One Robot Pays Back in a Year: Embodied AI's Data-Center Revenue Loop Goes Public

> A humanoid robot sells for 150,000 yuan. Data-collection centers repurchase the data it generates at 120–150 yuan per hour. A center running 100 robots produces 400 hours of real-machine data per day. At these prices, one robot pays for itself in about a year. These figures come from Lingyu Intelligence CEO Jin Ge in an interview with Dingjiao One. On September 10, 2026, Mech-Mind founder Shao Tianlan publicly named Galbot, pushing this revenue loop into open debate.

What One WeChat Post Said

On September 9–10, Shao Tianlan posted repeatedly on WeChat Moments, describing a type of company: based in Beijing or Shanghai, famous, highly valued, featured on the Spring Festival Gala, heavily using so-called "data-collection centers" and related-party transactions with local governments, investors, and suppliers to create "false, unsustainable revenue." He called this practice "illegal, immoral, and unwise." In the comments, he supplied a name: "for example, Galbot."

This is one party's allegation — there is currently no regulatory finding or audit report supporting it in public channels.

Galbot responded on the evening of September 10 with a statement titled "Embodied AI Is a Marathon; We Only Race Against Time." It did not directly address the data-center or related-party allegations, saying only "we do not join pointless wars of words and are not distracted by short-term noise," while citing its end-to-end model "Galbot AstraBrain," its embodied dataset, and two years of operational deployment in industrial, retail, and pharmacy settings.

How the Money Cycles

To understand the dispute, you need to know what a data-collection center is. Embodied foundation models require large amounts of real-machine data, which is expensive and slow for companies to gather alone. Local governments build public training facilities to supply data and create robotics clusters.

The loop Shao describes:

• Local governments / state-owned platforms fund data-collection centers
• Centers purchase equipment from robot companies (forming the companies' revenue)
• Robot companies buy the data back under data-procurement contracts
• Funds potentially flow back to related parties

The loop itself is not illegal. Related-party transactions are not inherently fake, and government procurement is not inherently without commercial value. What needs checking: contracts, pricing, delivery, payment collection, disclosure of relationships, and — crucially — repeat purchases. If a center buys equipment only once, the revenue is one-off; sustained third-party repurchase is another matter.

Two Companies' Books

Mech-Mind listed on the HKEX main board on September 1 at an issue price of HK$101.70, with its Hong Kong public offering ~3,835x oversubscribed, raising ~HK$2.353 billion. It broke below the issue price on day one (low of HK$75.40) and closed at HK$89.45 on September 10.

| 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 its 2025 revenue came through system integrators. Mech-Mind sells 3D vision, algorithms, and manipulation components — not whole robots.

Galbot, founded in Beijing in 2023, pursues a "whole robot + foundation model" route. Public information shows it completed a 2.5B yuan round in March 2026 at a valuation above 20B yuan. It claims its robots operate in pharmacies in 20+ cities, with a claimed pharmacy dispensing success rate above 95% (vendor-stated). Rumors that Galbot is a Mech-Mind customer were not confirmed.

Industry-Level Numbers Matter More

• H1 2026: 288 embodied/robotics funding rounds involving 226 companies, disclosed amounts exceeding 46 billion yuan — already surpassing all of 2025
• At least 15 large embodied data-collection facilities nationwide
• ~20,000 humanoid robots shipped globally in 2025; Chinese manufacturers accounted for 95%; 150+ humanoid robot companies exist in China
• Government procurement of humanoid robots, training systems, and demonstration projects reached at least $230M in H1 2026, versus ~$6M in the same period of 2024

This paints a specific structure: companies and capacity are growing fast; government investment, industrial funds, and demonstration procurement lead the way; end markets are still slowly searching for cost-covering scenarios.

The Timing Is Not Random

The feud followed an unverified regulatory rumor. On September 9, The Information reported (citing unnamed sources) that regulators had informally engaged in "window guidance" with some banks and investors: humanoid robot companies seeking IPOs would need to demonstrate sustainable revenue, a path to reducing losses, or genuine technological innovation. The report is unconfirmed.

Timeline of September 2026:

• Sep 1: Mech-Mind lists on HKEX; breaks below issue price on day one
• Sep 9: The Information reports IPO window guidance (unverified); Shao's first post
• Sep 10: Shao names Galbot in comments; Galbot issues a statement
• Sep 12: Reports say Galbot has reported the matter to police

Shao can discuss PMF credibly because of Mech-Mind's revenue structure — components sold to industrial customers and integrators have more traceable revenue. Whole-robot companies starting from policy-driven scenarios face easier revenue scrutiny. But these are different businesses; one's PMF cannot substitute for judgment on general-purpose robot business models.

Meanwhile, at CIFTIS: A Different Face

While the feud raged, the 2026 China International Fair for Trade in Services ran in Beijing (Sep 9–13), showcasing concrete deployments:

• A "unmanned post office + cultural-creative" dual-arm robot terminal from Jingchen Digital (Shijingshan), with ±0.03 mm positioning precision, completing an order-pay-grab-stamp-deliver loop, with bilingual greeting. The company claims a data-training cluster covering 680+ sub-scenarios.
• A NOLO Ego headset capturing 360° first-person operation data on-site (teaching robots when to turn in a marathon).
• Inspirai Robotics testing fruit-picking, tea-picking, and tea-processing robots, with eldercare as a focus.
• Two robots role-playing shopkeeper and assistant at a recreated Liulichang antique street, explaining 100+ artifacts — per the Beijing Cultural Heritage Bureau, trained robots can tell artifacts' stories better than human sellers.

These two faces are not contradictory. Demonstrating capability is the first half of the business; turning it into a reliable service repeated 10,000 times is the second half. The first half is progressing far faster than the second.

Money Is Also Flowing to the Data Layer

Three deals from the same week:

• Skild AI released S1, a robot foundation model that learns long-horizon tasks from a single video, built on NVIDIA's Blackwell, Isaac, and Omniverse stack, trained on mixed synthetic and real data.
• Mecka AI, robot training-data infrastructure, founded two years ago, is closing a round led by Sequoia at a valuation near $500M, months after its Series A.
• Maven Robotics emerged from stealth with a $100M Series A; its robots already run at customer sites, sold as hardware + software + deployment.

The direction of capital toward the data layer is itself a signal: one-shot learning is a model-layer capability claim whose raw material — hours of labeled real-machine experience at scale — almost no one can supply.

The Secondary Market Reads It Differently

• Unitree closed at 477 yuan/share on September 11, a market cap of 192.98B yuan — down from a peak of 444.9B yuan on its August 19 debut.
• Mech-Mind closed at HK$95 on September 11, still below its HK$101.70 issue price.

Primary markets price narrative; secondary markets ask about cash flow. When both are present, the question "is the revenue real?" inevitably surfaces.

The Right Question Is Not "Who Is Faking"

Separating "the phenomenon exists" from "a specific company inflated revenue this way" requires audit reports and transaction records — a gap not yet filled. A more accurate description of the industry: equipment has been delivered and application scenarios are growing, but efficiency, cost, and repurchase remain unverified. Robot training-center projects still depend heavily on subsidies; a clear profit model has not formed.

The Final Question

The hardest question in embodied AI is no longer whether a robot can complete a task once — the CIFTIS floor was full of demos. The truly unanswered questions are: can the same task be completed 10,000 times in a row, and who will keep paying for those 10,000 completions?

If Hong Kong or A-share IPO reviews for embodied AI do tighten, named companies will be required to disclose detailed related-party transaction information. That will be the real test of this dispute.

Sources

1. Tencent News / Hongxing Capital Bureau, "Mech-Mind CEO says some embodied AI companies 'stage' deals and create false revenue" (2026-09-12) — https://news.qq.com/rain/a/20260912A054X800 2. China Energy News, "Mech-Mind founder names Galbot; Galbot responds" — https://www.cnenergynews.cn/article/4TAeUdl1pwV 3. Sohu, "An embodied AI civil war ignited by one WeChat post" — 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