Most of the embodied AI industry is watching who shipped a robot that can backflip, or whose dexterous hand gained another degree of freedom. ByteDance has been moving in a different direction, and the pattern is now clear enough to read.
It is not building hardware. It is not chasing benchmark tables. It is positioning itself underneath the companies that do.
Three moves that fit together
The first is that founder Zhang Yiming is personally overseeing a real-time spatial video generation model, expected as early as October, with the application direction pointed squarely at robotics.
The second is organisational. In June, Seed Robotics, previously led by Li Hang, was folded into Zhou Chang’s remit and merged with multimodal interaction and world models. The stated purpose was technical closure and data reuse. Li Hang moved into an academic advisory role.
The third is hiring. ByteDance is reported to be searching hard for an L8-grade technical lead for embodied AI, reporting to Zhou Chang. An L8 posting at that company is not a research figurehead. It is an operator expected to turn research into infrastructure.
Selling shovels, and the water as well
What Zhang Yiming is supervising is not a robot body. It is closer to a game engine: a real-time model that generates a physically plausible, photoreal world in software, where robot policies can be trained at a volume that real-world collection cannot supply.
Robot training is starved of data. Collecting on real machines is slow and expensive, and every new task resets the count. Generating synthetic visual data at the source attacks the constraint directly, and it is the same logic that made Nvidia’s simulation business valuable.
The external capital is the mirror image. ByteDance led a RMB 1 billion A++ round in Zibianliang in January. That company is putting RMB 300 million into a headquarters in Hefei focused on model engineering and data factories, and reports put its valuation at RMB 20 billion. Meituan, Alibaba, ByteDance and Xiaomi, four groups that compete ferociously in consumer internet, all sit on the same cap table.
Read together, the strategy is one leg synthetic, one leg real: generate training data internally with the model Zhang Yiming is supervising, and collect physical interaction data externally through a data factory that reaches into factories and homes.
Why the valuation anchor is moving
The primary market has already started to reprice. Body companies that could raise on a single striking demo video are being marked down. Capital is concentrating in companies that address the data shortage or the engineering path to deployment, and Zibianliang’s RMB 20 billion valuation is being paid for data infrastructure, not for a robot.
The same transmission is visible in listed equities. The old trade was reducers, ball screws and motors, the mechanical body. Money now pays a higher premium for simulation platforms, data collection equipment and compute silicon. The reasoning is blunt: without data flowing into them, the hardware is scrap.
The comparison to the gold rush is almost too neat, but it holds. The most reliable business was never the mining. It was the shovels, and then the water.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-10/ART-8321205-8460-30708403.html.