Yoni Rechtman of Slow Ventures recently argued that the corpus is the economy: the internet was only the first training set, and the next will be drawn from the work people do for a living. We think he is right, and that the implication is larger than it first appears. The part of the economy that makes physical things, its factories, process lines and laboratories, largely left the United States a generation ago. If the next corpus is the economy, then the portion that could teach a model to build sits outside the reach of the labs that need it most.
We arrived at a narrower version of Yoni’s argument a few months before he published it, and we have spent the time since testing it. Over that period we have worked with fifty firms in China. What follows is what we have learned, and why we believe the gap is one of the largest unaddressed problems in AI training.
01
What models still cannot do
Language models learn most from material unlike anything they have already seen. A model’s ten-thousandth contract summary adds almost nothing, while a record of a problem it has never encountered can move it measurably. Researchers call such material out-of-distribution, and acquiring it has become one of the central costs of frontier training.
For most of the past decade, the internet supplied it for free. That abundance explains why models mastered writing and programming first: both fields had been publishing their work online for decades. When public text ran short, labs began paying experts to produce more. The market for expert data has since grown into a multibillion-dollar industry, but it draws overwhelmingly on a single population, American professionals whose work happens at a desk, and its output increasingly resembles itself.
The industrial economy is almost absent from that market. Ask a frontier model why a particular circuit board needed four revisions and it will produce a fluent answer that could apply to any board. The model has read a great deal about electronics. It has almost never observed an engineer debugging a specific board, because no one ever wrote that record down anywhere a model could find it.
Training data covers the knowledge economy, not the industrial one
02 22.5431° N, 114.0579° E
Tuesday in Shenzhen
I entered the field selling computer-use data to a frontier lab: labelled recordings of people performing work on screens. The business was good until the rest of the market discovered it, which took a few months.
The more valuable part of that period was the research conversations. The researchers I worked with kept returning to requests nobody could fill. They wanted CAD assemblies with their constraints intact, circuit layouts accompanied by their full revision history, and laboratory protocols that had failed. When I asked where else they had looked, the answer was nearly always the same. They had searched, and no supplier could provide material of that kind at volume.
I brought the list to Sandy, now my co-founder, who had run a fifty-person team spanning China and the United States at his previous company. He read it and laughed. Nothing on the list was rare, he said. It was “Tuesday in Shenzhen.”
Odulis began with Sandy’s observation: the material frontier labs considered hard-to-find was routine work on the other side of the Pacific.
03
Where the expertise went
A request can be difficult in San Francisco and ordinary in Shenzhen because of choices American industry made over two decades. Through the 1990s and 2000s, companies kept design at home and moved production abroad. Final assembly left first. Suppliers and tool shops followed, since they need to sit near the line, and much of the engineering eventually followed them.
The savings were real, but the costs were harder to see at the time. Manufacturing expertise accumulates through repetition. When production moved, the people accumulating that expertise moved with it. In many industries, the engineers who understand a process best no longer work in the United States.
The scale of the shift is substantial. China now produces roughly 28% of the world’s manufacturing value added, about $4.7 trillion a year, more than triple its share in 2004. The United States produces about 16%. More than four-fifths of global manufacturing now takes place outside America.
China more than tripled its share of world manufacturing in twenty years
Laboratory science followed a similar path with far less public attention. When a pharmaceutical company contracts its chemistry to a research organization abroad, it receives the finished compound. The record of which reactions failed, and how the chemists recovered, remains with the chemists.
American engineering and science remain formidable, and a great deal of excellent work is still done here. But training a model in a single field requires thousands of examples, and at that volume there is lots of work concentrated overseas. The result is a structural mismatch. Expertise frontier models lack sits outside the market frontier labs buy from.
04
Why no one has bought it
The obvious objection is that labs could hire this expertise themselves. In the United States, they mostly cannot. The engineers and scientists who could build industrial training data here are employed full-time and well paid, and every expert-data vendor in the market is already competing to recruit them. Each new vendor draws from the same small pool, which is a large part of why that market’s output has started to look alike.
China is in a different position. It has the deepest bench of practicing engineers and applied scientists in the world, and a meaningful share of that bench is already organized. Teams there build training data for domestic AI labs today, working alongside the factories, design houses and research groups whose expertise they draw on. The supply American labs say they cannot find already exists, in working form, on the other side of the Pacific.
What it lacks is a connection. A Chinese engineering firm has never sold to an American lab and has no way of learning what one will accept. The lab, for its part, cannot vet a firm eight thousand miles away, let alone judge whether its tasks hold up against a frontier model. Neither side was built to close that distance, and so the largest pool of industrial expertise in the world has stayed almost entirely outside the market for training data.
05
What we found
We expected locating the material to be the hardest part of the work. It proved to be the easiest.
Our first conversations were with Chinese data vendors, small teams already building training tasks for domestic AI labs. The vendors introduced us to the engineering firms they work with, and those firms introduced us to research groups and manufacturers. Each introduction produced several more.
The quality of the material exceeded our expectations. We found chip design data and PCB engineering work from teams that do both commercially, along with wet labs that run fully automated. Several teams had already built training environments on top of their own engineering records, and some of those environments were genuinely strong.
Nearly all of them, however, shared one weakness. None had been through the unwritten quality checks a team only learns by delivering to frontier labs, and no American lab had reviewed them. The result was work that looked finished and wasn’t. A task its authors thought brutal would fall to a frontier model on the first try, or its grader would wave through a wrong answer, and the team would never find out, because the models and researchers who could have told them were on the other side of the Pacific.
That finding reframed the problem for us. We had set out looking for expertise we assumed was scarce. Instead we found abundant expertise that had never been connected to the feedback loop that turns knowledge into training data.
Today we work with fifty partners across China.
06
The company in the middle
Other industries have faced the same structure, and each resolved it in a similar way. When Western apparel brands sought to source from thousands of Asian factories without managing each one, Li & Fung positioned itself between them. It selected the factory for every order and answered to the brand when an order went wrong, without ever owning a sewing machine. Clinical research developed a comparable institution. Because a drug company cannot inspect every hospital running its trial, contract research organizations perform the inspections and vouch for the data.
AI has already produced an early example. Several of the strongest open-weight models come from Chinese labs, and many American companies run them through domestic providers such as Baseten rather than sourcing them directly.
Three markets grew a company in the middle. The fourth has not.
In each case, fragmented producers met exacting buyers, and the intermediary that reconciled them built a durable, valuable business. Industrial training data has both conditions in abundance. It does not yet have the intermediary.
07
The Refinery
Odulis is built to fill that role. Internally, we call the operation the Refinery.
We do not wait for labs to send specifications, because we already know what labs need: the precise points at which their models fail. We search for those failures in the work our partners perform every day, and we build a program around each one we find. A single program might combine a design house’s drawings, a factory engineer’s account of why a part is difficult to produce, and a data team’s task construction. Labs contract with Odulis, and so does each partner, which makes us accountable for every delivery.
Most of our own effort goes into testing. Every task runs against current frontier models before it ships, and we work hard to make graders accept incorrect answers before a model can. When a task fails, it returns to the person who can explain why. We also measure the full cost of each accepted task, including rework, because that figure determines whether the operation is viable as a business.
All testing and lab-facing work takes place in the United States. Our partners contribute the expertise, while evaluation and the lab relationships remain here.
Raw expertise goes in. Lab-grade data comes out.
Inside the Refinery
- Find where models fail
- Combine partners into programs
- QA every task to lab grade against frontier models
- Rework what fails until it passes
08
What we believe
Our bet is that every step toward superintelligence will require more out-of-distribution data than the step before it. A model improves only on what it has not already seen, and the sources the industry relies on today are running short of novelty. Progress will depend more and more on whoever can bring in material from outside them.
The largest untapped reserve is the hard, physical work that the developed world sent abroad decades ago. It covers every industrial field and most of the planet’s factories and laboratories, and almost none of it has been turned into training data. That is a whole world of material in need of an integrator. We have positioned Odulis to be that integrator, with fifty partners already in place and the testing loop running beside the labs.
The position strengthens with use. Every program shows us which teams excel at which kinds of work and where tasks tend to break, and a competitor would have to learn that the way we are learning it, one program at a time.
Expert-built tasks are where the operation starts, and the same network reaches much further. It can extend to sensor data from automated factories and to the experimental output of chemistry and biology labs, each a source of data no model has seen and each able to move through the same Refinery.
We are laying the groundwork to source and refine that data for as long as models keep needing more of it, which we expect to be well into the era of superintelligence.
Sources
- Manufacturing value added by country and for the world, 2004 to 2024: World Bank, Manufacturing value added (NV.IND.MANF.CD).
- China’s manufacturing value added of $4.66 trillion in 2024, with its share more than tripled since 2004: IBTimes UK. World total of about $16.8 trillion: Cargoson.
- US at about 16.6% of world manufacturing value added (2025, World Bank and BEA data): Engimarket.
- China’s share estimated at 26 to 30%: ITIF, Feb 18, 2025.
- “The corpus is the economy”: Yoni Rechtman on X.
