132. A 3-Hour Interview with Xinghaitu Founder Gao Jiyang

Summary

• Xinghaitu has just closed a new financing round at a RMB10B valuation, roughly 30x higher than two years ago (about RMB300M after its January 2024 seed round). The round was led by CFO Tianqi, with strategic investors including Geely and BAIC, alongside multiple PE funds and public-to-private crossover investors; six existing shareholders followed pro rata, with three to four making super pro rata investments. Gao Jiyang rates the valuation “conservatively, fifth” in the market, behind 智源, 银河, and 宇树.
• The company’s strategy is Tesla-style vertical integration: full-stack hardware, a real-world data loop, and end-to-end models. “The long-term moat in embodied intelligence is built on a data loop with the physical world.” The robot is both the data carrier and the product form—hardware plus intelligence. Algorithms diffuse in only 2-3 months, while the hardware supply chain takes 12-18 months and the data system another 6-12 months—the algorithm moat is relatively small. “It invests heavily in innovation, but the barriers preventing imitation are very small.”
• The real-world data cost ledger is the hardest number in the entire episode: one hour of real-world data costs RMB200-250 all-in, based on 3-4 hours of labor plus depreciation on a robot priced at RMB100,000 over 1,000 hours. 100,000 hours—roughly the amount of time one person interacts with the physical world before age 18—costs only RMB25M. Data acquisition costs are 1:5 to 1:10 versus training costs, meaning low-quality synthetic data wastes money at the training stage. Nobody knows the right data recipe: “AI is ultimately an experimental science. You have to try it out.”
• The commercialization path runs from the developer market to the productivity market (with Macintosh, 拓竹, and 宇树 as reference points). Xinghaitu already has more than 150 developer customers globally. Its 2026 focus is shifting to applications, with bin picking and in-plant logistics SPS among the “pick anything, place to somewhere” scenarios it favors. His life bet: “ship 10,000 units in productivity applications.”
• Xu Huazhe’s departure was settled internally last August. After Zhao Hang took overall control of the foundation-model team and delivered a new out-of-the-box machine, Xinghaitu will invest in Huazhe’s first startup round focused on to C home applications. Gao’s characterization is blunt: “This is an unequivocal long-term positive for Xinghaitu,” rooted in a “fact-based, merit-through-battle culture.” But he rejects the idea that this means algorithms do not matter: “Algorithmic innovation cannot exist independently of the company’s overall infrastructure.”
• Waymo and Momenta were two sides of the same career lesson: Waymo was “an engineer’s paradise” but had “no founder,” lacking a concentrated top-down force; “Being wrong isn’t scary. What’s scary is having no concentrated, unified force.” Momenta was absolutely results-oriented, with the core algorithm team working until midnight and six days a week as the baseline. He gave up roughly $10M in options when he left to start Momenta, and “didn’t feel bad about it at all.”
• The real threat from Big Tech is business synergy, and embodied intelligence happens to have none: supply is entirely new—none of the auto-parts supply chain can be reused—and demand spans thousands of industries with little obvious coordination. Big Tech’s biggest gap is data, but buying data creates a paradox: “You need a company that understands foundation models to define the data system.” That creates a structural opening for startups with full-stack hardware capabilities.
• Why has embodied intelligence produced no Liang Wenfeng-style technological romanticist? “It may well be that robotics itself simply isn’t romantic.” The chain is long and the cycle is long: “We are naturally forced to go into the dirt and build a lot of things. There is no room for romance.” “This industry does not allow people like that to exist; if one does, they may suffer a great deal.” The robots in the videos look much better than they do in reality. “People still need to give it more time, but it will happen.”

Deep dive

1. The Opening Question: Why Does Chinese Embodied Intelligence Have No Technological Romanticists?

• Zhang Xiaojun’s opening question runs through the entire episode: why has embodied intelligence produced no one with the “heavy technological romanticism” of Liang Wenfeng or Yang Zhilin? She casts Gao Jiyang as the opposite of that romanticism—“extreme efficiency, engineering decomposition, and pragmatism.”
• The timing of the recording is itself a signal: co-founder Xu Huazhe is about to leave. Gao’s opening line becomes the episode’s central thesis: “Robotics is an industry with an extremely long chain. Sometimes you have to stick your head into the dirt.”

2. The Shijiazhuang Teenager: Diligence as a Switch You Can Turn On

• At Shijiazhuang Yudong Primary School and No. 27 Middle School, he was an ordinary student whose grades ranked “roughly fifth to tenth in the class.” During the summer before sixth-grade placement exams, he worked hard for the first time and ranked third in the entire grade—“the first time in my exam career that I did pretty well.”
• His own self-portrait is precise: “When I realize diligence is required, I become extremely diligent. When it isn’t required, I’m not particularly diligent.” The host summarizes his life as a series of “sprints,” which he accepts was true at least through college. After graduation, “my state changed—I stayed diligent all the time,” because “my sense of purpose became stronger.”

3. Tsinghua via Physics Competition: Summarization, Not Talent

• He studied physics competitions in the provincial science experimental class at Shijiazhuang No. 2 High School. “I’ve met genuinely talented people. Compared with them, I mainly relied on hard work”—if others solved a problem once, I solved it twice; if they solved it twice, I solved it four times. He mapped problem types to examinable concepts and built systematic summaries, a methodology that carried into every later exam and engineering problem.
• A national second prize was just enough to let him choose a major at Tsinghua. At the November 2010 competition in Xiamen, an admissions officer called the hotel room and told students to line up: “I’ll ask you 2 questions: Do you want to go to Tsinghua? Pick a major first.” He chose electronic engineering for a simple reason—“I thought chipmaking had a future, but I hadn’t thought it through very deeply.” His Chinese and English were poor—about 100 points combined across the two papers, while others scored more than 90 on one paper alone. “Without the physics competition, I would have been happy to get into Sichuan University or Xiamen University.”

4. Tsinghua’s Genius Education: Yang Zhilin as an Unreachable God

• Asked about his classmate Yang Zhilin, he says: “He was an unreachable god. Really. He had always been excellent, close to winning the top prize, and had already done outstanding work as an undergraduate.” Where did that ability come from? “I think it was genius.”
• Gao ranked in the “top 30-40%” of the electronics department. His classmate Han Yanjun was another level: “I once looked at his homework to learn from it, but I could barely understand it. I had to find someone else’s homework to use as a reference.” Tsinghua taught him “to be humble and low-key—to understand that there is always someone better.” His prematurely white hair also began there in his sophomore year. “It wasn’t dyed. It turned white naturally.”

5. Entrepreneurship Was Decided in College, but Mobile Internet “Had Nothing to Do with Me”

• He had decided in his second or third year of college that he would start a company. From 2011 to 2015, mobile internet was booming—campus social networks, ride-hailing, food delivery. “I vaguely sensed this was a huge opportunity, but it had nothing to do with me.” His reasoning was clean: “Every generation has its own opportunities. They had accumulated a lot of things, and it was their turn. I was just a college student who knew nothing and could do nothing. Even if I wanted to try, I couldn’t win.”
• The question therefore became: “What exactly should I build, and what should I do to prepare for it?” This reverse-engineering approach became the template for all his later career decisions.

6. The PhD Setback and Zeng Guofan: From Confucian Purity to Practical Achievement

• After a Stanford UGVR internship during the summer of his junior year, his professor did not write a strong recommendation. Every school he wanted rejected him; only USC, UCSD, and others that had not received recommendation letters gave him offers. It was “a small blow, a low point.” He responded by reading and reflecting: “I read history and biographies, looking for inspiration in other people’s difficulties.”
• What stayed with him was Zeng Guofan: from a Confucian purist in his twenties to a man of extraordinary practical achievement in his forties. He carried one realization from the book directly into his life: “When you want to do something, how many resources can you mobilize, how many people are willing to do it with you, and whether you can ultimately get it done—that is what matters most.” Applied to himself: “I couldn’t remain a Confucian purist—I couldn’t get into the best schools—so I had to think from the bottom up about what I actually wanted to do, and what to take as my goal, rather than treating superficial markers as the goal.”

7. The SenseTime Internship: A Neural-Network Revelation on a Bicycle

• In his senior year, he obtained an internship through a small-group discussion with Tang Xiao’ou. At the end of 2014, he joined SenseTime and trained its first neural network for pose estimation, working with Li Cheng and Lu Shu. He still credits SenseTime for “giving someone who had never encountered this a chance to learn.”
• The moment remains vivid. Cycling out of the Tsinghua Science Park startup building, he suddenly thought: “A computer can extract patterns from data by itself. We program by extracting patterns ourselves and writing them as if-else statements. A neural network extracts patterns from data itself, turning if-else statements into parameters, fully automatically. This is incredible. I have to do this in the future.” His conclusion at the time was that “people will probably do very little programming in the future.”

8. Finishing the PhD in Three and a Half Years: Three Ways to Publish at Top Conferences

• The goal came first: “I was doing a PhD to complete certain training, and the faster I completed it, the better.” He set a four-year graduation target, worked backward to 4-5 top-conference papers, and determined that he had to submit to CVPR by the end of his first year. He found that his “idea bandwidth exceeded my execution bandwidth,” so he shared ideas with labmates and co-authored papers, “putting everyone’s underutilized time to work. It was good for everyone.”
• His summary of paper-writing is that “almost every paper fits into 3 categories: first, the hole-digging type—raising a new question, building a dataset, or creating a benchmark; the strongest people do this. Second, improving performance. Third, achieving similar performance with lower cost or less labeled data.” He also used the simple tactic that “if others submitted 1 paper per cycle, I submitted 2, increasing the odds.” After 3 years, his Indian supervisor, in his 70s and “extremely nice,” let him graduate at the end of 2018.

9. Choosing an Industry: AI Had to Be the Industry’s “Absolute Variable”

• He researched industries the way he would research a job. His filter was: “The industry must have AI as its deepest variable—without AI, the industry would not exist; with AI, the industry comes into being—and it must be large enough.” SenseTime-style general-purpose AI plus services failed the test: “Would it turn into outsourcing? It would be hard to form a product, and delivery costs would remain high every time.” Advertising also failed: search and advertising work without AI, so AI is not an absolute variable. Cloud failed because “wrapping a model in an API is not fundamental.”
• His conclusion was “AI for the physical world,” with autonomous driving as its first form. Why not traditional robotics? “Control, optimization, SLAM, and similar technologies make robotics more machine-oriented than human-oriented. It is hard to make a robot work like a human from the ground up. If you want it to become human-like, you still need AI methods.” It came back to the same point: “The magic of AI is that it can replace humans in extracting patterns.”

10. A Decade of Autonomous Driving: The 2018 Architecture Was No Different from 2008

• Before joining Waymo, he read most of the papers since the DARPA Challenge. After joining, he especially liked reading Waymo’s codebase and its historical versions. His central discovery: “The architecture of autonomous driving in 2018 was no different from that of 2008—perception, localization, offline mapping, planning and control. The large pipeline had already been explained in 2008 papers.” The modules gradually shifted to neural networks, such as lidar moving from point-cloud clustering to network inference.
• But the underlying logic was robotics rather than AI. Robotics decomposes systems and focuses on corner cases, so it breaks the system into interpretable modules to solve individual cases. “AI’s methodology is data-driven and end-to-end—it is not good at solving 1 or 2 specific cases; it is good at moving a benchmark from 80 to 90 to 95, even if some cases get worse.” Waymo’s perception stack at the time had “literally dozens of models”: detection, tracking, multiple classifiers, and layers of scene understanding. Tesla began an AI-native redesign from 2018 to 2020, unifying perception in one model. “The evolution to BEV, end-to-end, and VLA has become mainstream today.”

11. Waymo’s Core Problem: Not Big-Company Disease, but “No Founder”

• Waymo had about 1,000 people when he joined and nearly 2,000 when he left. It “entered a big-company state too early, before creating value,” with its levels, engineering system, and culture all aligned to Google. But he says big-company disease was only the symptom: “The essence was that Waymo had no founder. Its founder was Google’s founder, who had no time to manage it, so the top-down force was missing. It was unlike Tesla, where Musk could say what to do and start doing it, even if he was wrong.”
• That became the starting point for his theory of organizations: “In this kind of industry, being wrong is not scary. You can iterate and adjust quickly; that’s fine. What’s scary is having no concentrated, unified force.” If Waymo had had a strong founder, what should it have done? “In 2018, it should have started another team with one infrastructure stack and one evaluation system, then redesigned the architecture AI-natively. That had nothing to do with whether it was an automaker.”

12. Four Business Models and Waymo vs. Tesla: The Difference Is Their Attitude Toward AI

• He divides autonomous driving into 4 business models: Waymo-style self-operated robotaxi fleets charging per ride; automakers selling cars plus software subscriptions; Momenta-style suppliers charging NRE plus licensing; and Huawei as the fourth—“in essence, earning the vehicle’s gross profit, redefining the car through top-tier assisted driving and cockpit experience, and adding brand and distribution.”
• He first corrects the host’s “science-driven” label for Waymo: “Waymo’s engineering DNA is as strong as Tesla’s. It came from Google, where I received my engineering training. The difference lies in their attitude toward AI and the speed and intensity with which they redesign systems around AI.” He went to Los Angeles last November to experience it: “It is much better than most Ubers. The business model is already close to working—it is an Uber for the AI era.” His respect for Waymo remains: “Once it chooses a long-term strategy, execution is very determined. There are still the right people inside. It was just slower to adjust.”

13. VectorNet: His First Collaboration with Zhao Hang, and His Codebase Habit

• His innovation at Waymo came while working on prediction. Maps had previously been rendered into images and processed with CNNs. “Convolution has a local field of view, while maps are long, so the information was not processed well.” They switched to vector representations of maps plus graph neural networks, using a lightweight self-attention operator inspired by Transformers but with a very light architecture because “we were given very little compute.” The results were strong. This became VectorNet, later adopted by many companies, and remains “work I’m quite satisfied with.”
• Zhao Hang remembers that when they were designing the approach, Zhao wanted to read the papers, while Gao said: “I’ll read through all our code over the weekend and know what’s going on.” Gao explains: “I don’t just like reading current code. I like tracing it from the past to the present—you can see an engineering system evolve, and there is logic in it.” Their complementarity was clear: “Zhao is more willing to think things through from first principles, while I look at the most effective solution from the problem’s perspective—one thinks from the supply side, the other from the demand side.” Zhao was “extremely emotionally stable,” while Gao “pushes hard when results fall short.”

14. Engineer Thinking Equals Decomposition Plus Measurement; Two Decisions That Took Him Away from Waymo

• After 2 years at Waymo, he says he “learned more than he contributed.” First, he learned how physical-world AI systems work across the cloud and edge, including their historical evolution. Second, his engineering mindset took shape: “decompose and measure—break a complex problem into subproblems, then keep breaking it down until it becomes lines of code; measure until it becomes unit tests, then work back up layer by layer to the top-level metrics.” The difference from physics competitions was that competitions were problem-solving games rather than systematic engineering, though both trained logical thinking to a high level.
• In the second half of 2020, he concluded that “my growth efficiency here was converging” and realized “I was too far from products and too far from operating a company.” He therefore made 2 decisions: build for mass production rather than robotaxis because they were “too slow,” and return to China to join a place with stronger autonomous-driving and AI capabilities. The candidates were Huawei, where he spoke with Chen Yilun and Su Qing, and Momenta. He did not choose Nio, Li Auto, or Xpeng: “Xpeng was already doing reasonably well. I like going to places that are not doing well and making something work.” His definition of a product also had no to C requirement: “If you create value directly for certain users and get them to use it, that is a product. To B is also a product.”

15. Choosing Momenta: An Organization Needs Someone Who Can Correct Its Errors; Deriving the Mass-Production Flywheel

• His reason for choosing Momenta was simple: “It had no big-company disease and was forceful enough. An organization can succeed while making mistakes, but someone has to say we were wrong and then change course. You cannot have nobody willing to correct errors.” Cao Xudong was “technically strong, a believer in AI, and unwavering about long-term goals.”
• He reconstructs Momenta’s mass-production path as a data-driven deduction: “AI definitely needs data. If you operate your own fleet, 1,000 cars is already a lot, and even covering 1 city is difficult. Autonomous driving needs to work everywhere. That naturally leads to mass-produced cars—install the software on production vehicles, first create value through assisted driving and parking, and form a cycle of data and commercial value, turning data acquisition into a commercially driven activity.” In 2018, when others were still talking about directly building robotaxis, this was “visionary and bold,” and Momenta stayed committed.

16. Being the Catfish: Rotating Through Modules to Build Transferable Ability

• Asked whether he was a catfish brought into the company, he says: “I probably was.” He joined at the end of 2020, just as Momenta secured SAIC IM’s first mass-production designation, and was “used flexibly”: perception, localization and parking, other work, then planning and control, where he helped convert both planning and localization from rule-based systems to deep learning, and finally delivered the high-speed and elevated-road NV production system to SAIC.
• The most important lesson was not solving any specific module: “It was that I could enter an unfamiliar field very quickly, use a fixed methodology to understand and decompose it, and then match people to the work. The match is not correct on day 1—you monitor and measure, expand what is working, contract what is not, and adjust.” Challenges made him “overwhelmingly excited,” for a clear reason: “If I cannot do this well, I have no chance of building a successful company.”

17. Two Firsts: Without Going to War, You Cannot Temper an Organization

• Momenta was facing “2 firsts” at once: turning something at demo level into a product-level system for the first time, and completing service delivery to a large B-end customer for the first time. The organization was inevitably unready. The architecture was not designed for mass production, and the team’s capabilities did not match the required intensity. The company therefore adjusted frequently, with both active and passive attrition and heavy personnel turnover.
• His assessment is unequivocal: “All of this was good. Without being tempered by battles like this, there is no way to train and temper a team.” An earlier adjustment, when Shaoqing—also called “Shengqing” or “Shaoxing”—and Wang Fan left, was in his view “Xudong changing the company culture—from a research lab into a company that truly builds products. Without changing the organization, you cannot secure production designations.”

18. Paradise and Battlefield: The Two Sides of Waymo and Momenta

• Waymo was “an engineer’s paradise”: the best infrastructure, the best colleagues, and generous, warm managers. “There were goals, but they weren’t coercive.” Some people left at 4 or 5 to pick up their children. Gao himself “worked overtime automatically,” reading all of Google’s documents and codebase after finishing his normal work.
• Momenta was “absolutely results-oriented: good results meant moving up, bad results meant adjustment, and the pressure was constant.” Directors and above faced customers directly, while China’s automaker culture was tough. “They would yell at you. Have you ever been yelled at?—That was completely normal. I didn’t care much. If they yelled, they yelled.” He heard things like “If you cannot do the job, we’ll replace you immediately” and “Bring Xudong here to explain.” His conclusion has 2 layers: “Purely from an engineer’s perspective, Momenta and China’s assisted-driving environment were not good. But for growth, China’s assisted-driving environment trains an engineer’s overall capabilities better—it lets you see the real state of the world. I prefer facing reality, even when the truth is harsh.”

19. Cao Xudong’s Two Sides: Saying the Truth Out Loud, and When Pressure Becomes Harm

• His highest praise for Xudong was strategic: “Xudong’s strongest point has never changed: strategic ability—he decided very early to pursue mass production and never wavered despite setbacks. That was Momenta’s most important advantage over so many autonomous-driving companies.” He compares him with Horizon Robotics’ Kai Ge: “Both had exceptional strategy and execution. The chip decision was made 3 years earlier; 1 year later and today’s outcome would not exist.”
• He is equally direct about the other side: “Xudong is pushy and aggressive—actually, not aggressive. He says the truth and expresses it in a very direct way, and when that pressure reaches a certain level it becomes harmful. Many colleagues left because of it. Later he selectively stopped showing that side to as many people.” Is Gao like that? “I have that side too.”

20. Giving Up Roughly $10M: There Was No Way to Keep Him

• He gave up all of his options when leaving Momenta to start a company. “At that point, it may have been $10M; at today’s point, it might be more.” Did he feel bad? “Not at all. What I care about most is the thing I want to do. Compared with that, everything else is not worth much. I cannot spend much money in daily life anyway.”
• Could Momenta have kept him? “Ultimately, it could not have kept me, no matter what, because this is my life’s mission.” As for why Momenta could not retain others, he first took responsibility and then offered context: “It could have done better at retaining and continuously developing truly top talent. But you cannot blame Xudong either—the industry is like this. The core algorithm team worked until midnight, with 6 days a week as the baseline. Working hard is one way of taking responsibility for everyone.” Momenta’s overtime intensity is still greater than that of Xinghaitu today.

21. Three Signals at the End of 2022: Time to Start

• Three signals aligned. First, GPT-3 and InstructGPT: “They made the world believe in AI again. It is not enough for only AI people to believe; more people have to believe before capital comes in.” Second, mass-produced assisted driving made edge intelligence possible: “The edge compute and sensors robots use are not very different from those in assisted driving.” Third, Tesla formally announced a humanoid robot.
• In December 2022, he had just turned 30 and was a Sagittarius. “All of these things came together, and I decided I definitely had to start.” But he still finished what he had started: delivering NV into production for SAIC in March or April 2023, test-driving every domestically sold car with NV functions to confirm that “it was doing reasonably well,” and only resigning in May. Finishing properly was itself part of his methodology.

22. Momenta’s Foundational Lesson: What It Means to Be Customer-Centric

• The biggest change in his 2 years back in China was learning what “customer-centric” means. “It does not mean mechanically doing whatever the customer says. It means genuinely looking at the need from the customer’s perspective, even helping uncover the need and proposing a better solution.” This applies internally too: every upstream team serving a downstream team must be customer-centric—support teams toward business teams, platform teams toward delivery teams.
• He names the opposite bluntly: “Being centered on your own growth, or on so-called technological leadership—at least when you are building a company, these are both wrong.” This is the key to understanding what follows about Xu Huazhe.

23. Starting Up: An Embarrassing BP and a Quickly Abandoned Delivery Robot

• After resigning, he drove around Tibet, then began preparing the BP and raising money in August. “Looking back now, that BP was embarrassingly bad.” But 2 things were clear from the beginning: embodied intelligence had to mean full-stack hardware plus intelligence, not intelligence alone because the long-term moat lay in the physical-world data loop; and the company could not be pure research—it had to create value in the real world.
• The first direction was last-mile delivery robots, essentially copying autonomous driving into delivery with some manipulation capability. “We rejected it very quickly.” In February or March 2024, he reviewed it himself: “This cannot work. It is too early. The hardware is immature, the supply chain does not exist, the intelligence is not ready, the customers are immature, and the market is not there.” The company shifted to wheeled dual-arm robots focused on manipulation and entered the developer market. “We were not as clear as we are today, and there was some luck involved. We got it right.”

24. The Angel Round: RMB30M, a RMB200M Valuation, and Zhu’s Exit

• The first round totaled RMB30M at a valuation of roughly RMB200M. IDG led, after classmate Li Yikang made the introduction, which led to Shao Hui and then Xiao Jun. Baidu Ventures followed. GSR invested the least; the deal was handled by Yutong, a senior alumna who later joined Kimi. After she left, “Boss Zhu exited. He had not invested much. If he wanted to exit, he exited; I was not particularly concerned.” After investing, IDG added: “What you want to do probably will not work. Think it through again.” That is what an angel investor does: accepts your mistakes and imperfections.
• Against today’s environment, where “you can casually tell a story and start at $200M,” he rejects the idea that the early valuation was cheap. “The angel and VC environment was much colder then. In 2023, robotics was not yet consensus, and people did not even understand what ’embodied intelligence’ meant. They were willing to invest, and I was already extremely grateful.” A fund from Tsinghua’s electronics department led by Professor Wang and Yao Song then added nearly RMB20M, bringing the post-money valuation to RMB300-400M.

25. Why Start with the Full Robot: Without Hardware, Algorithms Are Castles in the Air

• The host’s question goes to the heart of it: as an autonomous-driving AI team, why begin with the unfamiliar task of building the full robot? His answer is a double reverse-engineering exercise: “The long-term moat is built on a data loop with the physical world, so I must have a carrier for the data. In the short and medium term, our product is unlikely to be an algorithm or a so-called brain. It will probably be a physical entity combining the full robot and intelligence. Both points lead to the same conclusion: we had to build the hardware and supply chain.”
• Investors asked the same question throughout 2024, and his answer never changed: “If you directly build algorithms, those algorithms are castles in the air. They cannot become product value or commercial value.” However hard the start, “you cannot be afraid of moving slowly; you have to keep moving one step every day.”

26. Embodied Intelligence vs. Cars: Fortune and Misfortune Are Two Sides of the Same Coin

• Autonomous driving’s fortune and misfortune come from the fact that the auto industry has existed for 100 years. “The good fortune is that customers and commercialization paths are extremely clear—roughly 20 automakers globally. The misfortune is also there: the automaker sits between you and the data loop is not smooth. You cannot control the vehicle, and the experience delivered to customers may be discounted.”
• Embodied intelligence has the opposite pairing. “The misfortune is that there is no carrier. You have to build one yourself, overcoming the discomfort and learning how to do it. The fortune is also there: once you truly complete this step, you can serve thousands of industries downstream.” It is like taking the auto industry back more than 100 years, while also adding intelligent technology; both must be built at once. That is why Xinghaitu naturally has 2 tracks: a talent-density-driven intelligence team and a process-driven hardware supply-chain team.

27. Starting from Almost Zero: Finding Suppliers on Taobao and the Genius Yang Zeyi

• The company’s early awkwardness is preserved exactly. It bought other people’s products and disassembled them, “seeing a joint module for the first time.” When they found an unfamiliar part, they photographed it and used Taobao’s image search to find the supplier. They did not even know how to disassemble the machines. The structural lead later visited, thought “we were too pitiful,” and gave them a toolbox containing screwdrivers, a hammer, and tweezers.
• The turning point was Yang Zeyi, introduced by a FiveYue investor who did not ultimately invest. Born in 1997 and educated at SUSTech, Yang had started his own robotics training business. “When I spoke with Zeyi, it was the first time someone had explained the full robot system clearly to me. It was systematic and insightful. When you feel that way in an unfamiliar field, you are probably on the right track.” After several conversations and 2 trips to Shenzhen, he joined in January 2024 as chief electromechanical engineer. “He has overseen the broad architecture of all our products since then. He is incredibly hands-on—simply a genius.” His equity was granted “roughly in percentage points of the original shares.”

28. The Partner Mechanism: A Medium Hexagon Plus a Larger Hexagon

• His team-building theory is: “I want to be a hexagon of medium area myself, and I want the founding team to cover a larger area, forming a larger hexagon. Then the team is strong and balanced across the board.”
• The mechanism is continuously open. “It has existed since the company was founded by me, Tianwei, and Zhao Hang. Zeyi, Huazhe, commercial head Yu Lei, and CFO Tianqi, who joined a few months ago, all came in this way. I am relatively willing to give away equity. If you want the company to grow 10x or 20x in 5 years, every stage is a startup, so you need a mechanism that continuously absorbs good people. That is one reason we have managed to stand out slightly from a large group of companies.”

29. The Configuration: Not a Compromise, but “Intelligence Defines the Body”

• He directly rejects the idea that wheels plus 2 arms are a compromise: “I would not call it a compromise. Our philosophy is intelligence defines the body—start from the needs of intelligence and decide how to build the body. Our entry point is manipulation, so the 2 arms are the focus. Bipedal systems instead make intelligent evolution harder. Bipedal locomotion and control plus bimanual manipulation—loco-manipulation—has not been solved even today. Many real-world scenarios do not require stepping over obstacles; wheels are enough.” He partly accepts the host’s formulation that data defines the body: “Intelligence is the objective; data is the means.”
• Around March 2024, the company settled on a wheeled base plus torso and built R1. It found that academia genuinely had demand for it, which brought Xinghaitu into what he calls the “developer market.”

30. The Developer Market Is Not the University Market; It Is the Necessary Crossing of the Chasm

• The commercial strategy in 1 sentence is: “move from the developer market to the productivity market,” following the technology-product progression from innovators to the early majority. He offers 2 examples: Macintosh moving from geeks to designers to office workers; and 拓竹 3D printers moving from a toy for a small group of geeks to an enterprise essential, then a household purchase, and eventually into printing factories. “宇树 is actually the same model—universities, second-development companies, and gradually the entertainment market. There is a pattern behind this.”
• The developer market itself is a pyramid. At the top are academic developers such as Fei-Fei Li and PhDs at top U.S. universities. Next are research-oriented developers inside companies, including Physical Intelligence and “the 灵波 TBLA system we helped Ant build.” Below them are productivity-oriented developers focused on deployment and second development. As products mature, integrators will become the new developers, with end users coming last.

31. The Learning from Hardware: AI Requires Talent Density; Hardware Requires Process Rigor

• The commonality is that both are engineering problems requiring the same methodology: “decompose and measure, break complex problems into subproblems, and organize them properly.”
• The difference is one of the episode’s few detailed manufacturing discussions. “AI emphasizes talent density and ten-x engineers. Electromechanical hardware emphasizes the rigor of the entire R&D process. If the configuration design is wrong, everything downstream is wrong. Then come structure, wiring harnesses, embedded systems, and the robot software platform; EVT validates functions. Along the way you discover harness wear, insufficient structural strength, or recurring quality problems with a supplier’s incoming materials, and eliminate them through rigorous processes. DVT checks consistency and aging, then comes production. Robotics and consumer electronics work similarly.”

32. The Annual Main Line: The Central Contradiction Changes over Time

• He agrees with the host that embodied intelligence is a hexagonal competition, then contrasts it with foundation models. Language-model competition occurs mainly in the model itself: 90% of the data is external, channels and endpoints already exist. “In embodied intelligence, there is almost no supply chain, data is a desert, it is uncertain whether the algorithms work, channels are entirely offline, and the endpoint is the robot itself.” Algorithms and models are only a small part; much of the value chain is missing, along with the ability to continuously secure government and capital resources.
• The rhythm is therefore incremental: “2024: financing plus the full-robot supply chain; 2025: financing plus the data and intelligence system”—in August, the company made the first company-level open-source release in China, with 500 hours of self-collected high-quality teleoperation data and the G0 foundation model, and reached more than 150 customers that year. “2026: financing continues, while the business focus moves to scenarios and applications.” Is he satisfied with the data and intelligence? “Impossible. But the framework is in place. Zhao Hang’s intelligence team has a good atmosphere and good results, so I can spend less time there.”

33. The Real-World Data Ledger: RMB200-250 per Hour, RMB25M for 100,000 Hours

• Why insist on real-world data? The underlying principle is domain gap: “The problem belongs to a particular domain, so the training data should ideally come from that domain. Domain transfer was once a hot topic, but people stopped working on it—it is still more effective to use data from the relevant domain to solve the problem.” In autonomous driving, 99.9% of problems are solved with road-test data. They believe the sim-to-real gap in graphics-based simulation remains large.
• He turns the cost calculation into a guessing game, asking the host to estimate the cost of 1 hour of real-world data; she answers RMB1,000. His measured figure: 3-4 hours of labor plus robot depreciation. A RMB100,000 robot with a conservative 1,000-hour life—the final life limit is the gear, which either strips or loses precision—depreciates at RMB100 per hour. The all-in cost is RMB200-250. 10,000 hours cost RMB2.5M; 100,000 hours cost RMB25M, roughly the amount of time one person interacts with the physical world from birth to age 18. He dismantles Wang He’s calculation that 10,000 data-collection robots would cost RMB1B: “Why do you need to buy 10,000 robots? That question was never answered. I focus on the hourly cost and define the unit of measurement first.” Another calculation is equally important: data acquisition costs are roughly 1:5 to 1:10 versus training costs—low-quality data means wasting money at the training stage.

34. Who Says the Data Pyramid Has to Look That Way? The Data Recipe Must Be Discovered

• On the popular view of a pyramid with real-world data at the top and simulation and first-person data at the bottom, he pushes back: “The data pyramid should be defined by the needs of intelligence. The pyramid is right, but who says it has to look exactly like this or use exactly this ratio? Nobody.”
• He lists the data types: robot-centric data from real-machine teleoperation; human-centric data from UMI, exoskeletons, collection gloves and grippers, and POV headsets; third-person internet video; and simulation, divided into graphics-based simulation and genuine world-model generation. The mix is the industry’s biggest secret. “This is called the data recipe. It is the biggest secret of large language-model companies today. Is it 10,000 hours of real-machine teleoperation, 50,000 hours of UMI, or 200,000 hours of POV? You have to try it out—AI is ultimately an experimental science.” Xinghaitu’s job is to ensure that every category of data can be acquired without bottlenecks. The recipe it is training now is, candidly, “not especially clear even to me—roughly a mix of real-machine teleoperation, UMI, and POV.”

35. Rejecting the Conspiracy Theory, and Scaling Real-World Data

• The host presents a cynical theory: the company is betting on real-world data because its business model requires selling robots to laboratories so they can collect data. He flatly denies it: “Absolutely not. Embodied intelligence is a completely technology-driven startup—technological change drives product change, which drives business-model change. It is not that the business model is fixed first and therefore the technology is fixed. That logic is wrong.”
• Real-world data becomes scalable through 2 elements. First, it must enter real-world scenarios, rather than being collected in so-called material rooms. “We started this last year and should have been the first in China to collect data in real-world scenarios.” Second, the collection equipment must be distributed through crowdsourcing. Government support, company investment, and a viable business model will support the effort this year. This is also why he watches North America closely: “I think North America will enter no-body data collection earlier than China. UMI and collection gloves and grippers are very important directions.”

36. Three Supply-Side Metrics: Speed, Precision, and Generalization

• The top-level supply metrics are “speed—imitation learning will probably not exceed human speed, and 80-90% of human speed would be good; precision—first solve the centimeter level, with millimeter-level precision next; and generalization—how many new data points are needed to solve a new problem, measured by marginal cost.” They have achieved zero-shot generalization for grasping, or grasping anything. Folding works for towels, T-shirts, and shirts, while other clothing still requires retraining.
• These metrics become a demand-side filter. Speed requirements cannot be too high; AI will still make mistakes, so the cost of failure cannot be high—one mistake cannot cause irreversible loss. A good scenario must have breakout potential, meaning “you can move quickly to 10,000 units after completing one.” It must also be global: markets found only in less-developed regions are not ideal first markets because Europe and the U.S. have greater purchasing power. And global scenarios need to be sufficiently uniform: hotels and retail look different across countries, while many commercial-service scenarios do not generalize.

37. Five Action Types and Bin Picking: Why Traditional Solutions Fail

• The 5 broad action categories are: carry, pick, pack, fold, and operate equipment. “Many jobs in the real world are combinations of these 5 actions, with 20-40 actions in each job. Nobody performs several hundred types of actions in a day.”
• His first scenario is warehouse logistics and bin picking: tens of thousands of SKUs are stored by type, and workers take an order and go wherever they need to pick. “The quantity is too large, and traditional logistics solutions cannot solve it effectively.” Another is in-plant logistics in smart manufacturing, such as SPS kitting in auto plants, where an AGV makes a loop and one person pulls parts from every 3-5 racks to load a vehicle. To the objection that robotics companies could do this 10 years ago, he gives 2 answers. Kiva and clamp robots solve a different category. When there are many SKUs and an order selects a few, traditional systems have no answer. Collaborative arms can palletize but cannot generalize. Second, this is not customization: the problem reduces to “pick anything, place to somewhere,” a highly unified problem from the AI perspective. Whether someone pulls a mask, earbuds, or chewing gum from a pocket, a system trained on real-world data through VLA should be able to grasp it.

38. A Dual-System VLM plus VLA: Not an Academic Preference, but an Edge-Compute Reality

• He first clarifies the architecture. “Brain” is too vague; return to foundation models. One is the action foundation model, VLA, which takes vision and language as input and produces action. The other is the multimodal language model, VLM, which decomposes instructions and performs logical reasoning. The dual system is the combination of the 2.
• Why not a single end-to-end model? The answer is all about deployment constraints: “Inference models with tens or even hundreds of billions of parameters cannot be put on the edge, while putting them in the cloud creates unacceptable latency—the model executing actions has to run on the edge.” Many commercial scenarios involve only 20-30 actions, so directly calling the VLA language interface is enough. The VLM is genuinely needed for more general and flexible scenarios such as the home. The question remains what can create real commercial value. Which matters more, the body or the brain? “The model is definitely more important, but to build a good model, the full robot also has to be good. That is the relationship between them.”

39. The Confidence of a First-Tier Company: Four Stages from Open Source to Demo

• “Are you a first-tier company?” “If we look at China, I think we are. Judging by actual intelligence, we definitely are.” His evidence chain: in August 2025, Xinghaitu made China’s first company-level open-source release with 500 hours of self-collected data and the G0 foundation model. Competitors followed in September, December, and January. In January this year, it released the out-of-the-box G5 Plus, also called “G5 Plus,” integrated with the R E Lite robot, which he calls a domestic first.
• He gives the industry a measuring stick: “demo in the video → demo in the office → demo in the wild → application.” Xinghaitu’s grasp-anything capability is already demo in the wild: deployed in Singapore, South Korea, and the U.S., as well as at an investor annual meeting, where it could handle whatever objects people pulled from their pockets. Only after demo in the wild comes application.

40. Big Tech’s Real Advantage Is Business Synergy—but Embodied Intelligence Has None

• He reverses the question: “We should ask what advantages big companies have over us.” A strong VLA needs real-world manipulation data rather than internet data, algorithms, compute and capital, and talent. Big Tech has strong infrastructure, compute, and talent, but its biggest shortage is data. U.S. companies have told Tan Jie of Google DeepMind that they are anxious about the lack of data. China has a systematic advantage in data collection, while Chinese startups that build their own hardware can move faster than Big Tech. The paradox of buying data is decisive: “If you do not understand models, you cannot define a good data system, and the data you collect will be garbage. A company that understands foundation models must define the data and governance systems, then hand collection and acceptance to an operations team. The supplier’s quality depends on whether it understands models, and that is precisely the contradiction.”
• The other half of the Big Tech argument is that its greatest advantage in a new business is not talent or capital but existing business synergy. ByteDance’s language models can tap Feishu, Douyin, and traffic; an automaker’s in-house assisted-driving team directly controls data and users. Embodied intelligence has none of this: demand spans thousands of industries with little synergy, while supply is entirely new—no auto parts can be used directly on robots, and road data is of limited use for manipulation. What remains is competition in talent, organization, and capital, where each side has its own advantages. As for the route: full robot plus data collection plus end-to-end intelligence is the Tesla route. The difference is that cars can be sold, while robots themselves cannot yet be sold easily. That is the challenge.

41. Xu Huazhe’s Departure: A Fact-Based, Merit-Through-Battle Culture

• He begins by giving Huazhe full credit: “Huazhe is a highly influential scientist who understands many frontier problems very well.” Asked whether there was a values conflict between research and mass production, he first corrects the wording. It was not mass-production culture but how important it is to create value for customers. A culture that does not treat the customer as the company’s most important reason for existing creates many problems. Otherwise, a startup founded by technologists can easily become a large laboratory or research institute, betraying the point of building a company. He then downplays the disagreement: “There were not that many disagreements. It was about balancing pragmatic innovation—building customer value step by step—with more advanced innovation.”
• The facts were settled internally last August. Zhao Hang took unified control of the foundation-model team, and “that was also when we achieved many breakthroughs. The out-of-the-box release in January was delivered by Zhao’s team.” Zhao also led Huazhe in exploring to C home applications. The final decision was to support Huazhe’s startup, and Xinghaitu will invest in its first round. “This is a better choice for everyone under the circumstances.” Was it difficult? “It is never easy.”

42. Is Huazhe’s Departure Like Shaoqing’s Departure from Momenta? And the Rebuttal to “Algorithms Don’t Matter”

• The host draws a sharp comparison: is Huazhe’s departure from Xinghaitu to Gao what Shaoqing’s departure from Momenta was to Cao Xudong? Gao does not answer for Xudong, but gives his own view: “For Xinghaitu, this is absolutely a positive thing in the long term. The underlying values come down to 2 questions: what you choose and do not choose when facing tradeoffs, and who receives or does not receive value when allocating benefits. On these questions, we will hold firmly to our values and long-term strategy and will not compromise for short-term interests.”
• “Huazhe mainly worked on algorithms. Does that mean algorithmic innovation is not important at this stage?” “No. Our algorithmic innovation capabilities are extremely strong, as Zhao Hang has demonstrated. But algorithmic innovation cannot exist independently of the company’s overall infrastructure. You have to look at the entire value chain.”

43. The Diffusion-Cycle Thesis: Algorithms Are Learned in 2-3 Months, Making the Moat Obvious

• The episode’s most tradable framework measures moat by the “diffusion cycle”—how long it takes competitors to fully learn what you have built. “Hardware and supply chain: 12-18 months,” roughly the time needed to develop a new product. Customer channels: at least 6 months, longer for large customers. The data system: another 6-12 months on top of the hardware. But for a first-tier company with a strong engineering team, the algorithm diffusion cycle is 2-3 months—everything is open source now, and papers are public. The host summarizes it correctly: it requires heavy investment but has a small moat. “Exactly. The investment in innovation is extremely large, but the barriers preventing imitation are very small.”
• Does that mean a startup only needs to follow Big Tech? He rejects the conclusion: startups need to innovate more intelligently and with better ROI. Their value is “pragmatic innovation—not no innovation, but pragmatism first.” Idealism is right, and he considers himself an idealist, but idealism cannot become fantasy. Its foundation is calculating ROI every day: the long-term strategic contribution plus the short-term benefit of an action equals its total return.

44. The Technology Vision: Train a Robot the Way You Train an Employee

• Pressed by the host that everything sounds too practical, he gives the line that appears on the first page of every financing deck: “Train a robot the way you train an employee—with a few demonstrations followed by a few rounds of self-practice, the robot should be able to complete tasks autonomously and reliably in that scenario.”
• The product stack supporting that experience has 3 parts: “foundation model plus post-training tools plus the full robot,” allowing customers to use a robot employee the way they use a human employee. What does Xinghaitu bring that is different? His answer is restrained: more than 150 global developer customers are already using its hardware, data, and model products. “From this year onward, we hope the products will reach real users and enter productivity scenarios, raising productivity and bringing more happiness.”

45. Robots Are Not Romantic; the Company’s Closest Animal Is a Wolf

• Returning to the opening question, his answer is structural: “It may well be that robotics itself simply isn’t romantic. The chain is extremely long and the cycle is extremely long. If you are building a large language model or AI application, you do not need to worry about supply chains, data, or offline customers. You can calmly build the model and rely on social-media virality, so romance is naturally possible. We are naturally forced to go into the dirt and build many things. There is no room for romance; we have to be pragmatic. This industry does not allow such people to exist. If one does, they may suffer a great deal. The work shapes people differently: they need idealism, but also have to calculate returns and costs every day and balance many parties.”
• He hesitates before settling on the wolf: “Maybe a wolf, though that is not quite accurate. All the companies at the front of embodied intelligence are extremely wolf-like; there is no company that is not.” The defining moment came before the G0 Plus launch, when Zhao’s intelligence team worked with the hardware team “for roughly a month continuously, without weekends.” That determination to hit the target “moved me deeply, and made me proud.”

46. Selling the Dream, a RMB10B Valuation, and the Real Organizational Problems

• How do you tell whether a robotics company is selling a dream? “You cannot identify it in a single moment. You have to compare what it said 1 or 2 years ago, what it did this year, and what it delivers 1 year later.” Does Xinghaitu sell dreams? He answers candidly: “Necessary dream-selling is unavoidable—the world runs on belief. Employees, investors, suppliers, and customers have to believe we can do it. If describing the future counts as selling a dream, then we sell dreams every day. But we must work extremely hard to turn every description of the future into reality. That is not dream-selling in the same sense.”
• The latest financing has closed. Tianqi led it—“his financing ability is much stronger than mine; I mainly sell the story now.” The process had complications: “The issues around Huazhe certainly required explanation and caused some people to understand or not understand, but the objective result was very good.” Strategic investors included Geely and BAIC, alongside major PE funds and crossover investors. Six existing shareholders followed pro rata, and 3-4 invested super pro rata, including 凯辉, 基石资本, and 香河. The valuation is RMB10B, roughly 30x the approximately RMB300M valuation in January 2024. The organizational problem is precise: it is not because the valuation rose, but because the organization itself has become larger and more complex. Unless founders become inflated by valuation—which they have not; “we are much clearer-headed today than 2 years ago”—that is the core issue. The team grew from a dozen people to more than 200, roughly 20x, with partial adjustments every 3-5 months. The 2 real challenges are whether existing people, including Gao himself, can keep up with the rapidly expanding difficulty and scope, and whether the company can bring in even stronger people quickly enough. There is also inherent tension between the process discipline of hardware and supply chain and the talent density and innovation of intelligence. “I would not say we have harmonized it perfectly. Huazhe’s departure had little to do with this; our intelligence team is very strong today.” His methodology is “honesty and integrity with people, lean operations, saving where we should and spending where we should.” He adds that the company has RMB2B or even tens of billions of RMB on its balance sheet, and must spend and manage it well.

47. Learning from Peers: 宇树’s Supply Chain, PI’s Intelligence, 智源’s Operations

• He names 3 companies. “We constantly learn from 宇树’s hardware and supply chain—deep vertical integration, designing its own gears, housings, electromagnetic simulations, and motors. PI, or Physical Intelligence, is unquestionably the industry leader in intelligence. Its talent and capital density ensure it can keep working, but I think we operate more efficiently. As for 智源, I will not comment on its controversial practices. What I see is a mature management team’s operating discipline: it handles intellectual property exceptionally well, has a strong organization, is fact-based, and adjusts quickly. I respect 初然. I spoke with 邓总 very early on, and even then he had the mindset of a mature entrepreneur.”
• Does he have a startup mentor? “Not really. I learn whatever can be learned. I have not had the luck to meet someone who could teach me everything.”

48. Rapid-Fire Questions: The Positive Feedback He Wants Has Not Arrived

• What has been the biggest positive feedback? “The positive feedback I am waiting for has not arrived yet”—he means shipment volume. “Financing does not give me much positive feedback. Every successful round feels like responsibility on my shoulders: I have to spend and manage RMB2B or tens of billions properly, give colleagues good careers, and not waste the opportunities customers give us.” There has been little negative feedback either. “I formed this state very early: I do not particularly care what other people think of me. If I believe internally that we are right and fact-based, I am fine if people do not understand, approve, or even look down on us. So far, negative feedback has never shaken me.” If customers criticize the product? “I go to the customer site immediately. I am the first person responsible. Solve one problem in a way that solves a class of problems, lead by example, and require my partners to do the same.”
• The closing miscellany retains several notes. His life bet: “Starting a company in embodied intelligence is my life bet. I will do this for my entire life. The most important task now is to ship 10,000 units in productivity applications.” He recommends Lü Simian’s book on the Three Kingdoms: “Real history has its own logic. Nobody is a fool, and there are no absolutely idealistic heroes; everyone struggles, makes tradeoffs, and moves forward step by step in reality. Fortunately, this is not a zero-sum fight to the death; every company can do well.” The podcast that left the deepest impression was the episode with Li Yifan: “It showed that companies this successful today also built themselves step by step in the early days.” The industry secret known to insiders is: “The robots shown in videos perform much better than they do in reality. People need to give them more patience, but it will happen.”