168: 对话王新宇：美团龙珠怎么投科技？
- 01The "Passion-First" Investing Philosophy as a Systematic Edge
- 02The "Young Veteran" (年轻的大哥) Framework for Identifying Exceptional Founders
- 03China's Transition from Copycat to Global Leader in Hardware
Podcast: 晚点聊 LateTalk | Guest: Wang Xinyu, Partner & Head of Tech Investing at Meituan Longzhu
1. Key Themes
The "Passion-First" Investing Philosophy as a Systematic Edge
Wang Xinyu's core thesis is that founders who operate from genuine passion — not financial opportunity — represent a systematic, repeatable pattern that will increasingly define breakthrough companies in China. He traces this back to his own childhood building robots, writing software for fun, and competing in olympiads, and uses it as his primary lens for founder evaluation.
"The power of passion, the power of focus, will blossom across thousands of industries. This is a systematic, enormous opportunity." [00:00:28]
"If you truly love something from the bottom of your heart... and you have creative energy without any utilitarian motive to solve problems — that's essentially the same as entrepreneurship." [00:10:42]
The "Young Veteran" (年轻的大哥) Framework for Identifying Exceptional Founders
Wang Xinyu articulates a precise archetype he calls the "young veteran" — founders who are biologically young (born ~1995-1997) but have already accumulated 10+ years of deep, focused experience in their domain. This cohort grew up during China's golden economic decade, giving them unusual confidence and ambition combined with native fluency in mobile/AI technology.
"What is a 'young veteran'? One interpretation: biologically very young — born in '96 or '97, barely 30 years old. The 'veteran' part means they've already been doing something for 10 years in a particular domain." [01:00:02]
"The most representative distribution peak for AI application founders is those born in 1997, roughly within a three-sigma distribution of 1996-1998. We've invested in people born after 2000 too." [00:55:36]
"These post-95 founders are bolder in imagination, bolder in execution, iterate faster, and are more self-confident... they are AI native themselves." [00:56:37]
China's Transition from Copycat to Global Leader in Hardware
Wang Xinyu articulates a structural thesis: China is in the early stages of the same transformation Japan completed in consumer electronics from the 1960s-80s — moving from imitation to origination to global dominance. He sees embodied intelligence as the clearest current battleground where this transition is happening.
"When Chinese people could make phones, they could make all 3C electronics. iRobot — where is it today? Basically the whole world has become the stage for Chinese companies competing." [01:16:02]
"What we are doing today in embodied intelligence is not using 20% of someone else's cost to make something at 50-80% capability. We are in competition that leads human progress... We should use comparable money — or slightly less, with higher efficiency — to make something 120% as good." [01:14:36]
"2025 is very important for China and for my investing. 2025 gave us DeepSeek, Liang Wenfeng, and recently Zhang Xue and his motorcycles. I believe this paradigm will increasingly multiply in China." [01:15:33]
2. Contrarian Perspectives
China Is Under-Investing in Embodied Intelligence, Not Over-Investing
When most observers worry about a robotics bubble with 20+ companies valued over $1B in China, Wang Xinyu takes the opposite view — arguing China's total investment in the space is still insufficient given the scale of global competition.
"I still believe China's investment in this field is not too much, but too little. Even today, that conclusion holds." [01:13:42]
"Figure [the US humanoid company] has raised more than any single Chinese company. And if you count how much Tesla has invested in R&D in this space? Scale AI alone raised $14 billion." [01:14:12]
AGI Has Already Been Achieved
Wang Xinyu makes a bold, specific claim most AI researchers would disagree with: AGI is already here, defined by the threshold of exceeding average human knowledge and reasoning across the global population.
"If you measure by the average education level and IQ of 8 billion people globally — or even China's 1.4 billion — today's AI has already surpassed that average. That's 'general.' So I think AGI has already been reached." [00:51:16]
"I believe the world before and after November 2022 is fundamentally different. Even today, continuing toward AGI, there's still a long road ahead — but it follows a power law." [00:50:47]
The Manipulation/Dexterity Problem in Robotics Hasn't Converged, and That's Fine for Investors
While most robotics investors claim strong conviction on the winning approach to robot dexterity (real vs. synthetic data, simulation vs. real-world), Wang Xinyu argues the debate itself is a distraction — the equivalent of arguing about rocket fuel type instead of whether the rocket reaches space.
"People were debating liquid oxygen kerosene vs. liquid oxygen methane... I said at a panel: what matters is not the fuel. What matters is getting to space. Because Musk uses both — he uses one for landing rockets, another for Mars. First principles." [01:07:00]
"The important thing is not synthetic data or real data. Just like it's not liquid oxygen kerosene vs. methane. What's important is reaching space — that's the key." [01:07:18]
DAU Is a Obsolete Metric for Evaluating AI Applications
Wang Xinyu argues the entire analytical framework inherited from mobile internet — centered on DAU — is wrong for AI applications, and investors who use it are measuring with yesterday's ruler.
"DAU is already an old-era ruler. It should not be the most important metric for AI products — it doesn't even rank in the top three. Token usage is a very important metric." [00:53:11]
The "Team Assembly" (拼团) Origin Story of a Startup Doesn't Predict Failure
Contrary to common VC wisdom that opportunistically assembled teams ("the group got together because the sector is hot") are red flags, Wang Xinyu argues the origin is irrelevant — what matters is what the team pursues after formation.
"Whether it's a 'team assembly' doesn't matter much in my view... Where you come from is not the most critical thing. Where you're going is most critical. And how you get there." [01:41:06]
3. Companies Identified
Unitree Robotics (宇树科技) Quadruped and humanoid robot manufacturer; recently IPO'd Mentioned as a foundational portfolio company. Wang Xinyu met founder Wang Xinxin in his first week of work in July 2016. The combined Meituan ecosystem (Meituan direct + Longzhu) is now Unitree's largest external shareholder. Wang Xinyu's investment thesis shifted when he visited top US university robotics labs (Stanford, Berkeley, Harvard, MIT) at NeurIPS in late 2023 and found Unitree's quadruped dogs in every lab, with pre-orders for their first humanoid robot.
"All these robotics labs had Unitree's dogs. More importantly, at that point Unitree's first humanoid robot had already been built and they were beginning pre-sales. All the labs desperately wanted to buy." [01:03:22]
"My investment thesis was: I saw Huang Renxun [Jensen Huang] giving gaming GPUs to labs for AI training. If the world's best robotics PhD students are all using Unitree's humanoid for development and frontier research — will AI capability remain a problem?" [01:04:03]
Moonshot AI / Kimi (月之暗面) Chinese large language model startup; creator of Kimi chatbot Wang Xinyu led Longzhu's investment in July 2023 — the fund's only LLM investment. He had been trying to contact founder Yang Zhilin (Kimi) since late 2022, wrote "love letters," made introductions, before finally meeting in April 2023. The investment preceded Alibaba's $800M+ investment by under a year.
"When we pulled the trigger, all co-founders were there, the team had an initial shape, but there was no model yet... There was some pressure." [00:40:03]
"He had a very clear Tech Vision. On Day 1, he spoke about building a Super App. And this was when they didn't even have a model yet, let alone an app. But these things were very clear." [00:44:01]
Pony.ai (小马智行) Autonomous driving company; now publicly listed Wang Xinyu's first self-led investment at Kunlun Wanwei in early 2019 — at a $1.6B valuation, which was the highest in the world for an autonomous driving startup at the time. He rode hundreds of kilometers in Pony vehicles across multiple cities before investing.
"In 2019, giving it a $1.6B valuation meant it was the highest valuation in the world. Waymo's $175B was written on paper by investment banks, but nobody had invested at that value. Two US companies — Otto and Cruise — had been acquired at $1B each." [01:26:21]
Momenta Autonomous driving / AI for vehicles; described as a tier-1 supplier approach One of Wang Xinyu's early investments at GGV, alongside Pony.ai. Mentioned as representing the "Tesla route" of end-to-end autonomous driving vs. Waymo's HD-map route.
"Momenta and Pony are basically the two routes everyone debated most — Tesla route vs. Waymo route." [01:07:49]
DeepSeek (深度求索) Chinese AI research lab / model provider; creator of the R1 model Not a portfolio company but mentioned as a paradigm-shifting event. Wang Xinyu argues DeepSeek's R1 release on January 20, 2025 validated his core thesis that model capability is paramount and refocused the Chinese LLM industry.
"From my perspective, DeepSeek verified what I said: model capability is still the most critical thing. And there's still a lot to do in this field." [00:49:21]
Fauo (法奥) Collaborative robotic arm manufacturer Invested by Longzhu in 2021. Cited as an example of a later entrant to a "traditional" market that won by dramatically reducing price — first to market with a 20,000 RMB collaborative robot arm. Grew from 600 units sold the year of investment to over 10,000 units annually, becoming China's #1 in lightweight collaborative robots by shipment.
"They were a latecomer in this field too. They were the first to launch a 20,000 RMB collaborative robot arm that worked very well. The year we invested, they sold 600 units. Last year they passed 10,000 units — China's #1 in this category." [01:52:08]
Insta360 Action camera company; started as a VR camera maker Mentioned as an example of a company that pivoted from 360 VR video to find strong PMF in sports/action cameras. Wang Xinyu saw their products at CES 2017 while researching AR/VR.
"They started doing VR — multiple lenses, a spherical shape, 360 degrees. Their core capability became video stitching. Walking in this direction, they eventually found very strong PMF in sports and many other areas." [00:14:31]
4. People Identified
Wang Xinxin (王兴兴) — Founder, Unitree Robotics Met by Wang Xinyu in his first week on the job in July 2016. Described as fundamentally unchanged from then to now — driven purely by passion for robotics since childhood, building quadruped robots before MIT's open-source Cheetah project became famous. Embodies the "young veteran" archetype: barely 30 years old, but with nearly 30 years of practical hardware experience.
"His foundational core has not changed. That passion, that conviction — already evident back then. And it hasn't changed to this day." [01:35:01]
"He would discuss motor parameters with eyes lit up. He's truly a 'young veteran' — when he casually says 'I've been doing mechanical R&D for 20+ years,' you think: wait, aren't you only 30?" [01:00:30]
Yang Zhilin / Kimi (杨植麟) — Co-founder & CEO, Moonshot AI Wang Xinyu's most recent high-conviction bet. He spent months trying to reach Yang before their first meeting in April 2023. Key qualities: mature beyond his age, possesses "Tech Vision" (ability to see the full organizational and talent implications of building frontier AI), and had already been systematically working toward this goal through prior collaborations with Zhiyuan Institute and Huawei.
"He had a very clear Tech Vision. He not only had done the work but knew how this should be done — what kind of organization, what talent density, how those talents collaborate, what success they must pursue to retain these geniuses." [00:43:31]
"Even at that time point, he was very mature. And if you trace his past work — he had always been trying to do this thing. The collaborations at Zhiyuan Institute, early work with Huawei." [00:43:01]
Luo Tianchen (罗天成) — Co-founder, Pony.ai ("James") Described as a "young veteran" of autonomous driving whose generational timing (slightly older than Yang Zhilin) led him to tackle the hardest problems in computer vision and L4 autonomy. Wang Xinyu flew to Silicon Valley to build the relationship after Pony initially wasn't raising.
"Luo Tianchen's generational vintage led him to solve the hardest AI problem of his era — L4 autonomous driving problems within computer vision. Yang Zhilin's vintage led him to solve large language models in NLP." [01:20:58]
Liang Wenfeng (梁文峰) — Co-founder, DeepSeek Cited as the emblematic example of China's new paradigm: a deeply passionate, focused founder pursuing technical excellence for its own sake rather than for commercial imitation. Wang Xinyu uses him alongside Wang Xinxin and Zhang Xue (motorcycle designer) as proof the paradigm is accelerating.
"2025 gave us DeepSeek, Liang Wenfeng... I believe this paradigm will increasingly multiply in China — passion-driven, focus-driven work will blossom across thousands of industries." [01:15:33]
Zhou Yahui (周亚辉) — Founder, Kunlun Wanwei Wang Xinyu's second employer, described as forward-thinking for wanting to invest in AI as early as 2018. Known for quick, decisive judgment — after a 30-minute meeting with Pony.ai's founders, he immediately supported Wang Xinyu's conviction to pursue the investment.
"Old Zhou asked me: do you want to invest in this company? I said yes, but they're not raising. He said: then go to America and find them. That was it." [01:27:47]
5. Operating Insights
Use AI as an Independent IC (Investment Committee) Member
Wang Xinyu's team is experimenting with feeding all deal notes, meeting transcripts, and related materials for a company into an AI model and asking it to render an investment judgment as if it were an IC member. The results have been surprising in quality.
"We feed all the information we have about a company — meeting notes, related information — into AI and let it act as an investor, as an IC member. When it makes judgments, it often gives us surprises: risks it identifies, perspectives it raises, good-or-bad assessments that are genuinely useful." [01:59:23]
"Even in negotiations, when we encounter challenges and ask AI for advice, there are surprises. Not just negotiation tactics — also how to navigate difficult situations." [01:59:52]
Use AI to Audit Your Own Time Allocation Against Investment Strategy
Wang Xinyu manually audited his entire calendar year (4.5 hours on a high-speed train), categorizing every hour spent. He then ran the same exercise through AI on his team's CRM/project notes database. The two analyses matched with 90%+ accuracy — and the AI provided actionable guidance on which directions were over-weighted and which were neglected, which he then acted on.
"AI analyzed where I spent my time last year. I also did this manually — on a 4.5-hour train from Shanghai to Beijing, I went through every calendar entry. The match rate between AI's analysis and mine was over 90%. And the recommendations, especially the negative ones — where I spent too much or too little time — I've followed all of them this year." [02:00:50]
The "Validate at the Boundary" Method for Conviction on Deep Tech
Before committing $5M to Pony.ai at the world's highest valuation in 2019, Wang Xinyu rode hundreds of kilometers in Pony vehicles across 5+ cities, in varying conditions (rain, night, day), specifically to find the edge cases — and then assessed whether the technology trend pointed toward resolving those edges.
"I rode probably hundreds of kilometers in Pony's cars — in the US in Fremont, and in China in Beijing, Shanghai, Guangzhou, Nansha, day, night, rain, all different scenarios. You do a lot, and you can find the boundaries. When you see the boundaries and understand the technology, you know these edges can likely be solved." [01:25:52]
6. Overlooked Insights
The "Infrastructure Supplier" Endgame for Unitree May Be More Valuable Than the Humanoid Robot Maker Narrative
Wang Xinyu briefly but pointedly compares Unitree's potential trajectory to Jensen Huang giving away gaming GPUs to AI labs — suggesting Unitree could become the foundational hardware layer that the entire global robotics research and development ecosystem builds upon. This would make Unitree more analogous to NVIDIA than to Tesla. If correct, the valuation ceiling is dramatically higher than a "robot manufacturer" frame implies, and the competitive moat comes from ecosystem lock-in rather than unit economics.
"My investment thesis was like seeing Huang Renxun give gaming GPU cabinets to OpenAI. If the world's best robotics PhD students are all developing on Unitree's humanoid robot — will AI capability remain a problem?" [01:04:03]
"If Unitree can keep providing — not just China, but the entire world — hardware that is economical and excellent, then perhaps not everyone needs to build their own hardware." [01:48:18]
This "Jensen moment" analogy was mentioned only in passing, but it reframes Unitree from a robotics company into a potential platform company — an entirely different investment thesis with much larger return potential.
The 0.3% Penetration Stat Reveals AI Applications Are Still Pre-Cambrian
Wang Xinyu drops a striking data point almost as an aside: of Earth's 8 billion people, fewer than 0.3% have ever paid for AI services (defined as ~$20/month for premium access). This means the entire current AI ecosystem — all the hype, all the valuations, all the declared winners in AI applications — has been built on less than 1/300th of the eventual addressable user base. Almost no one has noticed that we are still in the pre-commercial phase of AI adoption, which means nearly every "winner" declared today is a placeholder.
"Of Earth's 8 billion people, those who have truly paid for AI — even $20 a month for the most advanced AI — account for no more than 0.3%." [02:03:46]
"Even if AI were frozen at today's level — no more progress — it would still transform human society at the scale of the internet and mobile internet combined. But obviously it won't stop here." [02:04:43]