Greetings from a world where…
the “Thou Mayest” discussion in East of Eden is magisterial
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: The Money Game of Embodied AI
Context: Whenever I give a talk on my thesis that China faces a substantial diffusion deficit in AI — one that’s backed by a lot of research — there’s always that guy in the audience who asks: But what about China’s embodied AI advantage?! This is the same dude who is eager to bring up all the books he’s read about China and technology (but only ones that trade on vibes and anecdotes, as opposed to data and sources). He’s the guy who’s been to China once or twice and visited everywhere (well, everywhere in the first-tier cities).
Here’s the thing, though, that guy is doing a lot. He’s hyping up Chinese robotics start-ups like Unitree, which listed on the Shanghai stock market last week. He’s advising the U.S. Federal Communications Commission to ban imports of foreign-made advanced robotic devices, yet another a counterproductive move in an archaic Fortress America strategy. He’s supremely confident about recursive self-improvement: after all, he was all in on AI 2027 before it changed to AI 2028, and once that forecast gets quietly revised, he’ll pivot quickly to AI 2030.
So, my guy, I dedicate this week’s issue to you. In an excellent, in-depth analysis published in LatePost[晚点], Zinan Li, Shen Yuan, Yumeng Xu puncture the hype surrounding China’s embodied AI sector. I previously translated another longform, deeply-sourced LatePost about the three-way contest among China’s big tech giants to build an AI Super-App (ChinAI #345).
Key Takeaways: The valuations of China’s robotics companies far exceed their actual capabilities. The piece reports on a variety of demonstrations and vignettes that illustrate this point.
From the article: “An investment banker who has participated in roadshows for three robotics companies recalls a due diligence visit in May to an embodied AI firm valued at over 20 billion RMB. Engineers had a robot demonstrate folding a towel; the process took 15 minutes, yet the task remained unfinished by the end.”
“As we wrapped up our reporting, our editor asked: How many robots working on a production line would it take to match one single BYD factory worker?” The reporters struggled to respond. Last month, at a bearing manufacturing plant, they had watched a robot take 70 seconds to extend its gripper to pick up a bearing from a tray and place it onto a nearby plastic bin: “Even if it were ten times faster, it still couldn’t enter factory use.” Earlier in the year, at another demonstration, the reporters watched a humanoid robot take 90 seconds to cradle a box, turn around, walk to a shelf, and set an item down.
Here’s the thing: of the two companies that build the robots for the above demonstrations, one is valued at around 6 billion USD, and the other, founded less than a year ago, is valued at over 1 billion USD. As the image below shows, seven Chinese robotics companies have surpassed 20-billion RMB valuations, with many just established in the last few years.
What is driving all these unsupported valuations? Moving money around to artificially inflate demand. Taking a strong view, this article determines that the current technology and financial numbers cannot justify the industry’s high valuations. Last month, an investor in this sector decided to resign. He stated: “If people say that failing to raise significant capital or achieve a high valuation early on means you lose your opportunity to play at the table, it sounds less like entrepreneurship and more like a game of Texas Hold’em.”
What’s taking place here is a Chinese twist on SpaceX buying Tesla Cybertrucks because no one else wanted them. These robotics companies often partner with local state-owned entities; under these arrangements, the SOE becomes a shareholder of the robotics firm, and the two parties establish a joint venture to build a data collection facility to train robots. The SOE also purchases a lot of robots from its partner, boosting the company’s revenues. For example, in one prospectus from a Chinese robotics firm, local government or state-affiliated entities accounted for 4 of its top 5 customers. So, now, you have a SOE that’s taking on three roles that all conflict with each other — client, investor, and operational partner.
These types of murky guanxi relationships do not reflect sustainable demand. From the article: “However, this type of revenue is viewed controversially by capital markets. We understand that when a leading robotics company filed for an IPO in the fourth quarter of last year, auditors required the exclusion of certain revenues deemed to stem from related-party transactions.”1
The article reminds us to look deeper into these revenue figures from Chinese robotics companies. Unitree, for instance, in 2025 generated less than 10% of its revenues from industrial applications (and half of that small percentage was for corporate tours). Over 70 percent of its revenue came from research applications!
What do these research applications look like? Newly built data collection facilities will buy robots from Unitree and other companies to take footage and gather movement trajectory data from robots performing various tasks, and then they will try and sell that data back to the robotics companies (which recoups part of the purchase cost). Li et al. report on what sometimes happens instead: “In reality, however, visitors to these data-collection facilities have observed robots sitting idle for long periods, with no buyers found for the data already collected. One facility, which had previously purchased 80 million RMB worth of robots, resold some of them to schools early this year because the time needed to recover their initial costs exceeded expectations.”
This nugget from the article made me laugh: “Li Yuanqing, Co-CEO of Lxson, remarked at a media briefing that regardless of the type of robot produced, a company can always sell at least 500 units, simply because 500 competitors will buy them for research purposes.”
The money has not translated into technical capabilities. Embodied AI does not work the same as large language models.
You have all these companies valued at sky-high amounts but not spending much on R&D (Unitree spent just 145 million RMB in R&D last year). Li et al. write, “This has created an unusual financial dynamic within the industry: companies continuously raise funds, yet struggle to rapidly convert that capital into technological advantages, leaving much of the money sitting idle. An investor who has backed companies in the embodied AI sector believes that, theoretically, companies valued at over 10 billion RMB have raised enough capital; yet, the fundraising race continues. ‘What is the point of raising financing? I don't get it.’”
Compared to LLMs, the embodied AI sector suffers from many disadvantages. For one, there is a scarcity of effective data, and there’s no clear pathway for how to scale up data collection. There’s minimal standardization when it comes to model structures and data requirements, not to mention benchmarks to measure progress.
The LatePost article concludes, “One investor in the embodied AI space summarized the shift in mindset over the past few years with a catchy sequence: questioning the bubble, understanding the bubble, embracing the bubble, and enjoying the bubble, but he didn't say what the fifth step was.”
Don’t be that guy who can’t figure out what comes next.
FULL TRANSLATION: The Money Game of Embodied AI
ChinAI Links (Four to Forward)
Should-read: China AI Bulletin 9
For the Safe AI Forum, Emmie Hine provides a very detailed and comprehensive roundup of China’s AI governance and safety landscape. I learned a lot about Chinese-language commentaries on the OpenAI-Hugging Face security incident as well as a a couple of interesting biosecurity benchmarks.
Should-read: Latitudes: Words or action?
Karin Fischer’s newsletter on international education is essential reading. In her latest issue, she covers the U.S. Department of Homeland Security’s continued efforts to undercut the U.S. national interest. The department has “signaled it will apply a more expansive interpretation of regulations governing curricular practical training, one that could make it difficult for international students to engage in academic-related work like co-ops, internships, and practicums.”
Should-read: State of the AI Economy
There’s some good numbers in this report from Exponential View. I would be more careful about some of these headline takeaways such as “AI is scaling three times faster than any IT wave,” where the numbers would change drastically depending on when you date “Year 0” of the AI wave. The authors date it as 2023, but why not 2012 (AlexNet submission to ImageNet) or AlphaGo (2016)?
Should-read: How (some) Chinese AI Practitioners View Model Distillation
Geopolitechs recently translated another great piece by LatePost that interviewed researchers and practitioners in the large model field from different companies. They gave their perspectives on controversies surrounding distillation.
Thank you for reading and engaging.
These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99
Emphasis mine.
You might be aware that Jordan Schneider has just interviewed people with a diametrically opposed view: https://chinatalk.substack.com/p/the-robots-are-here
The 500-competitor customer loop is hilarious and bleak. If buyers are mostly other labs trying to make data to resell into the same ecosystem, the revenue line starts smelling very weird.