The consensus on moats is wrong about AI products.

We hear that AI has no moat. And anyone can copy your ChatGPT wrapper.

But here's what they're missing!

Traditional software is deterministic. You can reverse-engineer features and rebuild them exactly. That's why, in the era of software, distribution has been a key moat.

But what about AI products? They're probabilistic in nature, it's hard to reverse engineer them by looking at output. The final product depends on hundreds of decisions

- How you cleaned your data

- Which edge cases you prioritized

- The specific prompts + policy you refined over months

- The proprietary data/ insights from the user feedback loops

- Latency

- Evals

- Tooling, and many more decisions

And this is where startups like Anthropic beat big-tech benchmarks on coding agents. They made thousands of small decisions that compound into something unique - and it's hard to reverse engineer all of these things by looking from outside.

The moat isn't the code. It's the speed of iterations, insights, and processes - all of which are tacit and path dependent. Use all this to build a better product and lock distribution, and by the time others catch up, you will have good moat built around various things.