Industry Watch 32: Embodied Robotics—Bubbles, Challenges, Value
Summary
Robotics enthusiasm has been jointly driven higher by policy support, its Spring Festival Gala debut, and investor attention, while valuations have already become expensive. The industry’s capabilities and market excitement remain badly out of sync: the government work report has named robotics a future industry, while marathon robots still “run slowly and stumble,” sometimes needing 2 people to lead them; Li Feng notes that “the industry gets a little hotter every quarter.”
Li Feng’s core investment thesis for embodied intelligence is China’s repeatedly validated formula of “soft technology plus a complex hardware supply chain.” Smartphones, electric vehicles, and drones all followed this path to global competitiveness; robotics adds motors, chips, sensors, and AI, and if it ultimately becomes a useful product, China could once again build the world’s largest industry and take it global.
The force that could truly extend the robotics demand curve is the tension between services-sector expansion and demographics over the next 10 to 15 years. Services account for roughly 80% or more of US GDP, versus 54%-55% in China; if incomes keep rising, demand for services will surge, but “the number of people who can afford services may far exceed the number of workers able to provide them,” making robots a potential technological solution to a structural economic problem.
The biggest near-term challenge for embodied intelligence, especially in manipulation, is the lack of real-world data containing force, temperature, material properties, and state changes beyond imitation and reinforcement learning. Historically, data on whether a bottle is soft or hard, whether a cup is hot or cold, or whether an object can be tilted was never collected—or was collected only in very small quantities; video-generation models produce objects that clip through one another because their generation process does not truly incorporate physical constraints. Li Feng remains cautious: “Virtual data alone… I guess it can help, but it should face substantial challenges.”
Because robot hardware has not yet settled into a standard form, investment in data faces the risk of being reset. If a hand has 20 degrees of freedom, there is no prior answer as to whether it should use 2, 3, or 5 fingers; once 2-finger data has finished training, switching to 3 fingers means “all the data is wasted—you have to start over with 3 fingers.” Hardware changes can therefore force repeated restarts in data accumulation.
For an industry large enough, a bubble is part of the expansion mechanism in Li Feng’s argument. The boom attracts capital, talent, and resources; the bust completes the process of creative destruction and leaves behind “the brave ones who remain” for the next cycle. Reflecting on his own failure to invest in Li Auto and its peers, he says that when a national and economic force far exceeds the industry itself, the resulting scale and speed can surpass what seemed rational at the time.
The 10-year horizon may generate extremely high returns, but only for investors who survive multiple valuation cycles rather than extrapolating historical cases directly. Li Auto’s ideal early valuation was about RMB700M, while other new-energy vehicle brands generally started at $500M; CATL was still in the private market in 2014-2015 at a valuation of “roughly RMB20B-plus”—though Li Feng said he could not remember whether that figure was accurate—and he estimated today’s return at several dozen times, below 100x, or close to 100x. No one at the time could have foreseen the eventual scale or the volatility along the way.
Deep dive
1. Robotics Was First Lifted by a Succession of Events; Investment Started Before Large Models
Li Feng begins with the market’s most genuine divide: optimists see robotics entering the government’s future-industry plans, while skeptics see marathon robots moving slowly, falling repeatedly, and needing people to lead them. Venture investors are likewise split between those selling their holdings and those continuing to build positions; the gap between demonstrations and industrial expectations remains enormous.
The excitement was then pushed higher by a succession of events: robots danced at the Spring Festival Gala, robotics-company representatives attended and spoke at a symposium on private enterprises, and the government work report again named robotics a future industry. Li Feng’s market observation is that “the industry gets a little hotter every quarter,” and companies have already become relatively expensive.
Fengrui began investing around 2022, when robotics was still unfashionable; it only “warmed up a little” by late 2024. The original technical rationale was not the large models that later became popular, but the progress robotics had already made through earlier-generation AI technologies such as reinforcement learning. Large models did not generate a broader wave of enthusiasm until after what he calls “the end of 2023.”
2. China’s Advantage Comes From “Soft Tech × Complex Hardware”; Bubbles Expand Large Industries
Li Feng’s industrial rule is that by combining high-value software technology with a long, complex hardware supply chain to redefine a useful new product, China has a chance to build the world’s largest industry and take it global. Smartphones, electric vehicles, and drones are three examples; robotics adds motors, chips, sensors, and AI to the formula.
For an industry large enough, Li Feng is categorical about the cycle: “Going through a bubble like this is inevitable, and going through the bursting of such a bubble is also inevitable. Otherwise, it cannot develop.” The boom throws vast amounts of resources into the sector, concentrating industrial capacity, technology, and talent; the bust completes the process of creative destruction and leaves behind “the brave ones who remain” for the next round.
He uses new-energy vehicles to revisit his own mistake. He once believed that entrepreneurs who had never built cars were violating commercial logic by entering the auto industry, so he invested only in electrification and intelligent-driving components such as solid-state batteries, sensors, and controllers, rather than betting on vehicle brands. He later realized that when national and economic forces are far stronger than the industry itself, the sector’s scale and development speed can break through what seemed like rational estimates at the time.
3. The Long-Term Demand Anchor Is the Gap Between Services Expansion and Demographics
Once China’s per-capita GDP entered the $10K-plus range, services should, according to the historical pattern Li Feng cites, have entered a period of rapid growth; policy stimulus and market opening are also pointing toward services consumption. Services account for roughly 80% or more of US GDP, versus 54%-55% in China, and China still needs further services-sector development as the effects of the pandemic fade.
The problem is that services are “extremely labor-intensive”: restaurants, travel and culture, entertainment, education, and other services both create jobs and depend on large numbers of service workers. If incomes and demand continue to rise over the next 10 to 15 years while demographics constrain supply, the economy could face the contradiction that “more people want to buy services, but there is no one available to serve them.”
E-commerce is Li Feng’s first example of a technological solution. After 2001, real-estate growth drove demand for renovations, appliances, and furniture, but offline retail expansion could not keep up; beginning around 2003, e-commerce used online channels, digitization, and technology to overcome the shortage of physical networks and grow faster than offline retail.
Ride-hailing is “half an example”: it remains part of the services sector, but uses digitization and technology to improve how supply is organized. China has an abundant supply of ride-hailing cars, which typically cost only 5%-10% more than taxis while offering black C-class sedans, water, charging cables, and a cleaner ride. Li Feng’s point is that seemingly intractable service-sector contradictions are often converted by technology into enormous markets.
4. The Key Bottleneck for Embodied Intelligence Is Data With Physical Constraints
Since 2015, China’s production and deployment of industrial robots—what Li Feng explains as manufacturing and use—has ranked first globally. Looking forward from that point, he also expects the world’s largest industrial-robot manufacturer, seller, and service provider to emerge in China. But embodied robots have accumulated far less manipulation data than autonomous driving: competition in autonomous driving began roughly 15 years ago, and cars have spent decades accumulating data under relatively simple control and motion objectives; today’s new-energy vehicles are also packed with countless sensors.
A machine must distinguish how much water is in a bottle, whether a container is soft or hard, whether it is a paper cup or an iron cup, whether it is hot or cold, and whether it can be tilted. Data containing such physical quantities and state changes was never collected in the past, or was collected only in very small amounts. Video-generation models that produce arms twisting backward or fists passing through walls expose the same problem: text and video generation do not incorporate the constraints of physical laws.
Humans develop intuition through experience. A person weighing 180 jin who sees a stool about to collapse naturally will not sit on it. If a robot does not understand the load-bearing limits of the human body or the scale of human interaction, its rigid movements could “either twist your arm off or break whatever it is handling.” Li Feng therefore makes only the cautious judgment that virtual data “can help,” while still presenting substantial challenges.
More difficult still, data is tied to hardware form. If a hand has 20 degrees of freedom, it may not need 5 fingers; 3 fingers might already cover 99% of fine-manipulation and force tasks. But moving from 2 fingers to 3, Li Feng says, would make all existing training data worthless; moving to 5 fingers later would require starting over again. Changes in hardware form can repeatedly reset the accumulation of data.
5. A 10-Year Horizon Can Rewrite Valuation Judgments, but Intermediate Cleansing Is Inevitable
Li Feng explains why Fengrui did not sell its robotics holdings even as peers were selling theirs: if bubbles and busts are part of how large industries expand, expensive valuations and immature capabilities today are not, by themselves, enough to invalidate long-term value. The historical cases, however, come with a condition: investors actually had to wait through every bubble and bust.
He recalls that around 2015, Li Auto’s cheapest valuation round was about RMB700M, while most other new-energy vehicle companies began with financing valuations of $500M or more; he did not participate because of the fund’s size and his risk assessment. CATL was still in the private market in 2014-2015 at a valuation of “roughly RMB20B-plus,” though Li Feng said he could not remember whether that figure was accurate, and added that almost nobody would have dared invest in its financing at the time. By his estimate, the return today would be several dozen times, below 100x, or close to 100x.
These examples are not forecasts for robotics returns, but reminders of the limits of imagination. Early investors may not have foreseen the outcome 10 years later, just as they could not have foreseen how many ups and downs would come in between. Li Feng ultimately leaves room for an “perhaps”: robotics still has a long road ahead, but the future we can imagine today may well fail to represent what robotics will embody 10 years from now.