168: Conversation with 王新宇: How Meituan Longzhu Invests in Tech
Summary
王新宇的核心判断不是“这是不是好公司”，而是“一个好的公司到底有多好”，最强的早期信号则是创始人能否把热爱变成长期专注。 When he first met 王兴兴 in 2016, he saw a robot dog that did not yet know what it could do; years later, the founder’s underlying state had barely changed: “He really loves this thing, and he’s made it very good.” That led him to conclude that China’s shift from copying to global leadership is not a story about a handful of companies, but a systemic opportunity across industries.
美团龙珠自 2021 年形成的“三纵三横”框架至今未变：AI、半导体、能源三种底层技术，交叉 EV、机器人、下一代终端三类场景。 Large models later became infrastructure capable of “filling every box,” reshaping applications and hardware above while influencing compute and energy below. 王兴 is just 1 equal partner at Longzhu; investments are decided through weighted voting and scoring, with “no sheriff” and no veto rights.
龙珠在基础大模型公司里只投了月之暗面，这笔 2023 年 7 月 2 日过 IC 的投资，本质上是在人、模型和产品都尚未被市场验证时押注杨植麟的 tech vision。 At the time, 智谱 and MiniMax were already valued at roughly $1.5B, while 月之暗面 did not yet have a model; Longzhu invested about $30M. 王新宇 was not focused on any single paper or benchmark, but on 杨植麟’s complete vision for talent density, research organization, continuous breakthroughs and a Super App. “Intuition is the starting point,” but conviction came from more than 6 months of subsequent contact, often monthly or more frequently.
王新宇到 2024 年中才认为 AI 应用进入可系统投资阶段，龙珠在 2025 年投了约十家早期公司，但他判断“今天还没有哪家 AI 应用公司真正成功了”。 DAU is “a ruler from the old era” and does not even rank among the top 3 metrics; token calls, ARR and directly delivered outcomes matter more. At the team level, he looks for founders with fast execution. The peak birth year in his sample is 1997: many accumulated 7 to 10 years of experience during the mobile internet’s high-growth period, making them “young veterans.”
具身智能是龙珠从 2023 年便开始重点研究的方向，但其下注方式并非把所有公司视为同一种资产。 Longzhu has backed only 宇树 in locomotion and whole-body control, while covering multiple approaches in still-unsettled manipulation, differentiated by “brains,” data, algorithms and use cases, and investing further in key suppliers such as 地瓜 and Sharpe. 王新宇 compares the debate over real versus simulated data to the debate over whether rockets should use LOX-kerosene or LOX-methane: “What matters is not … what matters is getting into space.”
面对百亿人民币估值公司快速增加，王新宇仍坚持中国投进具身智能的钱“不是太多，而是太少了”，但未来一两年会明显去伪存真。 His reference point is not any single Figure, but Tesla’s sustained R&D, the scale of capital in US AI deals and the competition for China and the US to jointly lead human progress. The goal should not be to use 20% of a rival’s cost to deliver 50%-80% of its capability, but to invest at the same order of magnitude and build a product at 120%. The real dividing line is not near-term unit sales, but whether world models, data and generalization can produce a robot’s GPT-3 or GPT-3.5 moment.
王新宇偏乐观地估计，具身智能让大众产生“这次真的能用”感受的时刻可能在一至两年内出现；即使不是，也可能大不了三年，而不是十年。 He acknowledges that autonomous driving took 6 years from investment conviction to consumer adoption, so whether it takes 3 or 6 years is not material for a long-term fund; what matters is whether the boundaries can keep being solved. His broader view is that even if AI were “locked” today and stopped improving, it would still drive an internet- or mobile-internet-scale transformation—and it plainly will not stop.
Deep dive
1. Investing Is Not Judging Companies Good or Bad, but Estimating How Good They Can Become
At the end of the episode, 王新宇 cites Michael Moritz to define the VC problem: “Our job as investors is not to judge whether a company is good or bad, but to judge how good a good company can become.” The hard part is estimating the ceiling of a good company, not merely proving that it clears the passing grade.
The starting point for that judgment is love. Someone who genuinely loves something will spend his youth creating and solving problems without a clear commercial payoff. At a fundamental level, that is the same impulse as entrepreneurship: providing other people with a better product or service.
He does not want to screen founders by education, résumé or a template based on past winners. A competition background can create affinity, but it is not the essence; the essence is recognizing “pure love” and whether it can compound into long-term focus and core capability.
2. Starting to Code at 8 or 9 Made Technology 王新宇’s Native Language
王新宇 started learning to code at 8 or 9 and participated in competitions similar to the Chinese primary-school Olympiad in informatics. Technology began as an interest, not a career plan.
In high school, he worked on software and hardware simultaneously in a robotics lab, competed and was admitted to Fudan through an early recommendation. Lab time was not all R&D; he also played Warcraft 3 with his classmates. That shared period of growing up later helped him understand the life trajectories and communication styles of technical founders.
This led him to develop the idea of a person’s “vintage”: the era in which someone is born often steers him toward the problems most worth solving in that era. What someone works on does not have to be right from Day One, but the times, information environment and knowledge base determine what he sees first.
3. Aimless Creation in a Fudan Dorm Later Became a Sample for Reading People
Entering Fudan in 2008, 王新宇 found his interests broadened by the general-education curriculum and Shanghai’s diversity. A major that ranked among the hardest to enter in Northeast China could be close to the bottom of the admissions scale in Shanghai. The contrast taught him early that no single evaluation system is reliable.
All 4 students in his undergraduate dorm were recommended admits, each strong in a different area. Together they built websites, software and electronics projects, and shot and edited 2 seasons of video, with 10 episodes per season; the most-watched episode drew more than 1M views on Youku. The scene that best captures that period was their reversing a capacitor to see exactly how it would explode—“Don’t try this at home,” but the curiosity was real.
They also solved small problems around them: turning off the lights without getting out of bed in winter, preventing bicycle theft, adding a small fan to military-training uniforms and building a remotely controlled camera car. No one thought about starting a company. 王新宇 says that if they had been born 5 to 10 years later, a different information environment might have led them naturally into entrepreneurship.
4. Founders’ Backgrounds Can Differ; the Urge to Keep Solving Problems Matters More
Having done competitions and robotics himself taught 王新宇 what kind of training technical talent might have received, but it did not make that background a screening criterion. He reminds himself that “people are quite different.” Investing should not be about finding one standardized type of person.
What a founder does on Day One may have nothing to do with the eventual success. More than where someone came from or what problem he initially chose, 王新宇 cares about how he thinks, whether he is willing to solve a problem over the long term and whether he can change direction quickly after making a mistake.
5. His First 3 Years at GGV Were an Apprenticeship in Investment Taste
王新宇 began interning at GGV in 2015 and joined full time in July 2016. Early in his career, he was effectively an apprentice: meeting large numbers of excellent founders alongside partners and deal teams, and first building a “cleaned” database of high-quality examples.
It is not true that he never looked at computer vision. 拓博智能 applied vision to wind-turbine blade inspection and unmanned retail. His more systematic work focused on AR/VR, including Magic Leap, ODG, Meta when it was still working on glasses, and Lumus.
Insta360 was a classic example of technology migrating across cycles. It began around VR and 360-degree capture, then found PMF in sports and other use cases through its video-stitching capability. 王新宇 summarizes the technology cycle this way: “Sometimes you overestimate what can happen in 2 or 3 years, and may underestimate what can happen in 10.”
6. Writing a 100-Page Report in 6 Months Was About Shortening VC’s Long Feedback Loop
At GGV, he set himself 6-month goals. To study AR/VR systematically, he spent the day sourcing deals and writing memos, then worked from roughly 6 p.m. until midnight after dinner for half a year. He ultimately produced 2 English reports totaling more than 100 pages; no one assigned the project.
The biggest problem for young people in VC, in his view, is not the workload but the length of the validation cycle. Failure can have many causes, while success is even harder to attribute: was it his judgment, the partner, the firm’s brand or something else? The answer can remain unclear for a long time.
Any opportunity that accelerated validation of whether he could complete the entire investment loop was therefore valuable. His move to Kunlun in 2018 was intended to answer a simpler question faster: “Can I do this job for the rest of my life?”
7. During AI’s Trough, “AI Plus Transportation” Led Him to Autonomous Driving
In 2018, 周亚辉 wanted to find someone to invest in AI, while 王新宇 had to build a team and define a hard-tech focus from scratch. The first wave of enthusiasm sparked by AlphaGo had already faded, and AI was in a clear trough. He spent roughly 6 months searching for the answer to “AI plus what?” before converging on transportation.
In early 2019, he invested in Pony.ai and later extended the thesis to Israeli lidar company Innoviz. He also invested in 追觅 around the same period, though it was not an AI thesis at the time. Together, these investments became his samples for understanding autonomous driving, critical sensors and consumer electronics.
8. Moving from Algorithms to Compute, Semiconductor Investing Was a Logical Deduction
王新宇 decomposed AI into compute, algorithms and data. He already had algorithm investments, while data was difficult to underwrite as a standalone asset at the time, so after the first quarter of 2019 he naturally shifted toward compute and proposed a systematic semiconductor research effort internally.
Although his undergraduate major was electronic engineering, his research focus was closer to computer vision than semiconductors. Learning chips was harder than learning autonomous driving and required entering a new knowledge base and industry network. The launch of China’s STAR Market at the end of 2019 created portfolio opportunities, but he stressed that the listing window was not the point: “Does the company you are investing in have core value?” was the real question.
He admits that he failed to identify the investment trend around Transformer and large models in 2018-2019. Kunlun began paying attention internally around the period after GPT-2 and around GPT-3, but the startup opportunity was still unclear. The true dividing line was November 2022: “Human society was different before and after November 2022.”
9. Repeated Learning Across Sectors Validates Investing Ability Better Than One Beautiful Exit
By around 2021, 王新宇 believed he had been validated repeatedly across different areas: 追觅 represented consumer electronics, Pony.ai represented autonomous driving and EV, and several semiconductor companies had entered the capital markets. A paradigm can transfer, provided an investor keeps learning and captures a new beta or a core company.
He does not treat an IPO exit as the only evidence. A company with core value may fail to list independently because of the many contingencies in entrepreneurship, yet be acquired or restart in another form. Conversely, a great company may emerge only when timing, team and market window all align.
Fund duration is not a hard constraint on technology investing. A dollar fund can run for more than 10 years, while the duration of an RMB fund depends on its LPs. Once a fund enters its exit period, tools such as an S Fund can let legacy LPs exit while a new vehicle continues to hold the assets.
10. Longzhu Was the One Name He Could Not Immediately Turn Down
During the pandemic, the founder of a headhunting firm traveled from Beijing to Shanghai and spent more than 2 hours eating and walking with 王新宇 before finally describing the opportunity. 王新宇 had planned to decline, but instead replied: “This is the only name I couldn’t immediately say no to.”
The initial attraction was Longzhu’s 2019 investment in Li Auto. Li Auto was one of the least-funded new-energy vehicle companies at the time, yet Longzhu invested and continued supporting it through a difficult period. 王新宇 believes the investment “deserves to be written into the history of the investment industry.”
The deeper fit was cultural. Before “patient capital” became a popular phrase, Longzhu was already repeatedly talking internally about being “patient for the long term.” After 王新宇 confirmed that he could do VC for the long run, the next question became whom to do it with and what opportunities to pursue over an even longer horizon.
In conversations with 王兴, the topics ran from rockets and satellites to new-energy vehicles, semiconductors, autonomous driving, robotics and AR/VR. What surprised 王新宇 was not merely the breadth of his curiosity, but whether he pursued it deeply: many of the fields were outside 王兴’s core business, yet he understood them to a considerable depth.
11. “Three Verticals, Three Horizontals” Is a Habit of Looking for Intersections, Not a Taxonomy
On Day One in 2021, Longzhu’s technology team established its “three verticals, three horizontals” framework. The verticals are AI, semiconductors and energy; the horizontals are EV, robotics and next-generation terminals. EV specifically means electric vehicle, and any vehicle powered by electricity is in scope, including drones.
“Next-generation terminals” was deliberately left vague because no one knew whether the form factor would be glasses, earbuds, a necklace or something else. The 9-box framework had not fundamentally changed from 2021 through the time of recording; what is more likely to change is the weight of each box, not the underlying logic.
Large models are more fundamental AI and can “paint across all 3 boxes.” They reshape autonomous driving and robotics while feeding back into semiconductors and energy. 王新宇 uses Jensen Huang’s “5-layer cake” analogy: the closer a technology is to infrastructure, the more layers it can influence.
Longzhu and Meituan’s strategic-investment arm do not operate under the same mechanism, and 王兴 is only 1 equal partner among others. Investment decisions do not merely count votes; they also assign intensity to support or opposition. But “no one has super rights”: there is no veto, no extra vote and “no sheriff.”
12. The 月之暗面 Investment Began with Pursuing 杨植麟 Before the Company Existed
By the end of 2022, Longzhu was already discussing internally who had the greatest potential to advance foundation models and even wanted to “pull 杨植麟 out to start a company.” He did not meet with them for a substantial part of the first quarter of 2023, so the team wrote long pitches, introduced talent and kept creating opportunities for contact.
The first formal meeting took place around April or May 2023. The deal entered internal decision-making in June, passed IC on July 2 and the fund pulled the trigger. 王新宇 later recalled that Longzhu invested roughly $30M, its only bet among large-model startups.
At the time, 智谱 and MiniMax were each valued at roughly $1.5B, while 阶跃 had not yet been founded. 月之暗面 consisted only of its co-founders and an embryonic team; the model had not been released. It was natural for the choice of a later-starting, younger team with clear strengths and weaknesses to create internal pressure.
The greater uncertainty came from the sector itself: how large would large models become, how would a business model form and would startups still have room to compete? There was no consensus. 王新宇 believes that “all kinds of skepticism are what’s normal,” and that early-stage investing derives much of its value from precisely this uncertainty.
13. 杨植麟 Dispelled the Assumption That a Technical Genius Cannot Run a Company
王新宇 acknowledges that he entered his first meeting with 杨植麟 carrying a clear bias: would he be arrogant, nerdy, technically capable but unable to understand strategy or company management? After a roughly 2-hour meeting, those negative questions had largely disappeared.
What impressed him was not a paper or a single technical achievement, but the tech vision. 杨植麟 could explain the talent density a large-model organization would require, how exceptional researchers would collaborate, what breakthroughs would have to be delivered continuously to retain them and what those people actually wanted.
Even before a model existed, 杨植麟 had clearly articulated the ambition to build a Super App. Looking back at his earlier collaboration at 致远 and some work with Huawei, 王新宇 saw someone who had been trying for years to evangelize technology and push it into real-world use—not an entrepreneur suddenly chasing the ChatGPT trend.
14. Intuition Starts the Process; More Than 6 Months of Repeated Contact Creates Conviction
王新宇 does not reject intuition, but defines it as “muscle memory” built through long training. When he first met “卷卷,” the founder spent 2 hours writing on a whiteboard. 王新宇 came out and told a colleague: “I didn’t really understand much, but I think this person is probably investable.”
Intuition must be followed by more data points. Even when he liked a founder from the first meeting, he often waited more than 6 months before investing, meeting monthly or even more frequently in the interim. If he did not understand the first explanation, he met the founder again until he could distinguish communication style from actual ability.
His principle is case by case, one matter at a time. He cannot be held hostage by past successes or scared off by past failures. In a given year, usually only 1 or 2 deals make him willing to pursue them relentlessly. That is not a quota; he acts only when he encounters that inner voice.
15. 月之暗面的 Ups and Downs Put Model Capability Back at the Center
王新宇 sees the post-investment role as “help without getting in the way.” An investor sits in the passenger seat and can offer experience on matters such as fundraising cadence, but should not direct a founder who understands the technology and product better. Giving advice and pushing the company are 2 different things; the final decision belongs to the company.
He strongly supported 月之暗面’s first large investment from Alibaba and even voluntarily stepped off the board. The capital was a “necessary but not sufficient condition” for the company’s development. He later saw a historical echo between Pony.ai and 月之暗面: the former secured $500M from Toyota about 1 year after investment, while the latter secured roughly $800M from Alibaba in less than 1 year.
In the third quarter of 2024, when the company was spending more aggressively to acquire users, 王新宇 suggested putting more attention back on the model. There was still “a very long way to go” from GPT-3.5 to higher capability. A growth strategy is not necessarily always wrong, but in his ranking of priorities, model progress remained foundational.
The shock from DeepSeek R1 validated that view: “People teaching people doesn’t work; events teach people in one shot.” He believes 月之暗面 subsequently returned to a stronger technical trajectory, with K2, K2.5 and other models improving its technical and global influence. He defines general intelligence against the average education and intelligence level of the global or Chinese population and, on that basis, argues that “AGI has already been achieved”; the next step is continued progress toward ASI.
16. AI Applications Became Investable, but No Final Winner Has Emerged
After investing in 月之暗面 in 2023, 王新宇 briefly believed that large models were “the most important, even the only important thing” in AI. It was not until mid-2024 that he had the team systematically study AI applications, initially focusing on technology- or distribution-driven 3D and video, as well as social, interactive content and games.
In 2025, Longzhu invested in roughly 10 AI application companies, many in their first 2 rounds and some directly in the first round. Most remain underwater because product forms, new demand and PMF are still being explored—not because they have mature products but are deliberately hiding them.
The experiments he cites include AI sales tools that directly deliver sales outcomes, coding, browsers, 3D and video, as well as 影谋 and OneTwoX. This is not a rejection of productivity tools; the team’s time is limited, so it can only prioritize the questions it most wants to understand.
His view on valuations and operating metrics is equally restrained: “Today, no AI application company has truly succeeded.” Some companies have decent ARR or DAU, but DAU is “a ruler from the old era” and does not rank among the top 3 metrics. Token calls, revenue and actual delivery outcomes are more worth tracking.
17. The Median AI-Application Founder Is a “Young Veteran” Around 30
王新宇 reviewed more than 100 AI application companies last year and only afterward noticed that founders’ birth years formed a roughly normal distribution peaking in 1997, with 1996 and 1998 the most concentrated years; there are already some post-2000 founders. This was an observation, not a decision to “invest in young people” followed by age-based sourcing.
陈曦 is his example of the type. Born in 1996, he joined Douyin in 2018, caught the product and company during their high-growth phase and now has roughly 7 years of experience. A young age does not mean shallow industry experience.
The deeper vintage comes from their formative environment. They were around 12 in 2008, and their values took shape during a decade of economic expansion and improving material conditions. A group-level confidence made some of them more willing to think big, act boldly and pursue what they strongly believed in.
王新宇 calls the target profile a “young veteran”: either someone under 30 biologically who has already spent 10 years in a field, or someone genuinely in his 40s or 50s with an unusually open, youthful mindset. 杨植麟 and 王兴兴 fit the first category well: “Without love, it is very hard to stay focused.”
18. 宇树 Became Investable When Robotics Labs Around the World Sent the Same Signal
Longzhu had identified embodied intelligence as a priority as early as 2023. The team reviewed more than 100 companies, while 王新宇 personally met roughly 20. It also had early contact with companies including 自变量 and 星海图, but at the time he candidly said he “still didn’t understand it.”
The turning point came during a NeurIPS trip at the end of 2023. He visited robotics labs at Stanford, Berkeley, Harvard and MIT and found that almost all of them had 宇树 robot dogs. More importantly, 宇树 had already built its first humanoid robot, and the world’s leading labs were eager to buy it.
王新宇 had previously questioned 宇树’s AI capabilities. He returned with a clear thesis: this resembled Jensen Huang putting gaming GPUs into labs for AI training. If the world’s best PhDs were using 宇树 hardware to study frontier problems, “wouldn’t the AI-capability question eventually be solved?” Longzhu invested in the first half of 2024.
19. Locomotion Has Converged; Manipulation Still Warrants Multiple Bets
Using a rough framework from early 2024, humanoid robotics can be divided into locomotion and manipulation. Locomotion has further evolved into whole-body control. Longzhu has invested only in 宇树 on this track and has not backed another hardware-platform company.
In 王新宇’s view, manipulation has still not converged. Startups hold different beliefs about the “brain,” data, algorithms and use cases, while the debate over real data, simulated data or a hybrid approach remains intense. Longzhu therefore covers different teams and methods.
He uses commercial spaceflight to explain the route debate. In 2018-2019, the industry argued over LOX-kerosene versus LOX-methane, but “what matters is not LOX-kerosene or LOX-methane; what matters is getting into space.” Different missions may naturally use different fuels; a route label cannot substitute for results.
For a founder, committing to one route and executing it seriously is necessary. But once he discovers it is wrong, he must switch quickly. “If you start a company, don’t be afraid of being proven wrong.” The ability to admit mistakes and change fast may itself be a core competitive advantage.
20. The Home May Arrive Before the “Factory, Then Store, Then Home” Sequence
By 2024, progress from overseas companies such as Google and Figure, together with advances by other teams, had pushed Longzhu’s conviction in embodied intelligence above the investment threshold.
Several home-focused robot demos changed its view of the path. 王新宇 was no longer certain that robots had to enter factories first, then commercial settings and only later homes. The home could be the starting point—and “faster than people think.”
妙动 illustrates how a company’s direction can change. Longzhu invested in 2 rounds, and 杨朔 had not yet joined at the first. 王新宇 warns that public labels often differ from what a team is actually exploring in the lab; a company should not be defined indefinitely by the category it used in its first fundraising round.
21. Critical Component Suppliers May Become the Chain Leaders
Beyond platforms and algorithms, Longzhu invested in key suppliers such as 地瓜 and Sharpe. The logic is that once an industry matures, the most important key supplier may capture more value, and the “chain leader” is not inherently the finished-product company.
王新宇 uses new-energy vehicles as a reference point. Conventional thinking says automakers lead the chain, but CATL’s profit can exceed that of all automakers. Value and bargaining power migrate with the bottleneck, and robotics could develop a similar structure.
妙动, 地瓜 and Sharpe are therefore not the same type of bet as 宇树, 星海图 and 自变量. Longzhu is trying to cover nodes that can control critical capabilities over the long term, rather than putting all its money into the single label of “software-hardware integrated humanoid robots.”
22. High Valuations Do Not Mean Excess Investment; China Needs Capital at the Same Scale
曼祺 says that, to his knowledge, China may already have roughly 20 embodied-intelligence companies valued above RMB10B. 王新宇 did not update the latest table. The team had once tracked valuation bands of $200M, $300M-$500M and $1B, but stopped as valuations rose too quickly and the bands ceased to be the most important question.
Even so, he maintains: “The money China has invested in this field is not too much; it is too little.” 曼祺 counters that Chinese companies’ cumulative financing may already exceed Figure’s. 王新宇 brings Tesla’s R&D spending, Scale AI’s roughly $14B valuation premium and the $1.4B financing supported by it into the comparison.
His point is not to defend any single valuation, but to reset the competitive objective. China should not build embodied intelligence at 20% of a rival’s cost while delivering 50%, 70% or 80% of its capability. If China and the US may both lead human progress, they should invest capital at the same order of magnitude, deliver a product at 120% of the benchmark and potentially lead through original innovation.
王兴兴, 梁文锋 and 张雪 in the motorcycle sector, all emerging in 2025, showed him a common pattern: not copying for a near-term commercial opportunity, but turning a long-term interest into a world-class capability. For investors, the fact that “love and focus will flower across every industry” is a systemic opportunity.
23. The Next Embodied-Intelligence Shakeout Will Be About Generalization, Not Unit Sales
王新宇 does not deny the bubble or the setbacks: “Every undertaking and every sector goes through setbacks.” The next 1 to 2 years will separate the real from the false, testing whether teams are genuinely pursuing single-point technical leadership and whether they can move toward commercialization. At the current stage, single-point technical leadership matters more.
He believes world models and data are close to industry consensus. The question is whether these approaches can lift robots’ generalizable, general-purpose capabilities by another level. It can be called a GPT-3 moment or a GPT-3.5 moment; what matters is that the public suddenly feels, “This thing is usable.”
Unit sales are not irrelevant, but their value depends on whether they generate real-world feedback and data that improve generalization. If shipments are merely a revenue number with no causal link to a capability leap, they are not the most important leading indicator.
The public-facing standard is concrete: robot demos should no longer need to be shown at 2x, 4x or 8x speed; they should remain smooth at 1x. The robot should do things in a way comparable to a person, rather than merely appearing to complete the task after editing.
24. Autonomous Driving’s 6-Year Delay Has Not Weakened His 3-Year Window for Embodied AI
王新宇 takes the optimistic view that embodied AI’s GPT-3 or GPT-3.5 moment could arrive within 1 to 2 years. Even if not, it may take no more than 3 years; “it won’t be off by as much as 10 years.” He also leaves room to be proven wrong, stressing that mistakes should be acknowledged and learning should be fast.
He sees FSD V12 as the public-perception moment when autonomous driving approached GPT-3.5. His own GPT-3 conviction came much earlier, when he invested in Pony.ai in the first half of 2019. At the time, he had already concluded: “My next generation definitely won’t need to learn how to drive.”
To understand the technology’s boundaries, he logged hundreds of kilometers testing Pony.ai vehicles in Fremont, Beijing, Shanghai and Guangzhou Nansha, across daytime, nighttime and rainy conditions. Seeing the boundary does not mean rejecting the technology; understanding why the boundary can be solved is what creates conviction.
In early 2019, he invested $50M in Pony.ai at a $1.6B valuation. Auto and Cruise had actual acquisition prices of roughly $1B each, while Waymo’s $17.5B was largely a paper valuation. About 1 year later, Toyota invested $500M at a valuation of roughly $2.5B, providing external validation of the thesis.
25. From Robot Dog to Humanoid, 王兴兴’s Underlying Drive Has Not Changed
In the first week after 王新宇 joined GGV full time in July 2016, he met 王兴兴, who had only recently started his company. When 王兴兴 received his first personal investment, he sent the SPA to 王新宇 to ask whether he could sign it. The 2 were “workplace childhood friends,” entering professional life at roughly the same time.
王兴兴 had already built a robot dog; he did not wait until MIT open-sourced one to begin. The 2 even seriously discussed what it could do: could it lead children around a park like a carousel? The answer was unclear, but 王新宇 saw that “he really loves this thing, and he’s made it very good.”
王兴兴 described humanoid robot H1 as “the dog standing up.” It both is and is not: the accumulated work on motors, control and whole-body control is indeed shared. 王新宇 had never understood what a robot dog could do until the humanoid product and demand from global labs emerged, resolving his questions about both applications and AI.
He believes 王兴兴 will “use small robots to make smaller robots and large robots to make larger robots,” just as he believes Musk genuinely wants to go to Mars. Pursuing the cutting edge inevitably brings failure and skepticism: “What rocket doesn’t fall?”
26. The Next 王兴兴 Will Emerge in Every Field That Is Not Yet World-Leading
When 曼祺 asked where the next 王兴兴 would appear, 王新宇’s answer was not a specific sector but “every sector”: AI applications, AI-native hardware and any industry in which China is not yet number 1 globally but could become number 1.
These people are both easy to recognize and difficult to invest in. They are often not successful by conventional standards and may fail to fit standard VC criteria for a long time. Even when an institution identifies the person, it is hard to know when success will come or whether it will come in this particular venture.
It is therefore normal, in probability terms, not to have invested in 王兴兴; any successful investment is an extremely low-probability event. VC’s social value is not requiring every firm to hit every winner, but maintaining enough diversity among founders and investors that genuinely different projects have at least someone willing to back them and help them continue.
China’s consumer-electronics evolution provides a reference point. Once a country can make smartphones, it may go on to build products across 3C, robot vacuums and other categories. Japan moved from copying in the 1960s to transformation in the 1970s and then to homegrown core technology and brands in the 1980s; China could make a similar shift over the next 5 to 10 years.
27. “Assembling a Team” Is Not the Original Sin; Opportunism Without an End State Is
王新宇 does not condemn “assembled” startups in hot sectors based on their starting point. Temporary combinations and institutional matchmaking are not decisive. What matters is what the team is pursuing when it comes together, where it intends to go and whether it can withstand the inevitable challenges and pain of entrepreneurship.
If the team is merely swept up by fundraising sentiment and short-term noise, with no vision, the probability of failure is high. Companies assembled around the goal of gaming new-energy vehicle subsidies around 2016 are a classic example of the wrong motivation.
Li Auto is the counterexample. Some of 李想’s eventual partners came from the later team and some from Autohome, but the name “Li Auto”—“car and home”—already stated the Day One vision. The first product’s lack of success did not change the fundamental problem the company wanted to solve.
Raising 1 or several rounds of capital is not even the beginning. What the team does with the money and whether it develops a genuine pursuit determine whether it can enter the final competition. Teams lacking that layer are often not eliminated by rivals; they trip over themselves first.
28. The Embodied-Intelligence Endgame Is Huge, but It Will Not Leave One Company or One Robot
王新宇 describes the end market from the demand side: “No one doesn’t need a driver, no one doesn’t need a secretary, no one doesn’t need a nanny, no one doesn’t need an assistant.” Most people simply cannot afford one today. If 8B people each needed 3 embodied-intelligence robots, the potential demand would reach 24B units.
Robots will also replace Dull, Dirty and Dangerous work, using technology to improve people’s lives. To B and To C adoption will spread simultaneously, eventually permeating daily life like smartphones, but that does not mean every product must take the same humanoid form.
The market will not be an extreme winner-take-all outcome. After a century of consolidation, global internal-combustion auto still has more than 10 major automakers; outside China, global smartphones still include Samsung, Google and Nothing. Embodied intelligence is unlikely to have CR1 or CR3 absorb the entire market, nor will it fragment into a collection of subscale “ant markets.”
Competition will persist, but the real war is not simply for share against peers. It is the ongoing delivery of better, cheaper products with stronger core advantages. New-energy vehicles show that intense competition can eventually put the technology dividend in ordinary consumers’ hands: even a RMB100K-level car can offer autonomous-driving functionality.
29. AI Will Rewrite Hardware, Chips and Investment Organizations—and Accelerate Its Spread to 8B People
Beyond “three verticals, three horizontals,” a key extension is AI-native hardware that would not exist without AI, including projects such as Luki and 极壳 mentioned on the show. The next generation of terminals may not uniformly replace the smartphone. TWS, Apple Watch and Oura Ring have shown that a device can create value without replacing the primary device; future experiences are more likely to combine LUI, cloud intelligence and endpoint capabilities.
Semiconductor and energy investing is also increasingly organized around AI. Longzhu recently invested in controlled nuclear fusion and in 2 companies pursuing different compute architectures, one of them closer to TPU. 王新宇 expects GPU to remain more general while large-model tasks are unsettled; once tasks stabilize, specialized architectures such as ASIC, DSA and LPU should win on compute, power consumption and cost. Chip innovation can still come from architecture, manufacturing and materials.
Longzhu’s technology team, which numbers roughly 10 people, has integrated AI into reports, deal research, IC risk assessment and transaction communications. 王新宇 also fed the year’s project notes, scores and owner information into an internal AI system to analyze time allocation. Its conclusions matched a 4.5-hour manual calendar review by more than 90%, and he has adopted its recommendations on areas where the team was under- or over-investing time.
In February 2015, reading an article about AGI and ASI at Burger King, he wrote: “If this thing is achieved within 10 years, I’ll remember how I felt reading this article at Burger King.” Today, he cites rough statistics suggesting that no more than 0.3% of the 8B people on Earth have actually paid for advanced AI. Even if AI were “locked” at its current level, it would still drive an internet- or mobile-internet-scale transformation. It will not stop, so the next 3 years, or 3.5 years, will move faster than the period since November 2022. The hardest task for investors remains not seeing change, but judging “how good it really is.”