Unitree Goes Public: 王兴兴's Luck, Investors' Lessons, with 田江川

Summary

• 田江川’s core postmortem: Unitree’s early funding crunch was its greatest stroke of luck. Without the capital to keep 王兴兴 in a comfortable domestic market, it forced him to go global on day one and serve top overseas labs deeply; flexible delivery, full payment upfront followed by a six-month wait, and gross margin that once reached 70% meant the company was profitable almost every year apart from one small loss — “not having raised that much capital was, for 兴兴, an extraordinarily lucky thing.”
• The decision to use motors instead of hydraulics was the starting point of Unitree’s generational moat. Motors cost roughly one-tenth as much as hydraulics, are far more standardized and can be scaled systematically; 田江川 compares the choice to “Apple turning a mainframe into a PC Macintosh that everyone could use.” It was not a technology innovation but an engineering paradigm shift — and at the time, “a decision that required tremendous courage.” “If I had a time machine, I would invest on the spot.”
• Unitree’s position is close to monopolistic: its hardware has become the industry standard for the world’s top labs. The Spring Festival Gala handkerchief-transfer project was co-developed by CMU, Nvidia and Unitree. Followers wait 3-6 months longer for open-source algorithms, which are not customized for their hardware, leaving them “crushed by Unitree on every dimension”; the remaining ways to break through are price and delivery time. Hardware is not winner-take-all, however, so room remains in specialized applications.
• The investor lesson: education bias and a “hand-built BP” caused Chuxin to miss Unitree at a RMB30M valuation in October 2017. One investor friend believed Unitree “couldn’t go far without visual algorithms,” yet the software capability was developed by labs building on its hardware. Chuxin later revised its framework to “judge heroes by what they do, not where they come from” and adopted a Poor, Smart and Desire model — “we are not fortune tellers”; investors should look at what founders have built, rather than define them in 30 minutes.
• 2025 will be a vintage year, and a bubble is inevitable. This is the origin of a technological paradigm, with the narrative controlled by a small group of players, and “the industry is still rewarding those willing to take bold risks.” Public markets are assigning leading embodied-AI companies a 40x PS “dream-market” multiple, sending secondary-market capital back into venture because listed targets are scarce. She rejects the idea that 2026 will be a shakeout year: the industry is non-zero-sum and hardware is fragmented, so, as in SaaS, a cohort of invisible champions will emerge.
• Cadence is Unitree’s coercive weapon: it consistently launches products 3-6 months before the market, visible but not yet reproducible by rivals. H1 was not the most complete product, but it captured the first-mover position in humanoids; Unitree moved decisively into humanoids in 2023 — “the humanoid is its B1 standing up” — with reusable components and an existing lab customer base. The Spring Festival Gala handkerchief transfer was the inflection point, demonstrating that “humanoid robots can do some things humans cannot.”
• The next phase of embodied AI begins when major players enter at 50,000 units shipped; the market expects humanoid shipments to exceed that level in 2026. A wave of autonomous-driving veterans is moving into embodied AI because “corner cases can be fatal” in autonomous driving, while many embodied settings have more tolerance for error; supply chains are reusable, and the addressable market is a RMB5T labor pool. Government funds are about capturing the “embodied dividend,” including the idea of taxing robot usage, while CVCs connect capital with top talent unwilling to settle into a large-company system — “capital is like water: it nourishes everything without competing for credit.”
• This wave of globalization is being led by China, and products have decoupled from brands. MIT classmates use DJI and Insta360 without knowing they are Chinese companies; the supply chain is a “systematic, overwhelming advantage.” Since 2023, Chuxin has invested only in companies global from day one. The cruelty of VC is that time is the ultimate constraint: a ski trip or a fishing trip can be enough to miss ByteDance or Pinduoduo.

Deep dive

1. Chuxin’s Starting Point: From “Gardener to China’s Richest” to the Beginner’s Mind

• 田江川 joined UBS investment banking in Hong Kong at 21. Her team called itself the “gardener to China’s richest,” having “created the richest entrepreneurs in a great many private-sector niches across China.” She later worked at a buyout fund and sold a company she founded to Warburg Pincus. In 2015, she concluded that “those experiences were too far removed from entrepreneurs” and founded early-stage VC firm Chuxin Capital.
• The name comes from the “beginner’s mind” championed by Steve Jobs: staying in love with and curious about the world. Chuxin’s slogan is “Walk alongside the different you” — an invitation to back nonconsensus founders trying to change their era.

2. Two Vintage Years: Cheap Nonconsensus in 2016, an Inevitable Bubble in 2025

• She calls 2016 the sweet-spot vintage. Technology was emerging in 2015-16 but “could not immediately be mapped to PMF,” leaving “a great deal of nonconsensus and many low-value assets that capital had not yet seen.” When she first met Unitree in October 2017, its valuation was only RMB30M.
• One sign of the times: companies that once called themselves automation collectively evolved into robotics, and now all call themselves dynamics. “The names themselves are an evolution of the era.”
• She is equally certain that 2025 will be a vintage year. This is “the origin of a technological paradigm”; mobile internet “was not even in the same league.” The bubble is inevitable because “the industry is still rewarding those willing to take bold risks,” much like the investors who backed DJI-affiliated teams in batches at valuations below RMB100M in 2023 and danced with the bubble.

3. 王兴兴 Was First to Take a Seat at the Table — and Was Never Opportunity-Driven

• She divides the field into engineers, industrial players spilling over from autonomous driving, and academics. Players are arriving from every direction, but 王兴兴 “entered extremely early and has stayed at the table”; he waited until the future he wanted arrived.
• That differs from the industrial players’ rational, granular dissection of Unitree’s shortcomings in household applications. 王兴兴 was more driven by conviction: the final line of his 2019 BP — “let robots into every household” — still moves her. Its timeline runs from drones to quadruped robots. “This group of people is less driven by opportunity,” which gives them the patience and resilience to stay the course.

4. Costa: The First Meeting and the Courage to Choose Motors Over Hydraulics

• The introduction came through an article in 机器之心 comparing Unitree with Boston Dynamics. Before the meeting, she sought a reference from a robotics lab at Harbin Institute of Technology. Hydraulic quadrupeds used in defense were expensive, extremely noisy and very large; the effort eventually went nowhere. 王兴兴’s motor-based route, by contrast, was “feasible from first principles.”
• The cost equation was compelling: motors cost roughly one-tenth as much as hydraulics, which also involve fuel. Motors are highly standardized; once developed in-house and produced at scale, they “can be shipped systematically in large volumes.”
• Her hindsight is that she recognized the 2 technical paths but failed to see that choosing motors was “a highly nonconsensus decision requiring tremendous courage.” “If I could take a time machine back with today’s knowledge, I would invest in 兴兴 on the spot.” She classifies it as an engineering paradigm shift rather than a technology innovation, comparable to “Apple turning a mainframe into a PC Macintosh that everyone could use.”

5. Why Chuxin Missed: Education, a Hand-Built BP and the DJI Question

• The first impression was excellent: “A highly ambitious, outstanding young person was doing something that could make the entire world marvel.” That stood out in an era filled with shared basketballs, shared pets and shared socks.
• But the eventual decision to drop the deal “was indeed somewhat related to his educational background.” And “if the market had a 100-point BP, his might have received a zero — it was arguably a hand-built BP. This was also Chuxin’s biggest lesson learned.”
• She also asked a naive question: “What happens if DJI builds your quadruped robot?” 王兴兴’s answer was: “If DJI builds it, we’ll be in real trouble. But DJI won’t — I’m the person responsible for this business at DJI.”
• The lesson became a talent philosophy: “judge heroes by what they do, not where they come from,” paired with the Poor, Smart and Desire framework. People from relatively poor families who are smart and ambitious are often better at targeting mass markets. She cites 黄峥’s view that someone not from an ordinary family might never have spotted the opportunity in Pinduoduo; a highly privileged, elite investor who starts a company may instead drift toward something narrower and more premium.

6. The Follow-Up Mechanism: Two Truths, Two Networks, and Term Sheets on the Spot

• Chuxin’s system is “two truths, two networks”: identify real problems and build a bench of real experts each year, then add a cognitive network and a trust-and-coordination network. In a hypercompetitive market, every Term Sheet needs to be prepared in advance and delivered on the spot. The only way to differentiate from other firms is “sufficiently strong conviction and the sincerity of being the first to recognize the founder.”
• Its vertical mapping of Alibaba found that the 2015 intake was the highest-caliber vintage, combining authority and responsibility in a role resembling an internal CEO. Chuxin invested in Deepexi at a RMB100M valuation; its earliest check has generated more than 100x, and the firm also found the founder of Yingdao when he was still a product manager. Its horizontal mapping of AI talent was similarly explicit: when investing in Leyan, her threshold was that if China had fewer than 10 people with NLP capabilities, she would invest.
• The breakthrough came from her habitual question: “What particularly investable projects or founders are around you?” 俞扬, a professor at Nanjing University’s AI lab, replied: “Have you looked at Unitree?” The lab had bought a Unitree dog and wanted more, but supply could not keep up.

7. The 2019 Signal and the 2021 Mass-Production Inflection

• The strongest signal was that the to-lab model was already supply-constrained: “It was clearly a Chinese company, yet Chinese labs could not buy its products because everything was being bought overseas.” Delivery required a roughly 6-month wait, leading her to conclude that Unitree was already profitable.
• The model was a PMF-before-PMF Kickstarter: customers paid in full upfront, then queued for delivery; gross margin reached 70% at one point. Apart from one year — possibly 2021 — when it lost several million renminbi, the company was profitable in almost every other year. “Capital is for scaling, not for helping founders with a halo keep searching for PMF.”
• That resolved her core question about young founders: “If he has already proven himself through this, then I think that is the best standard.” Sequoia was also preparing to invest. The launch of Go1 in 2021 was the turning point: Unitree began producing robot dogs at meaningful scale, brought many core components in-house and sharply improved product completeness after 5 years of co-development with top labs worldwide.

8. The Split with DEEP Robotics: DJI’s Full-Stack DNA Meets Ningbo’s Merchant Instinct

• DEEP Robotics and Unitree started in the same direction but gradually diverged. 田江川 attributes the difference to their “genes”: DJI DNA combined with the commercial instincts of Ningbo. “兴兴’s English is probably not very good, but he went overseas on day one.” By 2017, Unitree was already working with Google and Apple labs. She sees a generational difference: entrepreneurs born in the 1990s treat the entire world as their commercial territory, with a boldness the previous generation rarely had.
• 王兴兴 also turns self-developed core components into businesses. Unitree can source motors and lidar externally, and a friend who makes lidar even viewed Unitree as a competitor. That reflects DJI’s habit of turning full-stack technology into commercial products. His BP also benchmarked DJI: after drones, he would define quadruped robots.

9. The Dual-Currency Fund Logic of Going Global: Let the World’s Smartest Minds Develop on Its Platform

• Unitree’s lack of early capital pushed it toward a model resembling a dual-currency fund choosing between renminbi and dollars. Capable GPs generally prefer dollars because dollar capital is more generous and flexible. For 王兴兴, flexible delivery required targeting users who were more flexible, higher-value and more repeatable — users who were clearly overseas.
• The result was an ecosystem platform: “It is getting the smartest minds to conduct R&D on its hardware.” Most later embodied-AI startups no longer enter the research market because they would wait 3-6 months longer for open-source algorithms, and those algorithms would not be customized for their hardware. “You would be crushed by Unitree on every dimension.”
• In the United States in early 2025, she found many professors discussing DeepSeek and Unitree. “The reason the country eventually elevated them to this level was their international influence, not merely their influence in China.”

10. The Supposed Weakness — No Visual Algorithms — Became the Niche

• In 2017, she brought in an investor friend to co-invest. The friend believed that Unitree had essentially no reserves in robot vision and “might not go far without those software capabilities.” The outcome was the opposite: Unitree focused on hardware, while labs at MIT, Berkeley and Stanford bought the hardware, developed their own software and published papers. The hardware became “the industry’s hardware SOTA”: “Everyone turned the hardware into a fixed industry standard and competed on software and algorithm performance.”
• The main ways to break through are now price and delivery time. “There is almost no more room to cut price”; 众擎 has begun a price war, while “the overseas market still cares deeply about brands.” Efforts such as full-stack development and factory construction may leave competitors room in specific markets because hardware is not inherently winner-take-all — just as the new-energy vehicle market can support many brands.

11. Humanoids in 2023: “B1 Stood Up”; 6 Years in the Wilderness Created a Generational Lead

• Humanoids were a natural go-to-market move. The world’s best labs were already customers with large, recurring unmet needs around humanoids, while most of the quadruped’s core components could be reused. “You can think of the humanoid as its B1 standing up.”
• The wilderness had its own logic. Drawing on The Innovator’s Dilemma, she notes that few people worked on the category during the first 5-6 years. Unitree and DEEP Robotics were the main players in the same direction. “The road was difficult, but you also didn’t have that much competition.” Had the starting line been saturated with intense competition, Unitree might still have led, but not by the generational margin it enjoys today.
• Consumer-electronics giants left a window open. The 石头科技 and 昌敬 mindset was that “when a product ships fewer than 50,000 units, the sector is not even worth watching,” giving Unitree “a very long period to grow quietly.” At the Series B, many firms still hesitated to invest because of Xiaomi’s CyberOne, nicknamed Iron Egg.

12. The Details of Showing Up, and a Two-Tier View of Funds

• At 8:30 p.m. in Unitree’s office in 2023, she asked which roles needed support. 王兴兴 listed a long set of positions, then she asked who was currently doing them. The answer was always: “Me.” Later, he wrote that Chuxin had been “the investor that supported Unitree most firmly throughout,” a note she believes he probably wrote himself and may even have handwritten. He also attended Chuxin’s new-fund annual meeting to publicly support the firm. Many LPs who passed on the nonconsensus deal later regretted it deeply.
• Her fund hierarchy is straightforward: “excellent” means having 3 representative investments capable of generating 100x returns; “otherwise, you are basically wasting your time.” “Outstanding” means that, after generating returns, founders remain close friends, are willing to publicly support the fund and even become LPs to build the ecosystem. “Chuxin has not done very well on this front; we are still at a very early stage.”

13. The Gift of Scarcity: Not Raising Capital Was Lucky

• The host reflects that a decade ago and today each had its own luck and misfortune. 田江川’s counterintuitive conclusion is that “not having raised that much capital was a very good thing for 兴兴 — an extraordinarily lucky thing.” Had he had more money, he might have focused more heavily on China. Being cash-constrained forced Unitree to “go global very early” and serve overseas users deeply, which made the product so international.
• His early confidence on cost reduction also came from scarcity. He always believed that a startup “is itself a commercial enterprise and should be fully capable of making money and standing on its own.” A differentiated route forced in-house development; there might be no economies of scale early on, but “once volume comes through, it will definitely be highly economical at scale.” The supply chain also overlaps with drones and new-energy vehicles in many areas.
• The contrast with today is stark: a strong RoboMaster result today might prompt Chuxin to write a check immediately. But even a DJI intern transported back to that period would not necessarily command a high valuation — only faster fundraising. The value of the DJI brand as a reference was different 10 years ago than it is now.

14. Cadence as a Coercive Weapon: Always 3-6 Months Ahead of the Market

• Her observation is that “兴兴 is generally 3-6 months ahead of the market, launching products the market can see but competitors may not yet be able to build or catch up to.” H1 was not the most complete product, but he understood that the first priority was to seize the humanoid first-mover position and iterate afterward. “That is a trade-off.”
• On Xiaomi’s Iron Egg, she cites Elon Musk: competition may look intense because people are viewing the market statically; over a longer horizon, competitors may fail to persist. Iron Egg had no strong PMF and shipment volume was too low, making it easy to cut inside a performance-driven Xiaomi organization.
• The marketing was equally extreme. Footage of a robot being kicked became an industry role model. After DEEP Robotics released one flashy video, Unitree immediately followed with an even flashier version and quickly overwhelmed it on brand recognition. Boxing matches, IP collaborations and 王力宏 concerts showed the market for commercial performances. Her son’s birthday parties feature robot performances, and humanoids have appeared at friends’ weddings to present roses.

15. The Spring Festival Gala Inflection: Robots Did Something Humans Could Not

• 卫诗婕’s view is that Unitree’s first appearance at the Spring Festival Gala did not make a major splash because the dog was mainly performing. The humanoid handkerchief transfer created “a major conceptual shift,” making people realize that “there are movements robots can perform that humans cannot.”
• The technical backdrop matters. In 2024, only a handful of companies could make robots genuinely walk well and steadily. From 2024 to 2025, deep learning delivered a qualitative leap in robot motion control, while Unitree sat at the frontier of locomotion control. “At this highly watched turning point in the technological paradigm, the timing could only have been captured by it.”
• 卫诗婕 adds that the handkerchief-transfer algorithm was developed jointly by CMU, Nvidia and Unitree, calling it evidence that “Unitree is the Android of embodied AI.” 田江川 explains that Unitree is a universal hardware platform: top overseas labs and researchers develop on it, allowing the company to define the industry standard.

16. “Relieved I Didn’t Sell”: One Moment of Doubt, Then a Reality Check

• On the day Unitree went viral after the Spring Festival Gala, WeChat messages seeking existing shares came in at “40 or 50 a day,” without exaggeration. 田江川’s dominant reaction was “relief that I didn’t sell.” She had wavered in 2021 after buying a robot herself and finding that it was “completely unusable,” just as Xiaomi was preparing Iron Egg.
• The correction came on a Silicon Valley trip. Labs at MIT, Berkeley and Stanford were all using Unitree and publishing papers on it. “My feel for reality changed. That is also the meaning of embodied AI: I interacted with the physical world, and I revised my hypothesis from that year.”

17. After the Gala: Capital Spillover and Three Embodied-AI Factions

• The market structure changed as everyone tried to invest in Unitree. But Unitree was profitable, did not need the money and was selective; investors needed both industrial resources and credible capital backing. “When massive pools of capital cannot get into Unitree, they settle for the next best thing,” lifting the entire market.
• The factions are engineering-led companies such as Unitree and EngineAI, industrial players and academic teams. The industrial camp is led largely by autonomous-driving spillovers: trained teams with mature organizations targeting high-value applications directly. 卫诗婕 says one team may have shipped 6,000 units in only a few months. 田江川 compares these teams to divers: “Every movement is highly standardized and nearly perfect.”
• The autonomous-driving logic is straightforward: one reason autonomous driving is hard to commercialize is that “a corner case can be fatal,” while many embodied settings have more tolerance for error — folding clothes and blankets in a hotel, for example, is unlikely to be fatal. Supply chains are reusable, the market targets a RMB5T labor pool rather than a relatively fixed vehicle fleet, and veterans of autonomous driving’s short window may feel that “if we do not enter now, there will be no opportunity.”

18. The Capital-Market Mirror: Listings, Dream-Market Multiples and the “Ticket”

• Leading embodied-AI companies are pushing to list and capture momentum. The program also emphasizes Unitree’s uniqueness: it has exceptionally strong PMF, while “the vast majority of companies in the market do not have that condition.” Structurally, secondary-market capital is moving into early-stage venture. “The secondary market actually likes narratives more than the primary market. Nobody wants to miss the AI narrative, but the secondary market does not have many targets, and many of the companies listing are from the previous era.”
• How can a few thousand units support a valuation above RMB10B? Her answer is that the secondary market is assigning these companies 40x PS. The host calls it “dream-market valuation,” a multiple that rises even faster than P/E. A broadly accepted long-term view — that embodied intelligence will be primarily developed in China while serving the global market — is colliding with extremely scarce supply, “inevitably creating an inverted valuation structure.”
• Her explanation for not following up at the time is that “the speed at which foundation-model capabilities generalized into embodied AI was faster than everyone expected.” The industry was still framed around electrification and smartification, while foundation models were too capital-intensive. Today, “only people who are sufficiently smart, sufficiently ambitious or sufficiently accumulated from the past can get this ticket.”

19. What Unitree Got Right: 10x Better, Flexible Delivery and Product Obsession

• “Anything that is 10x better is an extremely strong signal and should be actively captured.” Whether it is 10x better performance or 10x more capital raised, the information content is a step-function lead.
• Her engineering intuition is that 王兴兴 likely grew up immersed in supply chains and hands-on assembly, giving him a feel for how far costs could be pushed down. Flexible delivery pulled PMF forward: “Some founders never find a PMF that customers will pay for. That is a classic way for a VC to lose money.”
• The product obsession extends internally. He interviews virtually every employee himself, with tailor-made questions and hands-on tests. “He is polishing the company the way he polishes a product,” a philosophy that fits MIT’s “Mind and Hand.”

20. 王兴兴: A Reserved, Sci-Fi Romantic — But Not a Top-1 Founder on First Impression

• Her three-part description is “an intensely product-oriented geek, a commercially minded businessman and a technological idealist.” Unitree means “the tree of universal technology.” But there is a contrast: “He is not particularly extroverted.” Casual conversation does not reveal the sci-fi romanticism behind ideas such as robots small enough to enter your bloodstream or large enough to build many other robots. That is why people can miss him. Judging only the person rather than what he has done, he would not rank top-1 among 10 founders; he is nevertheless “a genuine outlier.”
• That led to a methodological correction: “With many founders, we still need to look at what they have built rather than subjectively evaluate the person. We are not fortune tellers; it is impossible to define someone’s life trajectory in 30 minutes or an hour.”
• Her 6-dimensional framework points to decisions made correctly in complex environments, including the timing of Go1 and Go2 and the decisive move into humanoids in 2023; execution and focus, reflected in a lack of a lavish lifestyle and repeated refusals to attend Hupan and Hun Dun because he had no time; and continuous evolution. He began as a high-IQ, lower-EQ engineer, but at a recent dinner offered to pick up the bill and delivered rational insights on the upstream and downstream supply chain, showing the bearing of an industry leader. His resilience comes from love of the work: “Passion can withstand the long years.”

21. DJI, RoboMaster and “Clear-Eyed Ambition”

• Why do A-players spend time at DJI? The most respectful way to treat top talent is to surround them with other top talent, as in the Steve Jobs and Google examples. 李泽湘’s idea of putting a group of highly ambitious young people on an island follows the same logic. The key benefit is “taste in talent”: not knowing how to do everything yourself, but knowing what the best looks like and how it happened. That gave her the confidence to back DJI outliers with blank passports who nevertheless built global products.
• The paradox is that RoboMaster “may have cultivated DJI’s biggest competitors; it is a hotbed for creating competitors.” Talent density is high, and participants form startup teams during the competition. Many now choose entrepreneurship first or are persuaded by investors to start companies, creating a gap with DJI’s original intent in launching the competition. 昌敬’s advice for identifying young talent is simple: “Trace them back to when I was in college competing.”
• Chuxin responded by establishing a Shenzhen fund with RMB5M tickets. Its mandate is “you cannot have everything at once”: look for founders like 王兴兴 whose direction appears to have no commercial opportunity and whose market is tiny. “So what? If they build a product that makes the whole world marvel, you should invest.”

22. The Arrogance of Anti-Elitism: Silicon Valley’s “Bad Kids” Are Elites Too

• She disputes Silicon Valley’s taste for “disruptors, rebels and bad kids.” The United States is a highly elitist society, and its “bad kids” may be dropouts from MIT or Stanford. Profiles such as 王兴兴 and 俊杰 of 霸王茶姬 — “I could invest even if he had only finished middle school, but not finishing primary school would genuinely surprise me” — are precisely the profiles dollar funds’ talent filters can miss. Chuxin later formulated its “three non-” framework: nonconsensus, nonordinary and non-mainstream founders, who often have an edge in frontier innovation.
• One online theory holds that 王兴兴 could have attended Zhejiang University but performed poorly in English. 云安, the founder of the same-school tea chain 古茗, joked that he too could have gone to Zhejiang University but for a bad English score. 田江川’s conclusion is that English may itself be a signal of an outlier.
• She agrees that “excellence can be the enemy of greatness.” Highly accomplished people are consumed by excellence, and the opportunity cost of walking away is higher. Many talented investment bankers may remain bankers for life as salary and bonuses keep rising. Her own path ran from Olympiad math at Guangdong’s best school to becoming an internet-addicted teenager, after which her parents sent her to the UK to cut off the internet; her favorite Gu Long novel was The Happy Heroes, and she sees herself as a “female knight.” Some founders choose Chuxin partly because she has a reputation for bringing help in the hardest moments.

23. The Plain Logic of Government Funds and CVC

• At the national level, the question is: “If you believe the future is a world in which humans and machines coexist, should the population dividend become an embodied dividend?” Over dinner with 王兴兴, they even discussed a possible change in taxation: instead of taxing based on population, tax based on how many humanoid robots are in use. The identity of the taxpayer becomes a question of social ethics.
• At the CVC level, large companies cannot necessarily absorb the best people. “Ambitious people cannot be content within a particular system, regardless of whether the system itself is good.” Capital becomes the way to connect with them: “capital is like water; it nourishes everything without competing.”
• 田江川 considers JD.com and Meituan the most aggressive. In her view, JD.com has been unusually active this year and may have invested in more than 10 companies in short order. Domestic sales of embodied products depend heavily on JD.com’s channels because after-sales support is demanding, returns may be involved and users place a high value on trust. JD.com can capture incremental demand and feed it back to startup teams, while Meituan has a rich set of application scenarios.

24. Embodied AI’s PMF Has Not Arrived — and the Debate over a 2026 Shakeout

• The central reason To C lacks PMF is that manipulation is not yet good enough. “If it could provide a more tangible function, such as going to the refrigerator to bring you a can of Coke, I think consumers would at least pay the price of a robot vacuum for it.” The gap is in the brain and multimodal capability, but “I believe this can be achieved.”
• There is also a clear path forward, similar to SaaS: labor substitution and efficiency gains. “This is not an especially winner-take-all market. There will be many invisible champions in specialized fields” — companies that are not glamorous and may not have formidable technical barriers, but can generate strong profits.
• She rejects the idea that 2026 will be a shakeout year. “It is certainly difficult to build another full-stack hardware company now, but the industry as a whole still has enormous incremental demand and growth that can support many companies.” This is not a zero-sum game. Unlike foundation models, where software can be winner-take-all, hardware is inherently fragmented.

25. The US-China Landscape: Open Source and Closed Source Will Remain the Two Main Lines

• For the hardware body, “almost everyone is looking to China”; the supply chain is a structural advantage. Silicon Valley still has an opportunity in the robot’s “brain,” but the paradox is that embodied-AI scaling laws depend on massive data volumes: “The companies shipping the most units may hold the most data, and the data can in turn improve their brains.” The path to collecting data has not converged, so Chuxin has backed both Tsinghua-linked academic teams and industrial teams combining software and hardware.
• 卫诗婕 discusses Jensen’s GR00T and his emphasis on real-world data, including 270,000 hours of embodied data and an incremental 10,000 hours every week. 田江川 believes the industry has reached consensus that “data has become exceptionally central.”
• The United States is following the old script from The Cathedral and the Bazaar: “Under every major opportunity, 2 paths will always emerge — one open source and one closed source.” Tesla’s Optimus could be radically closed, like the iPhone, while Figure may offer a more open ecosystem, like Android. These will form the 2 main lines.
• Among China’s 3 leading players, AgiBot and Galbot are stronger in models, while Unitree is stronger in hardware. But “the landscape is certainly not fixed.” Major industrial players will enter, including BYD and Xiaomi, and potentially Meituan in the future. Startups will need to keep leading, and the pressure will be intense.

26. China-Led Globalization: Products and Brands Have Decoupled

• Her view is that “this wave of globalization is being led by China, so there will inevitably be a great deal of friction between China and the US. China benefits from this globalization, while US interests may be harmed, which is why the US advocates deglobalization.” Consumers are voting with their feet: her MIT classmates come from 40-50 countries and use DJI, Insta360 and Pudu without knowing they are Chinese companies. In the US, she also saw Xiaomi cars being chosen despite high tariffs.
• The foundation is comparative advantage: the combined innovation of China’s supply chain, engineers and talent dividend. Borrowing Warren Buffett’s farewell line — “I am grateful to have been born a white male in the United States that year” — she says: “I am grateful to have been born Chinese in China.”
• The operating principle follows: since 2023, Chuxin has required every software and hardware company to target the global market on day one. Most VCs now have little interest in China-only projects. Chinese giants may buy teams rather than acquire companies, while offering neither teams nor VCs much of a premium, preventing a virtuous ecosystem from forming. The biggest risk in going overseas is geopolitics: tariffs can quickly be passed on to consumers, become inflation and prove unaffordable for the US or other countries.

27. Incubation, Conviction and “Management Is a Science, Not an Art”

• Influenced by Sutter Hill’s incubation model, she breaks innovation into technology, product and capital — both talent capital and money. The key catalyst is conviction: “Only when you are truly certain can you convince a founder to pursue a direction and convince excellent people to join.” Entrepreneurship and investment share the same first principles; only the proportions of the inputs differ. Chuxin’s differentiation is “the ability to make innovation happen from 0 to 1, not merely to assist innovation.”
• MIT gave her a sharp realization: “For the first time, I truly felt that a technical role and a management role are 2 different things. I felt like I was running a company as a naked CEO.” In the US, management is closer to being licensed — a Microsoft Product Lead taking on a larger role might be sponsored by the company to attend an EMBA. In China, many managers are “unlicensed, managing by instinct and improvising blindly.”
• Management as a science shows up in exercises such as spending 2-3 hours discussing each person’s “user manual”: strengths, weaknesses, how to handle conflict and what they find unacceptable. She initially thought the exercise was less useful than copying ByteDance by sending out the manual and asking everyone to mark it read, but collaboration improved substantially afterward. The same applies to cultural scales: Koreans, Indians and Chinese are near the top in tolerance for direct feedback. Conflict may be cultural friction rather than a character flaw. “Even an institution as small as ours can feel why China and the US are decoupling.”

28. VC Is Brutal Manual Labor: Time Is the Ultimate Constraint

• The industry truth is that “for every investor, the biggest constraint is time. You may miss a huge company because you were briefly lazy at the moment you first thought of it.” She uses skiing to explain missing ByteDance and fishing to explain missing Pinduoduo. She spoke online with 俊杰 of 霸王茶姬 for more than 2 hours, but because she was in Beijing could not verify him in person; she sent her partner Max’s father to a store in Yunnan to taste the tea and provide feedback. “His father was not a typical milk-tea user,” making it an invalid sample. On another occasion, a meeting with 徐卓 of Pigeon was postponed at the last minute; Pigeon’s earliest round went to Sequoia.
• The GP generation is changing. As dollar capital dries up and renminbi capital takes the lead, “the market needs younger GPs” — a line she borrows from 开心消消乐’s 海宁. Succession splits into 2 paths: some recognize that “VC is me; after I retire, nothing remains, and LPs are investing in the key person,” while others build succession systems that preserve the soul of the brand. Chuxin is “relatively lucky” this cycle, with listed winners including Deepexi, Unitree, EHang and Geek+. 田江川 expects Unitree’s return to have a good chance of exceeding Deepexi’s. “Without this much luck, we would certainly face a shakeout.”
• Her self-critique is that Chuxin identified Unitree as an era-defining company but “seemed to do nothing afterward.” Today she would do more because “the meaning of bringing help in the hard times is always greater than adding polish when things are already going well”; the 2017 check may have mattered more than every subsequent check. In this industry, “talent and hard work are both indispensable.” Many results that look like accidental successes “actually come from hard work and from refusing to let opportunities pass.”