- Published on
The Model Is Not the Moat
- Authors

- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · The Model Is Not the Moat
0:00 / 0:00
Watch this article · The Model Is Not the Moat
The argument fits in one sentence, and Eliot Horowitz puts it that way himself: the value in robotics will not belong to whoever owns the biggest closed model. The model is not the moat. Your data and your know-how are.
That is a strange thing to hear from someone selling software, which is why the rest of his post is worth reading as economics rather than enthusiasm.
What stays yours, and what does not
He draws the line narrowly. Three things are the moat and should stay proprietary: the data your machines collect, the know-how in your process, and the deployment you actually run. Everything else — the models, the weights, the coding system — should be open.
Your data
collected on your own machines
Your know-how
the process, not the model
Your deployment
where it runs and what it touches
Everything else
models, weights, tooling — open
The reason is not sentiment. The public robotics data everyone trains on is small and shared, so models trained on it converge — and the difference then comes from what your own machines see. That data is spread across millions of deployments in factories, farms and warehouses that no model company can reach. Which, he argues, is a case for infrastructure owned by the people who hold the data, not by the few who would like to.
Two objections, answered
Scale wins
the data is scattered, not concentrated
Open weights exist
without tooling and deployment, weights do little
He takes on the two that come up first. If scale wins, whoever has the most data trains the best model — except the data that matters is not concentrated anywhere, and cannot be bought into one place. And if open weights already exist, that is a start rather than an answer: the tooling, the data pipelines and the deployment layer have to be open too, or the weights do very little for the person who needs them.

Where the training data actually comes from
Video, mostly. There is an enormous amount of footage of people doing things, and a great deal of footage made specifically to teach a task, plus a growing body of research on turning that into robot training data. Early results are promising enough that research groups are already training robot policies on video of everyday human work. That is the least certain part of the argument, and he treats it as the part with the most headroom.
The disclosure, which he makes himself
He is not neutral, and does not pretend to be. He has put his own money into a robotics company whose runtime is open source, whose modules he publishes, and which charges for cloud services — the same shape as the database business he built before. Business and open source are not opposed, he says; that is the model, not a compromise.
What he is arguing against is a specific outcome: millions of developers working for the ten companies that pulled the value behind a paywall. The alternative he wants is duller and larger — open physical AI in every business from a small machine shop to a shipping line to a commercial laundry.
What that changes for the rest of us
An open runtime moves the risk rather than removing it. If the model is a commodity and the runtime is public code, then the differentiator and the exposure both migrate to the deployment: what you run, what it can reach, and who can patch it when a flaw lands. That is a security problem before it is a licensing question, and it is the part of the open-robotics argument that tends to go unstated.
You not only get software — you get secure software. If you are putting a machine on a network with an open runtime underneath it, talk to us: [email protected].
Source: Eliot Horowitz, "Open Is How Robotics Wins", 18 November 2026. Positions are the author's; the disclosure of his investment is his own.