For two decades, the zero-marginal-cost assumption made us lazy. It convinced us that if we just built the software, the margins would engineer themselves. This dynamic produced 80%+ gross margins and justified why, at the peak, some companies traded at 20x revenue.
But AI breaks that assumption.
SaaS is a gym membership: pay once, visit forever. AI is a factory: every unit of output requires raw materials, energy, and labor. You don't scale a factory by just adding "users"; you scale it by managing yield. Every additional unit of usage consumes something real: tokens, inference, latency, and watts.
Consider an AI product priced at $100 per user/month:
A light user generates $5/month in inference cost. A power user generates $65/month. At small scale, the averages hide the problem. But when the top 10% of customers drive 80% of your compute costs, your unit economics collapse, even though your pricing looks "SaaS-like".
This creates hard truths for the AI founder:
Usage-based pricing isn’t optional: Flat per-seat pricing acts as a subsidy for your most expensive users.
Wrappers are Resellers: If your differentiation is just a UI on top of Anthropic, you aren't a software company. You are a low-margin services firm paying a "compute tax" to a vendor who is also your biggest competitor.
Gross margins matter again: Fast-growing AI apps are often operating at 25–40% gross margins, not 80%.
The Valuation Trap: SaaS multiples assume software economics. Do not raise at a $100M+ valuation if you have manufacturing-style COGS.
When you raise at SaaS multiples without SaaS margins, you are forced to "grow into" a valuation that physics won't allow. The result isn't just a missed target; it’s a consequential liquidation preference stack that can erase the common stock even in a "successful" exit.
The Bottom Line: In the SaaS era, margins were a gift of software physics. In the AI era, margins are a discipline you have to engineer.
But AI breaks that assumption.
SaaS is a gym membership: pay once, visit forever. AI is a factory: every unit of output requires raw materials, energy, and labor. You don't scale a factory by just adding "users"; you scale it by managing yield. Every additional unit of usage consumes something real: tokens, inference, latency, and watts.
Consider an AI product priced at $100 per user/month:
A light user generates $5/month in inference cost. A power user generates $65/month. At small scale, the averages hide the problem. But when the top 10% of customers drive 80% of your compute costs, your unit economics collapse, even though your pricing looks "SaaS-like".
This creates hard truths for the AI founder:
Usage-based pricing isn’t optional: Flat per-seat pricing acts as a subsidy for your most expensive users.
Wrappers are Resellers: If your differentiation is just a UI on top of Anthropic, you aren't a software company. You are a low-margin services firm paying a "compute tax" to a vendor who is also your biggest competitor.
Gross margins matter again: Fast-growing AI apps are often operating at 25–40% gross margins, not 80%.
The Valuation Trap: SaaS multiples assume software economics. Do not raise at a $100M+ valuation if you have manufacturing-style COGS.
When you raise at SaaS multiples without SaaS margins, you are forced to "grow into" a valuation that physics won't allow. The result isn't just a missed target; it’s a consequential liquidation preference stack that can erase the common stock even in a "successful" exit.
The Bottom Line: In the SaaS era, margins were a gift of software physics. In the AI era, margins are a discipline you have to engineer.