Who owns physical data?

In the flagship piece I argued that the binding constraint on physical-world AI is not architecture and not compute — it is data that doesn't exist yet and has to be captured by hand. That essay named the wedge in passing: unlike the internet, physical data has a geography. This one is about what that geography does — who ends up owning the data layer, and why ownership is decided less by who has the best model than by who is standing in the right place when the recording happens.

Start with the property that makes physical data strange. Internet data is placeless. A token scraped in Jakarta is byte-for-byte identical to one scraped in San Francisco; a sentence is a sentence wherever it was typed. That placelessness is why the entire language-model race could be run from a handful of GPU clusters anywhere on Earth — the corpus had already been gathered, for free, by everyone, and it didn't matter where you sat to train on it.

Physical-interaction data breaks that symmetry completely. A dataset of hands assembling a circuit board, an arm sorting bruised produce, a forklift threading a crowded dock — each one is recorded on site, in real time, by an operator who is physically there. You cannot scrape the motion of a wrist under load. You have to be in the room with the wrist. And that single constraint relocates the whole question of ownership from the cloud to a map.

The work is already somewhere — and it isn't the Bay Area

If physical data is only born where physical work happens, the first question is simply: where does the work happen? The answer is lopsided to the point of being almost funny. In 2024, Asia absorbed 74% of all new industrial-robot installations. China alone took 54% — roughly 295,000 machines in a single year — on an operational base above two million robots, about four and a half times Japan's. South Korea runs the densest robot workforce on the planet: 1,220 robots for every ten thousand human workers.¹

Read that map as a data map and it gets sharper. Every one of those machines is a sensor-laden platform doing contact-rich work all day — and every hour it runs is an hour of physical-interaction data that can, in principle, be captured. The richest training signal for the next generation of robots isn't sitting on a server in California. It is being generated, right now, on factory floors in Shenzhen, Suwon, Penang, and an arc of manufacturing towns most of the AI industry has never visited.

The historical rhyme — and where it breaks

We have watched AI's data layer migrate once before. The last data boom — labeling, ranking, RLHF, content moderation — quietly offshored to the Global South the moment it turned into a cost game. The data behind American frontier models was cleaned, rated, and made safe by workers in Nairobi, Manila, and Hyderabad, often for a few dollars an hour.² The lesson investors took from it was straightforward: when an AI input becomes a commodity, it flows downhill to wherever skilled labor is cheapest.

Physical capture rhymes with that — but it breaks the pattern in one decisive way. Labeling was data you could move. A photo to annotate, a transcript to rate, a passage to rank all travel over fiber at the speed of light; the worker can be anywhere a packet can reach. Physical capture cannot be moved at all. You cannot relocate a factory floor to a cheaper time zone, and you cannot pipe the act of bending metal across an ocean. The work and the recording of the work are welded to the same coordinates.

So the offshoring logic still applies — cost still wants the affordable operator — but it now points at a much smaller set of places: not "anywhere with cheap labor and a laptop," but "anywhere with the physical work and cheap skilled labor in the same building." That intersection is the prize, and it is far narrower than the labeling map ever was.

Why presence compounds into ownership

Location wouldn't matter much if it were a one-time advantage — show up, grab a dataset, ship it anywhere. It matters because presence compounds. Sit where the work is and you capture interaction data continuously, as a byproduct of operations that were happening regardless. That data trains better models; better models win more deployments; more deployments sit on more floors and capture still more data. The loop turns on its own, and it turns fastest for whoever is physically closest to the work.

A competitor an ocean away cannot crawl that loop — there is no crawler — so to match it they have to physically show up, stand up operations, and start their own recording from zero. By the time they do, the incumbent is a generation of data ahead and still pulling away. That is what makes capture geography a moat rather than a head start: the advantage is not the dataset you have today, it is the position that keeps minting fresh data tomorrow while rivals are still booking flights.

This is also where I should be honest about my own position, because it cuts directly toward where I sit. The "China + 1" shift — manufacturing spreading into Vietnam, India, and Mexico — is usually read as a supply-chain story. Read as a data story, it is a map of where the next capture sites are being built: the same density of hands-on work that made these places manufacturing hubs makes them the natural birthplaces of physical-AI data. From a Southeast Asian vantage point that looks like an edge rather than a handicap. I want to flag plainly that this is my positional thesis, not a neutral fact — I have an interest in the geography mattering. The rest of this piece is the part I'd defend even if I were sitting in Menlo Park.

And the capital is already pricing a data layer it expects someone to own. Robotics and physical-AI venture funding hit $27.6B in 2025, more than double the $13.7B of 2024.⁴ The clearest single signal is Physical Intelligence — a company whose entire pitch is the data-and-foundation-model layer for robots — repricing in barely sixteen months from a $2.4B round to a reported north of $11B:

That is not a model paying off — Physical Intelligence's models are good, but so are several rivals'. It is the market betting on whoever can build the data system underneath the models, and betting that the system compounds. The open question this whole piece turns on is where that system gets built, and therefore who gets to own it.

The counter-bets — given their due

I hold the geography view, but the honest version has to survive its three best objections. Each one is a real path by which the moat could turn out to be somewhere other than the map.

1. Synthetic data routes around the map

The strongest counter is that you won't need to record the factory floor at all — you'll generate it. World models like NVIDIA's Cosmos line already synthesize action data and photoreal, rare-event scenes precisely to cut how much real capture a robot needs (I went deep on this in Manufacturing reality). Push that far enough and geography stops mattering: you manufacture reality in a data center, anywhere on Earth, and the whole "be in the room with the wrist" argument dissolves. This is the bet I take most seriously, because if the sim-to-real gap closes hard, it doesn't dent the geography thesis — it deletes it.

2. The moat is the model or the distribution, not the data

The second objection grants that physical data is scarce but denies it's where the durable advantage lives. Maybe data is a transient bottleneck — painful now, commoditized later by pooled corpora and better sample-efficiency — and the lasting moat is the foundation model itself, or the deployed install base and the customer relationships around it. In that world the winner is whoever has the best brain or the widest distribution, and they can buy or rent capture from whoever happens to sit on the floors. Geography becomes a supplier detail, not a moat.

3. Capture de-localizes via teleop-from-anywhere

The third is the most technically specific: if an operator in Hanoi can puppeteer a robot in Ohio over low-latency links — or if capture migrates off factory floors onto consumer wearables that read hands from egocentric video — then capture detaches from the work. The skill stays affordable and remote, the robot stays where the work is, and "be present" quietly becomes "be online." If that holds, the edge diffuses to everywhere, which is the same as nowhere.

Why I still take the geography side

I weigh those seriously, and I still land on geography — for now, and for reasons I can state precisely enough to be proven wrong.

On synthetic data: it has to be grounded in, and validated against, the real thing, and the sim-to-real gap is widest exactly where the value is highest — dexterous, contact-rich, gloriously messy tasks that are the hardest to fake convincingly. Synthetic data is a multiplier on real capture, not a replacement for it; you still need a real-world anchor, and the anchor has a location. On the model-or-distribution objection: models are converging and increasingly portable, and distribution itself, in this domain, is built by deploying onto the floors where the work is — so both roads loop back through physical presence. On teleop-from-anywhere: the operator can be remote, but the robot, the task, the contact, and the sensors cannot — the expensive, scarce half of the loop stays welded to the site, and the latency and reliability needed for genuinely dexterous remote work are not casually solved.

So the bet, stated plainly and in the first person: physical-AI winners will be defined partly by where they capture data, and that map looks like the manufacturing map, not the internet's. The data layer of physical AI will be owned, disproportionately, by whoever is standing where the physical work and the skilled-but-affordable operators already are. That is a positional claim, and the position is mostly in Asia.

Where I'd change my mind

I'd abandon this thesis on a specific, falsifiable signal — and I'd rather name it than pretend the bet is safe. If a frontier robot policy gets trained predominantly on synthetic data and crosses the sim-to-real gap on contact-rich, dexterous tasks — matching teleop-trained systems in the real world without a proportional real-capture anchor — the geography argument is finished, and I'll say so. The softer version: if dexterous teleoperation goes reliably remote at scale, capture de-localizes and the edge diffuses to everyone with a network connection. I'm watching the sim-to-real benchmarks and the remote-dexterity demos as my tripwires, not the funding headlines.

Until one of those trips, the recording is being made by hand, in real time, in the places where the physical world actually gets work done. For thirty years the center of gravity in computing followed the data, and the data was the internet — which belonged to no one and everyone. The next dataset belongs to a place. Whoever stands there when the recording happens is the one who gets to own it.