3-minute read

The worst software businesses are smart. The best ones are dumb.

Let me explain.

There’s been ongoing debate of where value accrues in the AI era. Many believe it accrues to the app layer!

This has felt intuitive. Applications own the customer relationship, not AI models. In each application layer market, prices form in negotiations between application sellers and buyers with AI token COGS largely hidden. This follows standard practice: intelligence has typically been application layer special sauce and part of the service that customers pay for.

The trouble is that even as AI model costs have gone down, with agent frameworks like Claude Code, token consumption has gone up! This is Jevon’s Paradox in action (see @iamgingertrash’s application of it): even as inference costs have gone down, token consumption at the frontier of intelligence has gone up! [Notably, this Jevons Paradox is NOT the kind Aaron Levie imagined.]

This introduces a funny economic condition for application companies:

The more intelligence your application wields (i.e., the more tokens that your customers demand), the less money you’ll be able to make because the token costs will be so prohibitive.

This is likely why the Windsurf executive team jumped so abruptly to Google and China’s Moonshot’s Justin Wong wrote

I want to say: most Agent products are nothing without Claude. The fact that Windsurf was cut off by Claude further proves this point. In 2025, the upper limit of intelligence is still completely determined by the model.

It’s likely why Claude Code PMs Cat Wu and Boris Cherny are back at Anthropic after a 2 week stint at Cursor.

This puts intelligent application layer companies in dangerous territory, facing both

  • disintermediation risk: end-users can bring their own API keys and pay retail
  • margin squeeze: labs have superior margins and set the price ceiling

As a pure application layer startup, quite paradoxically, the smarter and more agentic your application is, the harder it is to make money.

This is why good software businesses are dumb.

ROIC to Smarts

The best dumb money businesses are those that provide a service where the provision of the service doesn’t scale with AI usage. Dumb businesses don’t price intelligence at all: their revenue is orthogonal to tokens burned.

Interestingly, this is contrary to past business practice where applications would routinely advertise how smart they are. In this new era though, where demand of bleeding edge models is unbounded, the smarter you are, the dumber your business is.

Smarts anywhere but the frontier will land in the danger zone of disintermediation risk and margin squeeze. At the same time, the smart frontier is always expanding: top models last only about 3 weeks at the frontier before being superseded.

But the most capable models are tremendous agents and per Jevons Paradox and Anthropic’s expanding revenue, it seems there’s no limit to the demand for intelligence.

For the rest of us, let’s get after that dumb money.