Alex Rampell on building defensible AI businesses: greenfield bingo, software-does-labor, and why the best entrepreneurs have revenge or redemption as their fuel

Summary

Alex Rampell, general partner at Andreessen Horowitz where he has focused on application-layer investing for ten years, lays out the three-category framework he presented to a16z's LPs for deploying capital into AI-era startups.

The Three Investment Categories

Category one: Greenfield Bingo. The core thesis is that incumbents like NetSuite, Workday, and Salesforce retain customers through lock-in, not loyalty. The opportunity is not to pry those hostages loose but to capture new companies before they default to legacy vendors. Rampell points to Mercury as the model: it never converted SVB customers until SVB collapsed. It simply won every new company that formed. AI-native ERP and vertical software plays follow the same logic, building better products for buyers who have no existing vendor relationship.

Category two: Software Does Labor. This is the category Rampell flags as most novel. Entire professions, trial attorneys, dental office receptionists, have never had purpose-built software because the addressable market looked too small to justify it. AI changes the unit economics. He cites Eve for trial attorneys and Tenor for dental receptionists as early examples. The analogy is Toast, which monetized restaurants not through software subscription fees alone but by bundling payroll and payment processing into an operating system restaurants actually needed. AI allows a similar move: tell a trial attorney the software handles cases they previously couldn't handle profitably, and they become a software buyer for the first time.

Category three: Walled Garden. Proprietary data creates durable defensibility even as underlying model quality converges. Rampell's illustration is Open Evidence, a ChatGPT-style interface trained on the full corpus of medical literature. His argument: a GPT-3.5-level model with every piece of medical knowledge ever recorded beats a hypothetical AGI with no medical data for any clinical query. A parallel example is Vilex, a roughly 25-year-old European data business that aggregated legal records across Spain and sold case-law access to firms like Wilson Sonsini. It now sells outcomes, not raw data, because it holds the only complete dataset.

The Defensibility Problem

The sharpest warning Rampell issues is about the speed asymmetry AI creates. The incumbent-versus-startup race has always been whether the startup gets distribution before the incumbent copies the innovation. His prior assumption was that incumbents normally win. With AI-assisted development, the copying cycle has compressed from three years to potentially three weeks, which raises the stakes on picking strategies, greenfield capture, labor substitution, or data moats, that are structurally hard to replicate regardless of how fast a competitor can ship code.

What Rampell Looks for in Founders

Across his career, which includes co-founding Affirm with Max Levchin in 2012 and an early investment in Mercury, Rampell has settled on five attributes: the ability to recruit talent into near-certain failure, fundraising credibility, customer acquisition under duress, deep knowledge of the history of one's own space, and what he calls the revenge or redemption driver. He uses Renaud Laplanche as the canonical example. Laplanche built LendingClub to public-company scale, was ousted by his board, walked away having made hundreds of millions of dollars, then founded Upgrade doing the same thing. Upgrade is now roughly ten times LendingClub's size by his account. The psychological fuel, not the financial return, is what he argues sustains founders when capital is gone and the company is failing. His framing: you need someone who would turn down a $500 million acquisition offer even when their personal take would exceed $100 million, because something other than money is driving the outcome they want.