The margin math of an AI feature at scale

Software people are used to margins that improve with scale. You build the thing once, and each new user costs almost nothing to serve, so margin widens as you grow. AI features break that habit. Every request costs real money in inference, so the cost line grows with every user you add. If your revenue per user does not grow at least as fast, your margin shrinks exactly as you succeed. Growth stops being pure upside and becomes something you have to price against.

Cost that grows with every user

A founder building an AI-native business intelligence platform put the shift plainly. In classic software, he said, your best day is when a million people show up, because they cost you almost nothing. With an AI product, a million people showing up is a bill. Each one of them is running inference, and inference is a variable cost that never goes to zero the way storage and bandwidth effectively did. So the question he had to answer, that he never had to answer before, was whether each user paid for the compute they consumed.

The math gets sharper inside agentic features. An agentic workflow can burn ten to thirty times the tokens of a single call per user intent, because it reasons in steps, calls tools, and retries. Run the arithmetic on a modest five-step agent at ten thousand runs a day and two cents a run, and you are at roughly $365,000 a year for one workflow. Output tokens make it worse, costing around five times what input tokens cost, so a chatty feature carries a heavier variable cost than a terse one. None of this shows up at pilot scale. All of it shows up when the feature works and usage climbs.

An engineering leader who scaled large consumer systems framed the danger as a mismatch of curves. His cost curve rose with usage in a straight, predictable line. His revenue curve did not, because a flat subscription does not care how many times a user hits the model. The gap between those two curves was his margin, and it narrowed every time an enthusiastic user leaned on the feature harder. His heaviest users were his least profitable, which is the opposite of how his business had ever worked.

Why the trap is invisible early

At small scale the variable cost is a rounding error, so margin looks fine and nobody models the slope. The problem is that margin at a thousand users tells you almost nothing about margin at a million, because the cost line has a slope the old software line did not. You cannot extrapolate from the comfortable early number. You have to project the cost curve against real growth and see where it crosses your revenue.

What to model before you scale

To keep an AI feature solvent as it grows, watch these:

Variable cost per user. The inference a typical user actually consumes, not the average across dormant accounts.
The heavy-user tail. What your most active users cost, since they can be unprofitable while looking like your best customers.
Revenue slope. Whether your pricing grows with usage or sits flat while cost climbs underneath it.
Agentic depth. How many model calls one user action triggers, because that is the real multiplier on the bill.

How we approach it at Density Labs

Our AI Opportunity Assessment is a fixed two-week engagement, priced at $2,500. When margin at scale is the worry, we model the variable cost per user against the revenue that user brings, then project both curves out to real volume. That is where you find out whether growth helps you or quietly sinks you. It is a far cheaper thing to learn on a whiteboard than on a production invoice.

Traditional software rewards scale. AI features charge for it. The teams that stay profitable are the ones that mapped where their cost curve crosses their revenue curve before they got there.