Gross Margin

What belongs in cost of revenue

For a subscription business, cost of revenue is everything required to keep the service running for customers who are already paying: compute, storage and bandwidth; monitoring, logging and security tooling that scales with usage; third-party APIs, data providers and model inference billed per call; payment processing and its fixed per-transaction fees; the support organisation; and the customer success headcount that serves existing accounts rather than closing new ones.

What does not belong: product engineering, general and administrative overhead, and sales and marketing — those sit below the gross-profit line and belong in CAC or operating expenses.

The judgement calls that move the number most

Customer success placement is the largest single lever, and as the worked example shows it can move reported margin by eight points or more without any change in spending. The usable test is what the team is paid to do: onboarding and adoption work that keeps existing customers alive is a cost of delivery; renewals and upsell work is closer to sales. Split the team's cost if it genuinely does both, decide once, and restate history when you change the rule.

Free tiers and trials consume real infrastructure. That cost is either an acquisition cost or a cost of revenue depending on how you treat non-paying users, but it is never zero, and a generous free tier can quietly take several points off margin.

Professional services — implementation, migration, bespoke work — usually run at far lower margin than software, sometimes near break-even. Blending them into one figure disguises both lines. Report subscription gross margin and services gross margin separately, and a blended figure only alongside them.

Why margin is the multiplier on everything else

Gross margin is not a standalone metric so much as a coefficient inside the others. LTV is ARPA over churn multiplied by margin. CAC payback is CAC over monthly revenue divided by margin. So a margin figure that is wrong by ten points carries a proportional error straight into both — and into the LTV:CAC ratio built on top of them. A company that misclassifies its support costs does not just report a flattering margin; it reports a flattering version of its entire unit economics, and it will over-invest in acquisition on the strength of it.

What moves gross margin over time

The traditional trajectory is upward: infrastructure gets cheaper per unit at scale, committed-use discounts land, support gets partly automated and fixed costs spread across a larger base.

Two forces push the other way, and both are worth watching monthly rather than annually. Mix shift toward usage-heavy customers raises variable cost faster than revenue when pricing is seat-based rather than usage-based. And AI features carry genuine marginal cost — inference is billed per token and does not amortise the way a database query does, so a product that adds a generous AI capability to an existing flat-rate plan converts a fixed cost base into a variable one. This is a structural change to subscription margins rather than a temporary expense, and it should be modelled per plan rather than absorbed into an overall figure.

Reporting it so it stays comparable

Define cost of revenue once, write the definition down, and restate history whenever you change it. Track margin by plan and by segment alongside the headline, because a blended margin hides the plan that is losing money on infrastructure. And check it against ARPU: at low price points, fixed per-transaction payment fees consume a share of revenue that becomes invisible in a blended percentage.