Three raises in one day, all for companies that don't build robots

On November 2, three companies that do not build robots announced funding on the same day: Jianzhi Robotics with a $100 million A+ round, GenRobot AI with a $100 million A+, and RoboScience with a Series A2 in the hundreds of millions of yuan. According to Zhicaishe, Jianzhi's cumulative funding now exceeds $300 million.

None of the three builds robots

Jianzhi's investor list includes SAIC, Baidu and Galbot. Three different kinds of player appear at once, each with its own reason: a carmaker wants production-line data, a model company wants training corpora, a robot company wants validation on real machines.

It does not sell machines. It sells three things: collecting human skill data, training and processing a Data Foundation Model, and evaluating robot fleets on real hardware. By the report's account, delivered "effective data" has grown from around 100,000 hours a quarter to the million-hour level.

GenRobot AI follows the same route, showing head-mounted and hand-held capture devices at its booth. RoboScience, in its earlier A2 round, offered a different pitch: it claims to have pushed the cost of a single training data item down to RMB 0.05, about one percent of conventional approaches.

Collecting in-house no longer adds up

Data used to be a cost line for robot makers: hire the people, build the sets, collect it yourself. Marginal cost does not fall — every extra hour means another hour of labour, equipment and space.

Outsourcing follows the logic of early cloud computing: building it yourself looks cheap until depreciation, staff and idle capacity are counted. A data company's value lies in turning capture, cleaning, labelling and evaluation into a pipeline, billed by the hour or the item.

One precondition makes these companies viable: enough robot makers exist. With a dozen customers, a data firm cannot afford collection studios and equipment; with a hundred, one pipeline spreads its cost and justifies separate financing.

That is why capital is willing to price "no-body" companies separately. Bodies are a hardware business measured on margin and shipments; data is an infrastructure business measured on repeat purchase and unit cost.

RMB 0.05 an item, for now

The most eye-catching number is the RMB 0.05 per data item. If it holds, collection stops being a moat and cleaning plus evaluation become the differentiator. But the same report raises the obvious question: how is "effective data" counted?

A hundred hours of messy footage may clean up to a handful of usable hours, and that conversion rate sets the real cost. There is no public standard, and every company quotes the number that flatters it. RMB 0.05 may cover one step of the process, or only one class of structured task.

The harder question is commercial: finishing a large-scale validation and having customers keep paying are separated by a gap no smaller than the technical one. Validation can be carried by one or two big projects; repeat purchase requires each batch of data to be more useful than the last.

The money changed lanes; the road is unfinished

For two years, embodied AI money went to bodies — whoever had the smoother robot video raised next. Now it is tiering: bodies, models and data each carry their own valuation logic.

That is healthy. Data pipelines are dirty work that nobody wanted to pay for separately, so robot makers carried it themselves. Now specialists do it and get funded, which means buyers are starting to accept its value on its own terms.

But the road is not finished. Data companies depend heavily on a few large customers, and those customers are building data teams of their own. If collecting in-house turns out cheaper, the business shrinks back. The numbers to watch are unglamorous: repeat purchase rate, and the actual delivered cost per data item — not the one in the pitch.

This piece summarizes public reporting; funding details follow the source and were not independently verified against company announcements. Not investment advice.