinely abnormal costs, but the gray-area items named (CS, trial infra, solutions engineering, model bills) are ongoing variable costs — so you manage the rest of opex around them rather than reclassifying to flatter the metric.

The Q2 framing itself is the load-bearing teaser for RDCO: usage-based revenue annualized from a trailing window is structurally different from contracted subscription ARR, and forcing it into a SaaS ARR comp misleads everyone. (The CFOs' specific "cleaner metric" recommendations are paywalled.)

Mapping against Ray Data Co

Strong. Three distinct hooks:

• Consumption-ARR question rhymes directly with Snowflake/usage-based data tooling. The founder's day-job orbit (phData, Snowflake) is the canonical consumption-billing world — credits consumed, not seats subscribed. The reader's pain (a "30-day run-rate annualized" that bounces with ramp and seasonality, mispriced against SaaS comps) is the exact analytical problem of measuring a Snowflake-style usage business. This pairs with the vault's existing usage-vs-subscription pricing and revenue-hierarchy notes — RDCO already has a thesis seam here.
• "Massive Anthropic bill" in gross margin maps to RDCO's own token economics. RDCO is a company whose variable cost is model spend (see the token-budget-as-employee-cost note). The Q1 answer — fully burden variable model cost into LTV regardless of COGS/opex classification, don't fool yourself on product economics — is the directly-applicable discipline if/when RDCO surfaces (Squarely, MAC, Sanity Check) ever carry their AI cost-of-goods as a real margin line.
• CFO-voice study for Sanity Check. The Mailbag format is a clean specimen of how operating CFOs actually talk — defensibility, auditor strawmen, refusing to relitigate, "don't fool yourself on product economics." That register is reusable raw material for a Sanity Check piece on metric-gaming, but only as an original re-frame (consumption-ARR-as-honesty-test), not a restatement of CJ's mailbag.

No project-changing threshold crossed and no new sponsor relationship; logging as status-only.

Related

• [[2026-05-11-mostlymetrics-usage-vs-subscription-pricing]] — CJ's prior take on when usage-based billing actually fits; direct predecessor to this issue's Q2.
• [[2026-05-03-mostlymetrics-revenue-hierarchy-is-it-cake]] — revenue-quality hierarchy; the "what counts as real ARR" lens this Mailbag's Q2 leans on.
• [[2026-05-10-mostlymetrics-token-budget-as-employee-cost]] — model spend as a first-class cost line; pairs with Q1's "massive Anthropic bill" in gross margin.
• [[2026-05-11-mostlymetrics-cac-payback-calculation]] — CJ's standalone CAC-payback method; same metric this issue's Q3 dissects.
• [[2026-05-31-mostly-metrics-ndr-net-dollar-retention-decline]] — adjacent disclosure-honesty theme (hiding a metric is itself a signal).