China's 90 data collection centers fuel 'circular financing' scandal in embodied AI, IPO crackdown follows

创业邦 · wechat · 2026-09-24

In under two years, China built at least 90 embodied-AI data collection centers, with local governments as the main driver — but the "circular financing" critique exploded in H2 2026.

The fundamentals: data is scarce

Robots need physical-interaction data; teleoperation data is highest fidelity but priciest: 600-700k RMB per robot, 10,000+ sqm facilities
Guojin Securities estimates useful embodied models need 10M hours of multimodal data; global supply covers under 5%

The circular financing model

FT labeled the scheme: government capital builds centers → centers buy robots from companies (tens to hundreds of millions RMB) → companies pay the money back as "data purchases," with companies as the only real customers
Melloddy founder Shao Tianlan publicly named Galbot (200亿+ valuation, IPO-bound) for fabricating revenue via related-party deals
CSRC window guidance: humanoid robot IPO applicants must prove sustainable revenue, narrowing losses, or real technical innovation

Market fallout

Unitree hit 1,100 RMB/share (4449亿 market cap) on debut, then halved within a month, wiping 2400亿+; Q1 2026 revenue growth fell from 332% to 68%, non-GAAP net profit down 52.55%
Three layers of bagholders: local governments, retail investors, and true believers in the industry

Two sides and telltale signs

Positive examples exist: Itshizhihang's robots running stably in Aptiv production lines; Xinzhijitile's tactile data route
Three litmus tests: de-risked repeat purchase rate, whether data actually improves models verifiably, and operating cash flow
Key difference from EV subsidies: subsidies supported real buyers, while data centers create buyers that otherwise don't exist — packaging R&D spend as revenue.

Original post →

China's 90 data collection centers fuel 'circular financing' scandal in embodied AI, IPO crackdown follows

More from Venture