You shipped the AI feature in March. Flat twenty dollars a seat, same as the rest of your plan, because that's how software has always been priced. Three months later your MRR is up and to the right, and so is the line you don't check as often: the invoice from whoever hosts your model. The two lines are climbing together, and nobody told you they were connected.
For fifteen years, software founders inherited one beautiful fact: the marginal cost of one more user was basically zero. You could add a thousand customers overnight and your server bill barely twitched. That fact is what made 80% gross margins normal, and it quietly shaped every pricing instinct you have. The moment your product started calling someone else's model, the fact stopped being true. You have a cost of goods sold now, for the first time, and most founders are still pricing as if they don't.
The margin you inherited belongs to a different business
The 2026 numbers are blunt about it. ICONIQ's data puts the average AI product gross margin around 52%, against the 80% benchmark that defined classic SaaS. Inference alone eats roughly 23% of revenue at scaling-stage AI companies, which means that for every million dollars of AI revenue, something like $230,000 walks out the door as raw compute before you've paid for a single other thing. That isn't a cost you optimize away in a later sprint. It's the shape of the business.
Why the average is the dangerous number
The 52% blended figure hides the part that actually hurts. Costs in an AI product don't spread evenly across your users, they pool around the heavy ones. The customer you're proudest of, the one living in the product all day and telling their friends about it, is also the one firing a hundred model calls where the average user fires five. On flat per-seat pricing, that power user can cost more to serve than they pay you, and they never surface as a problem anywhere. They surface as engagement, which is the one number you've been trained to celebrate.
| The old assumption | The new reality |
|---|---|
| One more user costs about zero | One more heavy user can cost more than they pay |
| Usage is something to maximize | Usage is something to maximize and meter |
| Margin is roughly uniform across users | Margin swings wildly from one account to the next |
| Price the seat | Price the seat plus what it consumes |
⚠️ The power-user trap: On flat pricing, your most engaged customers are often your least profitable ones, and every dashboard you own rewards you for going out and acquiring more of them.
Your growth dashboard is the wrong instrument
MRR climbing feels like proof the pricing works. It isn't, because MRR can't see cost. A revenue line going up while blended margin slides down is the exact pattern that ends with a founder discovering, a few thousand users too late, that the unit they've been scaling loses money at the unit level. Cursor, GitHub Copilot, and Devin all restructured their pricing in 2026 for this reason: per-seat plans that worked fine until autonomous agent sessions started burning multi-hour compute, and the cost side quietly detached from the revenue side.
Watch cost-to-serve, not just revenue
The instrument you actually need is cost-to-serve per account, and almost nobody builds it early, because classic SaaS never made you. Pull your model spend, attribute it down to the individual account, and set it beside what each account pays. The first time you run it, the result is usually uncomfortable: a handful of accounts are underwater, a long tail is wildly profitable, and the blended number you'd been quoting yourself was politely hiding both facts.
💡 Do this once: Take last month's model bill, split it across accounts by actual usage, and sort the list by margin. The shape of that sorted list tells you more about your pricing than your MRR chart ever has.
Price the heaviest user, not the average
The fix isn't to fear usage or throttle the people who love the product most. It's to make sure the heaviest plausible user is still profitable, and to stop pretending a flat seat covers an unbounded cost. In practice that means a hybrid: a subscription for the predictable base, plus usage-based pricing above a threshold, so the accounts generating the most cost are also the ones generating the most revenue. The whole industry drifted toward this shape in 2026 for one unglamorous reason. It's the only structure where your best customer and your margin point in the same direction.
| Pricing model | What it protects | Where it breaks |
|---|---|---|
| Flat per-seat | Simplicity, easy to sell | Power users on heavy models go underwater |
| Pure usage-based | Margin on every call | Unpredictable bills scare buyers off at signup |
| Hybrid (base plus metered) | Margin and predictability together | More work to design, more to explain |
🔑 The takeaway: The day you added a model call, you changed what kind of company you are. Software pricing assumed a cost that's no longer zero, and the founders who notice early get to reprice on their own terms instead of having the heavy users teach them.
None of this means AI products are bad businesses. It means they're a different business than the one your pricing instincts were trained on, with a real variable cost sitting in the middle of every transaction. Find your margin floor while you still have a few thousand users instead of a few hundred thousand. The math doesn't change with scale. It just gets a lot more expensive to fix.