Overview
A dispute between Chinese robotics founders has exposed the economics of embodied AI data collection. Lingyu Intelligence CEO Jin Ge, in an interview with Dingjiao One, laid out the numbers: a humanoid robot sells for 150,000 yuan; data collection centers repurchase its data at 120–150 yuan per hour; a center running 100 robots produces about 400 hours of real-robot data per day. At these prices, one robot pays for itself in roughly a year.
The Accusation
On September 9–10, 2026, Mech-Mind founder Shao Tianlan posted on WeChat Moments describing a class of companies—based in Beijing/Shanghai, famous, highly valued—that rely heavily on "data collection centers" and related-party transactions with local governments, investors, and suppliers to generate "fake, unsustainable revenue," calling this "illegal, unethical, and unwise." In the comments he named Galaxy General.
> This is one party's allegation. No regulator has made a determination and no audit supports it publicly.
Galaxy General responded with a statement, "Embodied intelligence is a long race," declining to engage directly, and citing its AstraBrain end-to-end model, its embodied dataset, and two years of operational validation in industrial, retail, and pharmacy deployments.
How the Money Loops
The alleged loop: local government/state platforms fund data collection centers → centers buy robots from vendors → vendors book revenue → centers repurchase the robots' data → capital flows back to related parties.
The loop itself is not illegal. Related-party transactions are not inherently fake; what must be verified: contracts, pricing, delivery, payment, disclosure of relationships, and—critically—repeat purchases by third parties.
The Two Companies' Books
Mech-Mind (3D vision, algorithms, software components; listed on HKEX September 1 at HK$101.70, ~3,835x oversubscribed retail offering, ~HK$2.353B raised, broke below issue price on debut):
| Item | 2023 | 2025 | Change | |---|---|---|---| | Revenue | 181M yuan | 389M yuan | More than doubled | | Gross margin | 39.1% | 64.6% | +25.5 pp | | Cumulative net loss | — | 1.044B yuan (3 years) | Loss expected in 2026 | | Overseas revenue share | — | 50.3% | — |
Over 80% of 2025 revenue came via system integrator channels; it does not build complete robots.
Galaxy General (founded 2023 in Beijing; full-robot + foundation model strategy) raised 2.5B yuan in March 2026 at a valuation above 20B yuan. It claims pharmacy deployments in 20+ cities with a claimed >95% success rate on medication retrieval (vendor-stated).
Industry-Level Numbers
- H1 2026: 288 embodied/robotics funding rounds involving 226 companies, disclosed amounts over 46B yuan—already exceeding all of 2025.
- At least 15 large embodied data collection facilities nationally.
- 2025 global humanoid shipments ~20,000 units, ~95% from Chinese vendors; 150+ humanoid companies in China.
- H1 2026 government procurement of humanoids, training systems, and pilot projects: at least $230M, versus ~$6M a year earlier.
- Skild AI launched S1, a foundation model learning long-horizon tasks from single videos, built on NVIDIA Blackwell/Isaac/Omniverse with mixed synthetic and real data.
- Mecka AI (robot training data infrastructure, two years old) is closing a Sequoia-led round at ~$500M valuation.
- Maven Robotics emerged from stealth with a $100M Series A, selling hardware, software, and deployment as a package.
Timing and Market Readings
The dispute followed an unconfirmed report (The Information, Sept 9) that regulators have informally signaled to banks and investors that IPO-bound humanoid makers must demonstrate sustainable revenue, shrinking losses, or genuine innovation. In secondary markets, Unitree closed at 477 yuan/share (192.98B yuan market cap) on Sept 11, down from a peak 444.9B yuan on its Aug 19 listing; Mech-Mind remained below issue price.
Same Week at CIFTIS
At the 2026 CIFTIS in Beijing (Sept 9–13): a Jingchen Zhishuang dual-arm "unmanned post office" kiosk (±0.03mm precision, full order-pay-grab-stamp-deliver loop); NOLO Ego headsets capturing 360° first-person operation data; Inspirai Robotics tested in fruit/tea picking with eldercare as a focus; and two robots role-playing shopkeeper and clerk at a recreated Liulichang antique street, narrating 100+ artifacts.
Money Flowing to the Data Layer
The Question Isn't "Who Is Faking It"
"The phenomenon exists" (data center subcontracting chains) does not entail the accusation (a specific company inflates revenue)—the gap requires audits and transaction records. A fairer description: hardware has been delivered and use cases are growing, but efficiency, cost, and repeat-purchase economics remain unproven, and robot training centers still depend on subsidies.
The Final Question
The hardest question is no longer whether robots can complete a task once—CIFTIS showed many demos—but whether the same task can be completed 10,000 times in a row, and who will keep paying for those 10,000 completions. If IPO scrutiny tightens, named companies will be required to disclose detailed related-party transactions; that will be the real test of this dispute.
References
1. Tencent News / Red Star Capital Bureau, 2026-09-12 — https://news.qq.com/rain/a/20260912A054X800 2. China Energy News — https://www.cnenergynews.cn/article/4TAeUdl1pwV 3. Sohu, "A WeChat Moment Ignites an Embodied AI Civil War" — https://www.sohu.com/a/1074882837_121777994 4. FutureX · Physical AI Daily, Issue 118 (09/13) — https://dev.to/future_x/futurex-physical-ai-daily-issue-118-0913-39je 5. Khyber Mail, "Robots Take Centre Stage at 2026 CIFTIS in Beijing," 2026-09-12 — https://khybermail.com/robots-take-centre-stage-at-2026-ciftis-in-beijing-envisioning-smarter-daily-living