Jack Flintoft

For the past 20 years, SaaS operated under the economics of high fixed cost + low marginal cost. When making a software product, you would sink massive swaths of capital into engineering in order to get your v1 out the door․ On the other side‚ after you had built your product‚ your marginal costs (the cost of serving the next additional customer) were near zero (hosting was cheap‚ and actually the major cost was finding customers to acquire)․

These old SaaS-onomics seem orthogonal to AI native economics now․ Since AI inference isn’t free (and in the case of AI-native companies‚ every customer is a direct‚ recurring cost to your model provider)‚ startups are being forced to think on the margin․ My Econ professor John List would be beaming at this — to him‚ the margin is all that matters․ Marginal thinking is about the next unit‚ not averages․ You shouldn’t ask whether 10 hours studying is good for your final exam but whether the eleventh hour gets you more value than an extra hour’s sleep․

This thinking is precisely why outcomes-based pricing is winning against per-seat licensing in startup-land (we must only look towards Fin or Sierra’s success here to see this thesis playing out successfully)․ Rather than lazily charging an enterprise $50 a month for a support seat that might sit idle‚ Sierra charges per resolved customer ticket․ Any time their AI agent really solves a problem‚ Sierra incurs the cost of inference‚ charges the business to have solved the problem‚ and pockets the direct margin on that one transaction^[1]․ Hallelujah, the unit economics finally align! (it’s the same reason that you pay lawyers when they win your case‚ not when they open one)

But the more interesting second-order effect is what this pressure is doing to the businesses themselves.

The old SaaS model (for the most part) rewarded you for being horizontally scaled and for providing thin value․ You could build a wide‚ shallow product‚ sell thousands of seats‚ and the math worked because your marginal cost was almost zero․ AI inference costs borne by the startup make that model brutal․ When inference costs eat into every query, wide and shallow products quickly become money-losing ones.

What survives then‚ is something closer to a Darwinian natural selection․ The businesses that will survive will be those that go deep into a workflow and produce real economic impact‚ for the inference bill in the AI native era doesn’t forgive vague ROI․

This is great news for customers․ Enterprise software for a long while could afford to be sticky without being great. Salesforce never had to think on the margin for existing accounts. Instead, they could collect millions of dollars off thousands of empty seat licenses lying around the enterprise․ That was easier than ripping out the software․ AI-native competitors‚ with their cost structure‚ just can’t afford that idle-ness․ They have to earn the margin․

And as it turns out‚ the inference bill might be the best thing that ever happened to software customers …

---

1. 0ver simplification‚ but to state the point