From 100,000 Hours to 1 Million Hours a Quarter: China's GenRobot Raises $100 Million to Feed Physical AI

The bottleneck in robotics is increasingly not the model but the data, and one Chinese company just raised nine figures to widen that pipe. GenRobot, known locally as Jianzhi Robotics, disclosed on November 2 that it closed a $100 million Series A+ during the third quarter of 2026. The round drew a notably industrial investor list: SAIC Motor, one of China's largest automakers, Malaysia's EastWorlds, Baidu, Shoucheng Holdings and the robot company GALBOT as strategic investors, with Guanghe Ventures, SUNHEM, CITIC Jinshi Investment, CCB International and Huatai Zijin among the financial backers. No valuation was disclosed.

GenRobot positions itself as data and evaluation infrastructure for embodied AI and world models. The business model is to capture what humans actually do — operations and interactions in homes, factories and logistics settings — and convert that human skill into structured data that robot foundation models can train on. The company describes a closed loop spanning large-scale human skill data collection, training and processing through its own Data Foundation Model, and validation on a cluster of real robots, rather than selling raw recordings.

The delivery figures are the company's own and come without third-party verification: quarterly delivery of usable training data has grown from the 100,000-hour scale to more than 1 million hours per quarter, and on current capacity and delivery plans GenRobot expects to reach 10 million hours a quarter in 2027. The jump matters because robot foundation model teams — in China and elsewhere — have repeatedly said that human demonstration data at volume, not compute, is what limits progress in manipulation tasks.

The hardware side is expanding in step. GenRobot says it will formally launch DAS 2.0 during the Conference on Robot Learning (CoRL), comprising two new capture systems. Ego2.0 pairs a first-person camera with a wrist-mounted unit and hard head-hand synchronization, and supports both body-free data collection and robot teleoperation. Dex2.0 was designed backwards from model data requirements rather than as a stack of sensors: it integrates a side-mounted bionic exoskeleton, in-joint magnetic encoders and semi-flexible full-hand electronic skin, and likewise runs in both body-free and teleoperation modes. The company's slogan for the launch — release means mass production — is a claim that research-grade capture rigs have become a shippable product line.

The investor list is the strategic story. An automaker putting money into robot training data is a statement about where manufacturing is heading: SAIC's interest tracks the broader push to move embodied AI from demo videos to production lines, a shift Chinese policymakers have been actively accelerating. Baidu's presence links the data layer to one of China's frontier model camps, and Shoucheng Holdings has become a recurring name in the country's AI infrastructure financing wave.

GenRobot's market framing, laid out in the announcement, is that physical AI is entering a growth phase in which general AI giants — not just robotics startups — are entering systematically, and that demand will shift from frontier labs to industry, and from training data to production data. Each stage transition, the company argues, expands the addressable market by orders of magnitude. That is a pitch, not a forecast, but it matches what the past year of Chinese embodied-AI financing has actually looked like.

The caveats are the usual ones for a Chinese growth-stage announcement: the round's figures, delivery volumes and capacity targets are all company-provided, the disclosure came months after the round actually closed, and no independent benchmark exists for data quality — hours delivered is not the same as hours that improve a model. The test will come at CoRL, where DAS 2.0 ships, and later, when customers' robot models either do or do not get better because of the data they bought.