SaaS Apocalypse

The mechanism, as Horowitz states it

The framework holds that AI coding tools collapse the build-versus-buy calculus for software, putting structural downward pressure on SaaS pricing and terminal value. Ben Horowitz supplies the term and the mechanism.¹ The lock-ins that justified premium SaaS multiples, he argues, are dissolving on three fronts. Migration pain shrinks because AI can reproduce a competitor's product at machine speed, turning years of work into weeks. Data lock-in weakens because clean exports plus AI-assisted migration remove the friction of switching. And UI lock-in fades because future "users" are increasingly agents, and agents are flexible about how they operate an interface. When all three erode together, Horowitz's claim is that the market re-prices the terminal value of undifferentiated software downward. The framework develops directly out of no moat in software.

The operator's playbook, as Petersen runs it

Ryan Petersen turns the thesis into a procurement strategy at Flexport.² Selling SaaS to tech-forward buyers, he argues, gets harder: "the negotiation we're going to have with Salesforce is going to be a lot different than the last one... we can build stuff for ourselves," and he already replaces some SaaS in-house. His central tactic is a credible-threat shakedown. He has the procurement team build a case study of each SaaS they did replace and how long it took, then walks the vendor list asking each to cut rates or be, in his phrase, vibe-replaced. His estimate is roughly 20% out of almost everybody, and he concedes it is partly a bluff because he cannot and will not replace them all. As a proof point he cites a health-insurance founder who replaced $600k a year of Salesforce in three weeks with an in-built tool.

The exception, in Petersen's telling, is stickiness rather than features. Slack survives because nobody wants to build their own Slack, yet Flexport still incubated an AI-first Slack competitor aimed at green-field markets with large net-new customer creation each year, where there is no incumbent to rip out. His own brake on the framework is maintenance: he does not want to spend engineering effort maintaining a sprawl of in-house SaaS clones, so the realistic outcome is selective replacement paired with broad price extraction through the threat.

The counterweight

A second reading on the Cheeky Pint podcast frames the outcome as a re-pricing of undifferentiated SaaS rather than an extinction.³ Telling a model to clone Ramp, on this account, yields a skin-deep shell because there are, in the phrase used, nine million edge cases the clone misses. The survivors line up with the no moat in software list: buyers who are hostages rather than customers of deeply embedded systems of record, products defended by network effects, and businesses defended by a proprietary data moat. Rational targets are features that became companies; systems of record resist longer. This sharpens rather than refutes Horowitz and Petersen: the apocalypse hits the undifferentiated layer, and the live question becomes what invisible depth or data an incumbent has that an agent cannot reproduce.

Open questions

The 20% figure is asserted, not measured, and the threat only works while it stays credible, which erodes if maintenance costs bite. The counterweight can over-rotate the other way, since claiming edge cases rationalizes any incumbent; the honest test is whether those edge cases are load-bearing or merely accreted. And the proof points carry selection bias, since a CRM replaced in three weeks is a relatively replaceable category, while deeply integrated systems of record may resist far longer.