AI Profit Matrix
The AI Profit Matrix is a two-by-two framework for AI unit economics, developed at Potio. It plots consumption of a product’s unit of delivery against gross profit per unit of consumption, and sorts any AI product or feature into one of four quadrants: Dead Zone, Token Furnace, Boutique or Flywheel.
The two underlying metrics come from Kyle Poyar. Consumption is whatever the product ships: tokens, audio hours, API calls, images generated. Gross profit per unit is profit in currency on each of those units, not a margin percentage.
| Quadrant | Consumption | Gross profit per unit | What to do |
|---|---|---|---|
| Dead Zone | Low | Low | Reposition the feature or switch it off |
| Token Furnace | High | Low or negative | Fix unit economics before scaling further |
| Boutique | Low | High | Spend margin on buying consumption |
| Flywheel | High | High | Protect the loop that got you there |
Replit shows why revenue alone misleads. Its AI agents took ARR from $300M to $525M during 2025 while gross margin went from 36% to negative 14%. On an ARR chart that is a winner. On the matrix it sits deep in the Token Furnace.
Potio’s take: An AI company reporting success only in ARR is flying with half the dashboard dark. Two companies at 55% gross margin can be heading in opposite directions, one with per-unit profit doubling and the other with it halving. The matrix makes that direction visible, and nobody drifts into the Flywheel by accident.
How this plays out
Each quadrant needs a different fix, and the Token Furnace needs the most work. Serge recommends three fixes for a furnace, in this order:
- Charge for the unit that drives your cost. If every request costs you tokens but customers pay per seat, a heavy user costs more without paying more. Moving the price onto the same unit as the cost, such as tokens, runs or credits, means revenue rises when cost rises.
- Lower the cost of each completed task. Send simple requests to smaller, cheaper models and cache repeated answers. Measure the cost of the whole task rather than a single model call, because a cheap model that needs four retries costs more than a pricier one that gets it right first time.
- Put limits on heavy usage. Usage caps, monthly quotas or tiers priced by effort stop a handful of power users from driving the whole compute bill.
A Boutique has the opposite problem. Each unit is profitable, but too few people use it, and pricing will not fix that. Product changes will: make the output easy to share so one user’s work brings in the next user, and remove the caps and approval steps that teach customers to ration.
Raising prices on a furnace often backfires. Higher prices suppress usage, so the product ends up with a better margin on a smaller base. That is a move left across the matrix, into the Boutique, rather than up.
How do you calculate gross profit per unit for an AI product?
Take revenue attributable to the unit and subtract the full cost of delivering it: inference, GPU hosting, embeddings and retrieval, observability, data egress and any support headcount the feature created. Without per-request logs that tie model, token count and customer together, you get a total you cannot break down.
Is a low gross margin always bad for an AI company?
No. Cursor and Lovable both run gross margins in the 50% to 60% range and still compound, because gross profit dollars grow faster than costs. Direction beats level: a 55% margin improving every quarter is healthier than a 70% margin decaying.
More on this: AI Unit Economics: How to Escape the Token Furnace
Pricing consulting for
SaaS and AI companies.
I'm a 3x founder and former CEO of Toggl. I work hands-on with SaaS & AI teams to fix pricing, packaging and monetization.
Book a call