An hour of robot training data cost about $340 in early 2024. In March this year it cost $118.
The work behind that number is unglamorous and specific. Chinese data vendors pay people to cook and fold laundry while a head-mounted camera records their hands, and one woman in Suqian told Rest of World she earned 20 yuan an hour, roughly three dollars, filming her own chores six hours a day.
Nobody had paid her to do that work before.
A 65 percent fall in twenty-six months describes a market being industrialised out from under its suppliers.
The price curve contains the whole thesis for any Southeast Asian firm now selling collection hours into this market. A 65 percent fall in twenty-six months describes a market being industrialised out from under its suppliers.
Two forces push the price down at once. China has stood up more than forty government-backed robot training centres, around twenty of them already running, and JD.com alone wants ten million hours of demonstration footage inside two years, drawing on 100,000 of its own employees and half a million external workers. Supply at that scale sets the price for everybody.
The second force is worse for a supplier, because it removes the need to buy hours at all.
Teams at Carnegie Mellon and Stanford reported this year that models trained on 40 percent synthetic data matched models trained on 100 percent real demonstrations, on tasks neither had seen. Synthetic generation still needs real seeds, so demand thins rather than vanishes. A firm priced against thinning demand loses twice, on volume and on rate.
One number in the same research has held still, and it is the one worth building on.
A person collecting demonstrations produces fewer than 200 a day, and no amount of capital has moved that ceiling. Rigs cost between $50,000 and $150,000 and they do not raise it either, because the constraint sits in the hands and the attention rather than in the equipment.
Throughput is a property of people, and people take a year to get good.
Throughput is a property of people, and people take a year to get good.
That is where the durable asset sits. Operators in the Philippines bill between $22 and $55 an hour against $65 to $120 for American ones, and that gap is the arbitrage, which is to say the part that closes. A retained roster that knows how to stage a task, reset a scene between takes and log the result to spec is the thing a buyer cannot conjure in a quarter.
Singapore is about to make the distinction literal. The National Robotics Programme put S$635 million into the sector between 2022 and 2025, and its second phase, expected in the back half of this year at around S$800 million, carries a national robotics data standard: a shared format for operational data collected by government-deployed robots.1
A standard outlives the price war beneath it.
The move for a regional collection firm is to build to that format before it publishes rather than after, and to sell certification against it instead of footage. Hours are quoted per hour. Compliance is quoted per contract, and contracts renew.
The workers in China’s training centres already have a name for themselves. They call themselves cyber-laborers, and one told Rest of World the work was a fine gig, though a bit boring.
The boring part is what the region is being paid to learn, and it is the part that survives the price.
Footnotes
-
The standard covers government-deployed robots only, which is a narrow slice of the market and most of the reason it will be adopted at all. ↩