The New AI P&L
Unit economics in a token-based world
CAC to LTV.
For a generation of SaaS founders, unit economics meant one thing: CAC to LTV.
Acquire a customer for less than they’ll pay you over time. Win.
That framing really worked, and for lots of companies, still works great today. But it relies on a certain understanding of costs.
Traditional SaaS had a beautiful scaling property: expensive to build, nearly free to distribute. You invested heavily upfront—engineering, design, infrastructure—then served each additional customer for almost nothing.
Gross margin was assumed to be 80%+ and safely ignored. Early stage startups never built elaborate models tracking the cost of serving power users versus casual ones. It didn’t matter enough.
From Users to Features
AI shifts the unit of analysis. The question isn’t just “can I acquire this customer profitably?” It’s now “can I deliver this feature profitably?”
Every AI-powered interaction has a cost. For some simple writing feature (“draft an email”) - it’s still functionally zero. Older models work fine and are good enough.
But as we start building AI-first interfaces, that bring long chains of thought and high token counts from inference-hungry models, things change.
For the first time since the Saas revolution, cost matters.
Suddenly the economics are granular in a way SaaS founders never had to think about.
The old P&L: acquire users, assume margin. The new P&L: acquire users, then win or lose on every feature they touch.
The Question Every Investor Should Ask
For your highest-usage AI features, what is your contribution margin?
Every founder should know the answer to this question cold.
Not gross margin. Not blended average across the product.
The specific margin on the specific features that drive retention and willingness-to-pay.
Most founders can’t answer this.
They usually know overall COGS. They know what they’re paying OpenAI or Anthropic in aggregate. But they haven’t mapped costs to features, or features to value.
And it’s not hard to do! Something as simple as distinct API keys for distinct features achieves this, and the AI APIs typically provide good financial reporting on a key-by-key basis.
The Value Exchange
This is the new table stakes.
If you don’t know which AI features are profitable and which are subsidized, you don’t yet understand your business.
Investors should push here too. Revenue growth without contribution margin clarity is a warning sign. The question isn’t “are they growing?”—it’s “as usage scales, does margin expand or compress?”
Companies that can’t answer are losing money on every unit, and planning to make it up in volume.
This pushes founders toward a harder question than “how do I get more users?” The question becomes: is the value exchange high enough on each interaction to pencil out?
A feature that users love but costs more to deliver than they’ll pay for isn’t an asset. It’s a leak. Growth makes it worse.
Originally published on Griff’s Substack on 27 February 2026.