In Physical AI, the data is the moat — not the robot

A robotics startup is reportedly offering to clean your apartment for free. There's a catch, and it's a revealing one: they record first-person footage the whole time they do it. The cleaning is the cost. The footage is the product.

It sounds gimmicky until you sit with it. Then it becomes the clearest illustration I've seen of where the real value in “Physical AI” — robots, embodied systems, machines that act in the world — is actually going to sit. Not in the robot. In the data.

Why footage of someone cleaning a flat is worth more than the cleaning

To teach a robot to tidy a real home, you need thousands of examples of hands doing real tasks in real, messy rooms — picking up a sock, wiping a counter, working around a chair that's in the wrong place. That kind of data barely exists at scale. You can't scrape it off the internet. You have to go out into the physical world and capture it, one apartment at a time.

So whoever captures it first owns a head start that a competitor can't simply buy later. The free cleaning isn't generosity. It's a data-collection campaign wearing an apron.

The cleaning is the cost of acquiring something far more valuable: real-world data nobody else has.

Software already ate the easy data. The physical world is what's left.

Language models had it easy, in one specific sense: the internet was a giant, free training corpus sitting there waiting. Decades of human text, already digitised. The physical world offers no such gift. There is no pre-existing, free dataset of “hands doing chores” or “forklifts navigating a warehouse at 4 p.m.”

That's why the race in Physical AI is, underneath, a race to manufacture data: cars logging miles, warehouse arms logging picks, kitchens logging cooks, apartments being cleaned. The model architectures are increasingly shared and copyable. The data you collected by being out in the world is not.

This pattern isn't new — and it isn't only about robots

We've seen the shape before. Self-driving programmes spent years driving simply to gather driving data. Plenty of ordinary products quietly log how they're used and turn that into an advantage rivals can't replicate. The move is always the same: give away the service, keep the data.

And the lesson generalises far beyond robotics, which is the part I want you to take away if you run a normal business.

You're probably sitting on a moat you treat as exhaust

Most companies generate valuable real-world data every day and throw it away without noticing: support chat transcripts, recorded sales calls, site-inspection photos, repair logs, returns and complaints, the dozen revisions a design went through before the client said yes. It feels like exhaust — a by-product. In the AI era it's closer to ore.

Three questions worth asking this quarter:

• What real-world data does only my business see? The stuff a competitor can't get from Google or a public dataset.
• Am I capturing it in a usable form — or letting it evaporate? A phone call that isn't transcribed is a lesson taught to no one.
• Could a model trained on it do something a rival can't copy? That's the test for whether it's a moat or just storage.

The honest caveat: consent isn't optional

Recording the inside of people's homes — or your customers' conversations — raises real questions about consent and privacy, and they deserve straight answers, not fine print. The companies that win the data race and keep their reputation will be the ones who collect openly, with permission, and are clear about what they keep and why. A moat built on data people didn't know they were giving you is a liability waiting to happen.