Overview
Software became the most profitable business model in history for one reason: once built, serving each additional user cost almost nothing. Gross margins of 80–90% were normal. AI quietly breaks this and that has big consequences for software economics.
The old magic: near-zero marginal cost
Classic SaaS spends a lot to build the product, then serves millions of users at trivial incremental cost (a bit of storage and bandwidth). That's why software margins dwarfed physical businesses revenue scaled while cost of goods barely moved. It was, effectively, infinite leverage.
What AI changes
Every AI-powered action a generation, a summary, an agent run consumes real compute that costs real money, every single time. AI features carry a recurring cost of goods sold that scales with usage. A heavily-used AI feature can quietly turn a high-margin product into a low-margin one. The more successful the feature, the bigger the bill.
The flat-fee trap
Many products still charge a flat monthly fee while their AI usage costs vary wildly per customer. Power users can consume far more compute than they pay for, making your best-engaged accounts your least profitable. This mismatch is a quiet margin killer.
How software adapts
Survivors do two things: re-architect pricing toward usage-based or tiered models that tie revenue to AI cost, and engineer inference costs down (caching, smaller/routed models, prompt efficiency). The discipline that hardware and services businesses always had watching cost of goods is returning to software.
What this means for you
If you build software: track gross margin per feature and per customer, align pricing to AI cost, and optimize inference. Don't assume software's traditional 80%+ margins still hold once AI is in the loop. If you invest: scrutinize AI-era gross margins, not just revenue growth.
Honest limits
Falling model prices partly offset this, and not all software is AI-heavy. But the structural point stands: AI reintroduces meaningful marginal cost to software, ending the era when margins could be ignored.
