Your Heaviest User Is Your Thinnest Margin

A founder I work with sells an AI writing tool, flat rate, $40 a seat a month. Last month one of his marquee accounts went from a handful of sessions a week to running the product all day. He was thrilled. Real adoption, the kind every product deck promises. I asked what that account costs him to serve now. He didn't know. We pulled it. At their current usage, the seats they pay $40 for cost him $52 in model calls. His best logo was running at negative gross margin.

He'd built a SaaS pricing model on top of a cost base that doesn't behave like SaaS.

Engagement used to be free. Now it has a meter.

For twenty years the SaaS rule held: once you've built the product, serving one more session costs close to nothing. Usage was pure upside. A customer who logged in every day was your best account, full stop, because the price was flat and the marginal cost was a rounding error. Drive engagement, win.

AI breaks that rule at the cost line. Every prompt is a model call, and a model call costs real money that scales with how much the customer uses. The math making the rounds this year is blunt. A power user who grows prompt size and frequency can push token cost from $25 to $40 on a $100 account, dropping gross margin from 75% to 60% with no change to their plan. The heavier the use, the thinner the margin. That's the exact customer your old playbook told you to celebrate.

Your blended margin is hiding the spread

Your P&L won't show you this directly. Blended gross margin looks fine because light users subsidize heavy ones. Most accounts barely touch the product, their cost to serve is near zero, and they paper over the handful running negative. The average reads like a healthy 68%. Underneath it you've got one cohort above 90% and another below zero, and you're managing to a number that describes neither.

Two customers on the same plan, paying you the same dollar, can cost wildly different amounts to serve. Flat pricing on a metered cost means your light users are quietly funding your power users. That holds until the power users outgrow them.

The risk isn't today's blended number. It's the direction it's moving. As the product gets better and adoption deepens, the heavy cohort grows, and success drags the average the wrong way. You win the engagement battle and lose the margin war. It shows up as a slow bleed nobody can pin to one cause.

Compute margin per account, then act

Pull cost to serve by account for a single month. Model calls, inference, whatever your AI vendor meters, divided down to the customer. The underwater cohort surfaces fast, and it's usually a small group of heavy users doing most of the damage.

Then you have a few moves, and you don't need all of them. Cache aggressively, since cached input tokens run roughly 90% cheaper and a stable system prompt can cut per-query cost by an order of magnitude. Add a usage component above a threshold so the heaviest accounts pay for what they burn. Or cap the flat plan and route true power users to a tier that prices the cost back in.

What you can't do is run an all-you-can-eat plan on a bill that grows every time someone eats. That math only works while nobody's hungry.