Embodied AI data collection grows into a crowdsourced industry chain
A TMTPost-published analysis says embodied AI is bottlenecked by physical-world data: Mifeng Technology CEO Yao Maoqing put the need at roughly 100 million hours of real operation data for a "GPT-3.5-level" robotics model, versus only a few hundred thousand hours globally, while CAICT estimates about 1.1 trillion tokens are required. It reports 15 data-only startups raised about 4.47 billion yuan in 2026, China's data annotation market is expected to top 15 billion yuan (up over 27% year on year), Mifeng's "Mifeng Pai" crowdsourcing platform (device rental 39 yuan a day, 19 yuan in a promotion, about 20 yuan per verified hour, 40 cities) launched in September 2026, and Figure's Index reached 69,943 weekly contributors. It also flags prices falling from over 200 yuan to 100-plus yuan per hour with some data at a few yuan, multi-layer subcontracting, consent captured in employment agreements, and National Data Administration pilot cities.
The bottleneck for embodied AI is shifting from models to data supply, and how pricing, privacy and labor rules are written will decide whether this chain becomes a healthy ecosystem.