Yu Kai Recounts 30 Years: A Jianghu of People Coming and Going

2025/07/07 · Source link · AI Summary archive

Summary

Yu Kai’s core methodology runs through the entire episode: “Consensus is either wrong or worthless.” When most people move right, always look left. He entered deep learning when it was obscure, returned to Baidu when no one was coming back to China, and built automotive-grade chips while China’s “AI Four Tigers” were making software. Horizon Robotics was built on two principles—compete where there is no competition, and endure the lonely stretch. Today, more than 90% of Chinese automakers are customers, and the company had 102 institutional investors before going public.
His early call on Nvidia is the episode’s hardest investment evidence. In 2012, Baidu was Nvidia’s largest AI-application customer globally; “even Jensen Huang had not realized in 2012 that Nvidia was sitting on a gold mine” (the book says Huang only realized it in 2014). When he founded Horizon in July 2015, he made three investments: Nvidia, Tesla, and his all-in commitment to Horizon. NVDA’s market cap was $10.7B then; “now it’s $3T.” He has never sold.
His first-hand account of Hinton’s secret auction is a rare reconstruction. Baidu opened at $12M to scare off rivals, with an authorization ceiling of $24M; the final price reached $44M. The fourth bidder, which he only learned 10 years later, was DeepMind—a one-year-old company that entered with stock and “already had the bearing of a king.” He wanted to hire both of Hinton’s students, but gave later “father of ChatGPT” Ilia the lowest score. Four companies and 3 people around that table formed the entire outline of today’s AI landscape.
His countercyclical capital strategy reads like a textbook. During the 2020-21 capital frenzy, Horizon raised $1.6B without increasing its valuation by a single cent, betting that “the bubble would definitely burst.” During the following 3-year capital winter, most companies ran out of cash while Horizon held “RMB10B-plus” and focused on products. The framework is “the life gate and the death gate”: spend 70-80% of your energy identifying the death gate, and “never dance on the edge of a cliff.” Strategy is fundamentally about not betting—just as Buffett says the essence of investing is not losing money.
The most tradable contrarian call: automakers will not build autonomous driving in-house. Autonomous driving is a standardized feature, like a phone’s baseband—“even Apple doesn’t develop its own baseband.” The experience is the same for everyone: getting from A to B safely, comfortably, and efficiently. “You can’t drive with Guo Degang’s style or Lin Zhiling’s style.” Automakers should spend their energy on “emotional value,” where there is no common standard. The sole exception is Li Auto: it agrees autonomous driving is not worth developing in-house, but wants to use it as a way to build toward AI, benchmarking Google, Microsoft, and OpenAI.
The endgame timeline is explicit. Autonomous driving is “a problem that is about to be solved”: 100% hands off in 3 years, eyes off in 5, mind off in 10, with L4 arriving within 5-10 years. “The battle may be over within the next 3 years.” He also discarded his earlier willingness to remain second: “The long-distance strategy is to stay second for the long haul, then sprint ahead on the final lap. We’re on the final lap now.” Journey 6P is entering the highest-compute tier for the first time this year—“there is only one player in this tier”—with the goal of making HSD the world’s best passenger-vehicle autonomous-driving system by year-end.
The next hardware disruption will come from a new computing architecture. The human brain delivers 5,000T of compute at 20 watts—10 times Journey 6P—while a single 6P chip already consumes more than 100 watts. Robots cannot support the von Neumann architecture; the next 10 years will require large-scale innovation in compute-memory integration, compilers, operating systems, and instruction sets. Horizon is not yet actively pursuing robotic use cases, choosing to build an ecosystem instead: “To be honest, we’re still figuring out this strategy.”
The organizational and entrepreneurial lesson is that scientists usually fail on the first question of business fundamentals: “Who is the customer?” They spray in every direction across 360 degrees. Horizon spent 5 years completing the 0-to-0.1 trial-and-error phase in Zeng Ming’s framework. In 2019, it chose nine out of ten, cut 600 people—half the workforce—while holding more than RMB3B in cash, and concluded: “If we hadn’t made that decision, we would definitely have died and never even made it to the table.” The auto industry exploded the following year. “Dumb luck is also one of the core competitive advantages in entrepreneurship.”

Deep dive

1. The Opening Thesis: The Essence of To B Is Putting Yourself in Others’ Shoes

Xiaojun opened with the obvious question: How can a scientist be so good at dealing with people? Yu Kai’s answer came down to the business model: “To C products require you to make something for yourself; To B products require more putting yourself in others’ shoes.” Think from the other person’s perspective. “If you asked me to make a To C product, I might not be able to.”
He also classifies Baidu Research as To B: Search, advertising, and community product teams would knock on the door and ask, “What do you need? What can we solve for you with deep learning?” Research itself was a delivery business.

2. Every Circle Starts by Breaking Through as a Nobody

At the start of his academic career, he was “absolutely a nobody,” chasing senior scholars with papers while “the big shots wouldn’t even pay attention to me.” He even mocked the fortune-teller’s verdict: “Before 24, unknown and laboring in vain.”
He directly compared that experience with selling chips: In the early days of Journey 2, Horizon knocked on automakers’ doors asking, “Why don’t they understand us?” Every level-up was a breakthrough. “That mindset was just like a young person moving to Beijing to make it.”

3. From Painter to Physics: A Company’s Core Competitiveness Is Taste

Until middle school, he wanted to attend Zhejiang Academy of Fine Arts and become a painter, drawn especially to Impressionism. He switched to physics in high school. Art never became his profession, but it still nourishes him: “A company’s core competitiveness is still taste—what you choose to do, what you choose not to do, and whether your trade-offs follow a standard.”
The evidence is Horizon’s SuperDrive HMI, which just won a German iF Design Award—“probably a first for this industry.” Technology, the humanities, and aesthetics have to be combined.

4. 1995: Machine Learning Found Me

As a sophomore, he read an English paper on speech processing with neural networks. “It was like being electrocuted…I fell in love.” He found the relevant books and read them obsessively through the night. “Sometimes I feel that machine learning actually found me; it wasn’t me who found it.”
He told himself, “I’m going to work in machine learning for the rest of my life.” From his undergraduate thesis to Siemens, NEC, Baidu, and Horizon, 3 countries changed, but “one thing never did.”

5. Don’t Follow the Crowd: The Hardest Course, Ranked First

He enrolled in information physics but was reassigned to the renamed Department of Electronic Science and Engineering, a program changed to make employment easier. He was “extremely unhappy”—renaming it for jobs ran counter to his interests. He continued taking mathematical physics and ranked first in his entire class in Nanjing University’s hardest course, Liang Kunmiao’s Mathematical Physics Methods. “The harder the course, the more I liked it.”
He connects that thread to the present: Deep learning is not abstract mathematics floating in the air; many early practitioners came from physics. “Last year Jeff Hinton won the Nobel Prize in Physics.” First-principles reasoning is the physicist’s way of thinking: derive from the essence. The conclusion may be counterintuitive, but it is often closer to the truth—and that matters enormously in business.

6. Munich: The Class-Cutting King and the Nutrients of Philosophy

In 2000, he went to the University of Munich on a German scholarship and used it to bargain his way through Professor Xu Guolin, finishing a 2-year master’s degree—“still the only one from Nanjing University to have done it.” Choosing Europe over the US was another contrarian move: UIUC and CMU were “cornfields” and “steel-industry cities”; Munich had the Alps, castles, and bars. “America is one giant boring countryside.”
He skipped “60-70% of university classes.” One professor asked in class, “Who are you?” to a roomful of laughter. He used the saved time to read informal books on society, history, religion, and aesthetics. He believes that was the foundation of first-principles thinking.

7. Advice for the Young: Don’t Study AI—Study Philosophy

His unexpected judgment: “These days I don’t really recommend that young people study AI or computer science. This discipline may basically be coming to an end.” Like databases and communications before it, AI was once a hot field; now AI can write its own code, so rushing into the hottest area is not necessarily wise.
His own generation is the cautionary example. When he returned to China in 2012, he could not recruit AI talent: AI students at Peking University, Tsinghua, and the Chinese Academy of Sciences had moved into finance, telecom regulators, or government jobs. “They chose such a good direction when they were young—why didn’t they stick with it? Because it was hard to find work. But I don’t care.” His alternative: “Study philosophy—it makes people more open, more lucid, wiser, and less troubled.”

8. NIPS 2002 in Vancouver: Two Titans Arguing and Sutton Eating Alone

At his first NIPS in 2002, there were “only around 300 people, nobody cared” (now there are tens of thousands). At a Japanese restaurant, Yann LeCun and Jeff Hinton across from him “argued nonstop.” Getting to the essence of a question and repeatedly drilling down was normal in academia; it did not prevent them from being good friends.
The most vivid scene: “A person sat across from me, completely silent. Nobody paid attention to him. He just ate alone.” His field—reinforcement learning—was not respected by the group. “His name was Richard Sutton,” who won the Turing Award just 2 years ago. What makes scholars lovable? “They seek truth and follow their passion. They don’t give up just because the mainstream doesn’t applaud them.”

9. NEC Lab: One of the World’s 4 or 5 Deep-Learning Strongholds

He joined NEC’s US research lab in 2006, the home of SVM pioneer Vapnik and a center of shallow learning as well as deep learning. “At the time, there were probably only 4 or 5 institutes in the world doing deep learning.” The code he inherited had been written personally by Yann LeCun. Torch was also developed at NEC and was “more popular than Google’s TensorFlow” at the time.
His ambition then was to become “a scientist people respected.” He published more than 100 papers, working through nights and dreaming about deriving formulas in his sleep. His self-mockery is unfiltered: “Sometimes, late at night, I’ll pull out my old papers for a little self-admiration. Still do.”

10. The First to Return in 2012: Opening the AI Talent Repatriation Wave

“Back then, nobody doing AI in the US was coming back.” Pay was higher and research conditions were better. He considers himself the first senior Chinese AI scholar abroad to return to China and the first to join an internet company. Baidu recruited him through a headhunter, with compensation that “fully matched the US—there was no premium at all.”
His logic was simple: China’s internet sector was already showing world-class potential, with data and users—the best setting for AI—while no one in China had built a machine-learning research team. “I wanted to be the first.” US-trained PhDs later began joining BAT and Huawei in waves. “All of that started when I joined Baidu and built the team.”

11. Those 10 ImageNet Points: Only I Knew How Hard They Were

He won the first ImageNet competition, which drew more than 30 teams. The next year, believing there was no new method and nobody could surpass him, he “didn’t even pay attention.” In the third edition, Hinton and 2 students raised accuracy from 75% to 85%. “Most people had no feeling for those 10 points. I was struck by lightning because nobody understood better than me how hard that competition was.”
He immediately wrote to propose a collaboration. Hinton asked for $1M in research funding, and he agreed on the spot. Then “the old man was not only good at research—he had excellent commercial instincts too”: Hinton wrote back asking whether he minded if he also asked other companies. “Of course I minded, but I had to pretend to be magnanimous.”

12. The Secret Lake Tahoe Auction: $12M Opening Bid, $44M Final Price

In the secret auction in December 2012, he rated his chances as poor—not only because the others had deeper pockets, but because Hinton had never been to China, knew only him, and could not fly because of a back injury. “Of course, you could pick him up by private jet, if Baidu was willing to pay.” His strategy was to bid first: “Put up a high enough number to scare everyone else away.” The opening bid was $12M.
Microsoft and DeepMind dropped out after bidding slightly above $20M. His authorization ceiling was $24M; anything higher required repeated approval from China. He ultimately competed with Google up to $44M. “The old man may have developed a conscience—the money had gone beyond what he could imagine.” He stopped bidding and went to Google.

13. Reviewing the Table: DeepMind’s Bearing and the Ilia Misread

He had always assumed the fourth bidder was IBM. It was only 10 years later, when a journalist wrote a book, that he learned it was DeepMind—a London startup founded only a year earlier that entered the auction with stock. “A small startup daring to join a world-class auction…even then it already had the bearing of a king. Its later rise is not surprising at all.”
He had wanted to hire both of Hinton’s students. Alex was a hacker who could make GPUs sing, solid and methodical but “later faded into obscurity.” Ilia was intellectually lively, but “I wondered whether he was all talk. I gave Ilia the lowest score.” He later became OpenAI’s chief scientist and the “father of ChatGPT.” His conclusion: 4 companies and 3 people at that table formed today’s AI leadership landscape. “The other company was Horizon, which has also fought its way at that table to this day.”

14. Losing the Auction, Winning Anyway—and 7 or 8 Hours of Circling with Li Deng

On the flight home, he ran into Microsoft’s Li Deng. Each suspected the other had participated in the auction, but professional ethics prevented either from asking directly, so they “went around in circles for 7 or 8 hours”: “Hey, why didn’t that old man Jeff show up at the venue? What did he go do?” By the time they landed, he knew the other side had bid.
Losing the auction made him a winner: “I let Robin Li see that world-class companies were willing to pay so much for 3 people. It proved how important this was.” In January 2013, Baidu established its Institute of Deep Learning—the first in China. The auction also “drove up AI talent compensation worldwide”: before then, $100K a year was considered good for a returnee; afterward, the benchmark shot toward $1M.

15. Recruiting Andrew Ng: Selling Sugar Water or Unlimited GPUs

In early 2014, he ran into Andrew Ng at breakfast at the Sheraton in Palo Alto. Ng looked “a little dazed” while working on Coursera, and his repeated CEO changes suggested trouble. Yu borrowed Steve Jobs’ famous Pepsi pitch: “Do you want to keep selling sugar water, or come back and build technology that changes the world?”
What really moved Ng was another line: “Andrew, if you join Baidu, you can buy as many GPUs as you want. You can train neural networks of unlimited size.” Google Brain’s first generation was built entirely on CPUs, following Jeff Dean’s attempt to replicate MapReduce’s success. The inability to buy GPUs was one of Ng’s frustrations at Google Brain. They finished negotiations in a month. Big deal.
The cruel turn came later. Ng joined Baidu in May 2014. When Yu flew to Silicon Valley in April to arrange his onboarding, he confessed: “Buddy, I’m actually leaving Baidu.” Ng was shocked: “You lured me over, and I only know you at Baidu. I can barely speak Chinese.” After being persuaded to stay, Yu spent exactly one year helping him get started before leaving at the end of May 2015.

16. The GPU Call: Baidu Was Nvidia’s Largest AI Customer Globally

In 2012, he bought GPUs aggressively at Baidu for parallel training. Nvidia China’s sales team told him: “You are Nvidia’s largest AI-application customer in the world.” “Even Jensen Huang had not realized that Nvidia was sitting on a gold mine.” He cites The Nvidia Way: Huang only realized in 2014 how suitable GPUs were for deep learning.
From the efficiency gap between GPUs and CPUs, he made the key deduction: “Software and hardware are inseparable.” Autonomous driving and robotics therefore needed dedicated hardware. “The derivation was completely natural and logical. Whether or not Horizon had been founded, this was where the world was inevitably going.”

17. Leaving Baidu: 3 Questions and a 5-Hour Retention Dinner

Before leaving, he asked himself 3 questions. Would each wave of innovation be bigger than the last? (Sina at $1B, Shanda at $5B, Baidu at $50B, Tencent at $500B—“each multiplied by 10.”) Did he want to watch from the sidelines or jump into the current? (He had argued with a professor over whether playing football was more fun than coaching it: “I thought playing was more fun.”) The answer was: “Make choices, don’t make comparisons.” He walked away from a stock grant so large he had “never seen anything like it,” without taking a cent or waiting for it to vest.
That night, Robin invited him to dinner. They talked for 5 hours in a private room at Xishan Hot Spring Hotel. An old hand had given him the key advice: “If you’re determined to leave, never say anything bad about the company. The moment you do, the boss will say he can fix it, and you’ll have no way out.” So for 5 hours he said only good things about Baidu. “By the end, neither of us could say anything. We walked from the second floor to the first. He got in his car; I walked away.”

18. Build an Era, Not a Product: Wintel, ARM, and NVIDIA

Why not build a car directly? “Which had a bigger impact on the world, Lenovo and Dell or Microsoft and Intel? I want to push an era.” The PC era was led not by Dell, Lenovo, or HP, but by Wintel. The mobile era’s foundation was ARM, Android, and Qualcomm. In the AI era, the server side is NVIDIA. “Enabling every car in the world to become safer and smarter” is more exciting than building a brand.
He is equally direct about his limits: “If you asked me to build a specific brand, I don’t think I could beat Xiang Ge. What am I good at? Getting customers done.” One aside: “Xiang Ge is the person in the auto industry most likely to build Apple.” His iteration ability, understanding, and vision are all exceptional.
Chips were extremely unfashionable then: capital-intensive, long-cycle, and slow to make money. Investors did not want to fund them; in 2019, many CEOs of the first batch of STAR Market chip companies were “grandpas in their 60s.” Baidu management could not understand joint software-hardware optimization. “If software is so profitable, why make such boring chips?” He says, “If Baidu had supported me then, I would not have left to start a company.”

19. The Contrarian Method: Consensus Is Either Wrong or Worthless

When he left in 2015, all 4 leading Chinese AI startups were doing software and algorithms, while he chose chips. The logic was complete: “Consensus is either wrong—most people think about the future by inertia, but the future is nonlinear, so linear forecasts must be wrong—or consensus is right, but because it is consensus, you have no differentiation and no value. So consensus is either wrong or worthless.”
The resulting 3 questions are the episode’s central theme: “What is your business secret? What have you seen that others haven’t? Does this world have a bug—some narrow gate to the future that most people are ignoring?”
The company’s values condensed into 8 words: global customers, and enduring loneliness. “Most people want to go where it’s crowded. We’re willing to go where no one has been…We turn something non-mainstream into the mainstream because we make it mainstream.” He also criticized the exam-oriented competition of the Thousand团 War and Hundred Models War: “A hundred companies rush into one field at the drop of a hat.”

20. The 3 Investments of 2015: Nvidia, Tesla, and Horizon

“When I founded Horizon in July 2015, I made 3 investments: I bought Nvidia, I bought Tesla, and then I put my whole heart into Horizon.” On the day Horizon was founded, he checked Nvidia’s market cap: $10.7B. “Now it’s $3T.” Did he keep holding? “Of course.”
He persuaded several friends to buy Nvidia. One, now a Google researcher and one of the early proponents of chain-of-thought reasoning, later told him: “Brother, do you know? My position in my family today is all because of that one sentence.”
His choice of direction was also a process of elimination. By 2015, he judged Nvidia’s cloud momentum and ecosystem “unshakeable.” “We didn’t care about servers or cloud computing. Nvidia could take it.” The opportunity was robots away from the cloud and everywhere in the physical world. “Cars are the first robots to land.” Hence Horizon Robotics. Most of the later chip startups rushed into the cloud. “We’re doing pretty well now. We don’t have much competition.”

21. The First Round Without a Deck, the Second Rejected by 50 or 60 Funds

His first investor was Linear Capital’s Wang Huai, who had planted the seed over dinner in 2013: “If you ever want to start a company, remember to call me.” He later brought in Hillhouse, Liu Qin, and others. “I raised the first round without writing a single page of a deck. I thought, life is so easy.”
The second round, in the first half of 2016, was a shock: “I met with 50 or 60 institutions, and not one placed an order.” Investors asked why he was not doing facial recognition or security, where results could appear in 3 months. “I have no interest.” Chips were like “throwing a stone into a pond and getting barely a ripple. We couldn’t even make a sound.” He talked until he was parched and exhausted; no one was interested.

22. The Funnel Model: Never Meet an Investor in Their Office the First Time

From the darkest period came his iron rule for fundraising: Investors make decisions through a funnel—hear about you, learn about you, become interested, research you, visit and compare, then place an order. When you knock on their door, they are at the top of the funnel and you must spend huge effort persuading them. So he fixed the rule: “The first time you meet an investor, never meet in their office. Always meet in mine. If they knock on my door, they are already lower in the funnel and close to making a move.”
The execution was almost theatrical. Of 10 firms that came knocking, he first turned away 2 or 3: “I don’t have time. I have a lot of customers.” Then he acted like “a focused, socially awkward scientist” and drove away another 3 or 4. The remaining 4 arrived saying, “I’ve done deep research on your field and visited A, B, C, and D.” Their willingness to invest was high, and each wanted exclusivity. He refused: “I need to negotiate with all of them. I’ll go with whoever moves fastest and offers terms I’m most comfortable with.” The round was done. “It was a strategy adjustment, not a change in myself.”

23. Intel Hedges Mobileye with Horizon; China’s Chip Capital Misses the Wave

Intel first partnered with Horizon on an advanced-driver-assistance POC. After acquiring Mobileye in 2017, CFO and later CEO Bob Swan was not convinced Mobileye could win in China, so Intel decided to “bet on both teams.” Its investment in Horizon was “a hedge against the China risk of acquiring Mobileye.” Yu considers the strategy correct.
SAIC and SK Hynix entered from late 2018 to early 2019, with rounds of several hundred million dollars. The telling footnote: “Most domestic capital investing in chips did not invest in us. They couldn’t understand Horizon’s model.”

24. The First 5 Years in Darkness: 360-Degree Shooting and “Losing Your Mind”

“It felt completely dark…The business was not going well, the organization was not going well, people were leaving, morale was low, and we could not see the direction forward.” The root problem was typical of scientist-founders: starting with a grand vision and technology but “having no concept of business scenarios.” Across automotive and AIoT, they explored toys, appliances, and air conditioners. Resources were spread across too many directions, each impossible to penetrate. It was shooting in every direction rather than concentrating every cannon on one mountain.
The most painful detail came after the SAIC and SK round, when the company was not short of cash: “I actually felt uneasy. I didn’t have that exhilarating feeling.” At year-end, when criticizing a business head, “I had no confidence myself, because I knew the strategic target was wrong. The fault was mine.”

25. Choose 1 of 10: The One-Month Revelation That Cut 600 People

In 2019, he launched 6 months of internal strategic debate. “Dozens of meetings—some ran until dawn without a conclusion, some colleagues cried in the room.” Trade-offs are “a deeply anti-human thing.” By November, he was certain: cut every business except automotive. “Rather than dig so many small pits, we should dig one large oil well.” The company had around 1,200-1,300 people; 600 were cut, nearly half.
In execution, a late-night realization changed the plan. The original idea was to make some cuts before the New Year and some after. He woke from a dream and realized that staggering the layoffs was “essentially about making ourselves comfortable, not helping employees.” So they completed the cuts in one month, from December 15, 2019 to January 15, 2020, while paying year-end bonuses and N+1 severance. He stayed and personally spoke to 10 or 20 people. “This was the company’s fault, not a failure of anyone’s performance.”
2 details followed. Those being laid off worried less about their own jobs than “how to communicate with customers and protect customer interests.” “Horizon has made many mistakes in its history, but at least its values are sound.” Outside reports said Horizon could not survive; in reality, “we had more than RMB3B in cash.” It was an intentional choice, not a decision made at the edge of a cliff.

26. 3 Questions of Business Fundamentals and Zeng Ming’s Framework: 5 Years to Go from 0 to 0.1

His 3 first-principles questions for business are: Who is your customer? What is the customer’s pain point and need? What do you have that is difficult for others to replicate in meeting it? “The first question is more important than the second, and the second more important than the third. Scientist-founders usually fail on the first question: They don’t know who the customer is, so they shoot in every direction.” Li Auto is the positive example: its customer is the father, and everything is built around him.
Under Zeng Ming’s 4-stage strategy framework—0 to 0.1 trial and error, 0.1 to 1 formation, 1 to 10 breakout, and 10 to N maturity—opportunity-driven startups such as Xiaomi complete 0 to 0.1 on day one. Vision-driven scientist-founders want to “restore justice to the world,” but “you need to see an injustice before you can draw your sword.” Horizon took 5 years to find its opening: all in on automotive.
His counterfactual is unequivocal: “If we hadn’t made that decision in 2019, we definitely would have died and never made it to the table. The only reason you can do one thing well is that you put more effort into it than others—not that you’re smarter. Without saturation bombing, there is no chance.” The auto industry exploded in 2020 and mass production arrived. “Dumb luck is also one of the core competitive advantages in entrepreneurship.”

27. Changan: Creating Mutual Success at the Hardest Moment

In 2018, Horizon and Changan established a joint lab and “formed a revolutionary friendship.” In 2019, they launched Journey 2, the first automotive-grade chip. In 2020, Changan UNI-T achieved the first mass-production deployment—selling more than 10,000 units in its first month and becoming a nationwide hit.
He stresses the timing: Changan was going through its hardest period. Changan Ford had long ceased to work as a “cash cow that could no longer provide transfusions.” Forced to stand on its own, Changan became China’s top domestic brand by sales in 2020. “We helped precisely when it was at its hardest.” During joint development in the peak summer heat, both teams were so exhausted they slept outdoors at the site. When a colleague suffered heatstroke, Yu and Changan executives visited him in hospital. He also told his team to “lose elegantly and inconspicuously on purpose” when playing football with customers.

28. Li Xiang: The Classmate in a Military Coat Who Couldn’t Climb the Mountain

On the first night of a Hupan class, the group climbed a tea mountain in Hangzhou. Li Xiang “couldn’t make it up…He waited for us at the foot of the mountain in a military coat.” Li Auto was struggling to raise money, had only a few months of cash, and he did not dare tell the company. “The pressure had crushed his body.” During that climb in early 2019, Li suggested that Yu focus on automotive. Yu “remembered it for a long time, but took more than half a year to understand it.”
The partnership began with Mobileye’s arrogance. Li Auto wanted to localize Li ONE for Chinese road conditions, including construction cones, but foreign incumbents said, “We’re the best. Use us or don’t.” Li decided to switch. “Many automakers had run into similar problems but didn’t dare to switch. Horizon was a nobody—why trust it? Xiang Ge dared to make the call. That took courage, even wisdom.” The program began in September 2020; Journey 3 replaced Mobileye in the Li ONE facelift and entered mass production in May 2021. Replacing Mobileye in 8 months set an industry record. Monthly sales rose from 2,000-3,000 to 7,000, then above 10,000. Journey 5’s first mass-production customer was also Li Auto, with the L7, L8, and L6 all becoming hits. Journey 6 will enter Li Auto production again in May.
Yu’s assessment of Li: “A learning and evolving machine. Every time I see him after 3 months, he has iterated again.” It is not just the product; organization, strategy, and talent all evolve. “From the perspective of someone who has worked in AI for 30 years, his thinking in that conversation was exceptionally sound.”

29. $1.6B in Flat-Valuation Financing: The Life Gate and the Death Gate

From the second half of 2020 to the first half of 2021, capital was euphoric: SPAC listings and SaaS concepts fetched huge valuations. He judged that “the fundamentals simply could not stand, and the bubble would definitely burst.” So he went the other way. After exhausting the roughly $800M authorized by shareholders, he applied to expand the amount and “raised another $800M at full speed.” The entire $1.6B came without increasing the valuation by a single cent. Most companies were raising $200M-$300M and repeatedly marking up their valuations.
In the following 3-year capital winter, “most companies could not raise money. We had RMB10B-plus in hand and were not anxious at all.” His method: “Most entrepreneurs focus on the life gate. Horizon always spends 70-80% of its energy thinking about where the death gate is, then staying as far from it as possible.” The by-product was a smooth IPO: because the private-market valuation had not been inflated, “every shareholder made money,” unlike companies that fought with their shareholders before listing.
Asked whether repeatedly going to the edge of the cliff, as Jensen Huang did, was necessary—and whether staying away from the cliff was too conservative—he replied: “Something may look risky to others, but if you see its essence clearly and attack it with thunderous force, you know in your heart that it carries no risk. Other people thought chips were hugely risky. I was absolutely certain.”

30. The Tenfold Speed Gap: Sprinting While Running Software and Hardware in Parallel

What is Horizon’s death gate today? “The product is not good enough. Autonomous driving has entered an era of tenfold, generational-speed change…There are not many such moments in technology history. Once one arrives, the only thing you can do is sprint.” The model is Wintel in the 1990s: 286, 386, and 486 successively crushed competitors. “IBM thought 286 was enough; by the time it realized it had to sprint, it was too late.” Companies that failed to keep up with Android and Qualcomm in mobile were also eliminated.
The structural advantage is that “only Tesla, Huawei, and Horizon are doing both chips and software.” Software and hardware development have shifted from serial to parallel, so “our iteration speed has to be the fastest in the industry.” His positioning: “Horizon has reached the stage of challenging for the World Cup final.”

31. 4 Chip Generations: From 28nm to 560T, Advancing with Trepidation

Every generation is difficult because every generation overturns the previous one: Journey 2 at 28nm, although the team had previously built only 40nm chips; Journey 5 at 16nm, with compute jumping from 4-5T to 128T; and Journey 6 with 6 chips in one generation, reaching a maximum of 560T. “Every step up was taken with trepidation, as if walking on thin ice.”
His warning to later entrants followed naturally: “After 10 years in automotive-grade chips, I have a sense of awe. For most chip companies, the first and second generations are tuition payments.” His battlefield metaphor is pure Horizon: “When the battle opened, we had built all the fortifications and bunkers in the center of the battlefield and mounted all the heavy artillery. The other troops were still on the way.”

32. The Toughest Customer Is XPeng; BYD Slipped Through the Door

More than 90% of Chinese automakers are now customers, and Horizon has become a mainstream autonomous-driving compute provider. But “I still haven’t won XPeng.” He understands the reason: XPeng treats autonomous driving as a core competitive advantage. “If I become the support for it, its fulcrum disappears. That naturally makes it the hardest to win.” His strategy is slow cooking: “Check in from time to time—How are things? Sales are good; autonomous driving has upgraded again? Great, great. Any chance for us? Open a window for us?” XPeng gives him vague answers. His self-reflection is telling: “Being rejected by customers is normal. It still means we’re not good enough. If my autonomous driving were 10 times better than XPeng’s, it would use mine. The customer can mistreat me a thousand times, and I’ll still treat it like my first love.”
BYD was a classic window-of-opportunity play. In 2022, he met Wang Chuanfu and reached a strategic partnership. “Pragmatic teams naturally find it easier to reach agreement—Li Auto is pragmatic, Changan is pragmatic.” The real breakthrough came when another supplier encountered delivery problems. “The door opened a crack. We slipped through in one burst, then pried the whole door open and gave competitors no chance.” Wang later personally endorsed Horizon—the first time an automaker had ever stood up publicly for a supplier.
The current position: Changan, Geely, and BYD have all announced autonomous-driving parity this year. “We are the largest supplier behind it, the only domestic supplier.” Many other automakers have not yet announced, but Horizon is already the most important supplier behind them.

33. The Contrarian Prediction: Automakers Will Not Develop Autonomous Driving In-House

The sharpest industrial call in the episode: “Most automakers disagree with me. Automakers will not develop autonomous driving in-house in the future because it is a standardized feature.” His baseband analogy: Have you ever seen a phone company develop its own calling function? Motorola and Nokia once developed their own basebands; today, “even Apple does not develop its own baseband.” The autonomous-driving experience is the same for everyone: getting from A to B safely, comfortably, and efficiently. “You can’t drive with Guo Degang’s style or Lin Zhiling’s style.” There can be rankings, but not truly different experiences.
Automakers should spend their energy where phone makers compete on cameras. “Photography has no common standard. It is emotional value.” That is why Xiaomi and Huawei invest so heavily. Standardized autonomous driving should be handed to suppliers. “We can revisit this view 5 years from now.”
Asked whether this meant Li Auto could not build AI, he explained the mutual understanding: “I’ve discussed this with Li Xiang. He agrees autonomous driving is a standard feature and not worth developing in-house. But Li Auto’s ideal is not to be a car company. It wants to compete with Google, Microsoft, and OpenAI one day. Autonomous driving is its path to AI, its way of building capabilities through combat.” Would he accept Li Auto eventually ending the partnership for that reason? “Of course. Building cars is only one step for him.”

34. Robots and the Next Chip: 20 Watts, 5,000T, and the End of von Neumann

Robots are the next decade’s opportunity, but he unusually admits that the strategy is unfinished: “Autonomous driving has basically converged…Robotics should have an ecosystem like CUDA, but to be honest, we’re still figuring out this strategy.” The schedule: sprint for the next 5 years to become number one in autonomous driving and 10 times better than number two; then develop an open robotics ecosystem over 5-10 years. The robotics strategy remains to build an ecosystem and wait for the opportunity, not actively pursue use cases.
The technical roadmap is the most concrete section: Data centers have nearly infinite power and cooling, while cars and robots run on batteries through extreme cold and heat with constrained power. “The human brain can deliver 5,000T of compute at 20 watts, 10 times Journey 6P, while a single 6P chip consumes more than 100 watts.” The answer is a large-scale redesign: move away from the current von Neumann architecture and fully merge compute with memory. Compilers, operating systems, and instruction sets will all be completely different. Autonomous driving is barely acceptable on today’s technology curve, “but it absolutely will not work for robots.”

35. The Timeline: Hands Off in 3 Years, Eyes Off in 5, Mind Off in 10

His prediction: “Autonomous driving is a problem that is about to be solved. L4 within 5-10 years. 100% hands off in 3 years, 100% eyes off in 5 years, 100% mind off in 10 years.” An engineer from Haidian will get into a car after work on Friday, play games and sleep, then watch the sunrise over the Qingdao coast the next morning. For robots, useful robots will appear in limited scenarios within 5 years; today’s humanoids dancing and boxing are “performances.” A form of general-purpose robot will arrive in 10 years.
The 2025 battle: Journey 6P is entering the highest compute tier for the first time. “You know there is only one player in this tier.” The goal is to win this battle and enter the main hall. By year-end, HSD, Horizon SuperDrive, should be “the best passenger-vehicle autonomous-driving system in the world.”
Xiaojun challenged him with his old line that being second was fine. He denied it while remaining consistent: “The long-distance strategy is to stay second for the long haul, then sprint ahead on the final lap. We’re on the final lap now.” There will not be many players at the table. “The battle may be over within the next 3 years.”

36. Liu Bang, Rhett Butler, and Margin: Strategy Is Fundamentally About Not Betting

His leadership role model is Liu Bang: “He knew how to maneuver, when to advance and retreat, and how to stay flexible, but his will was extremely strong.” His favorite movie character is Rhett Butler in Gone with the Wind: He moved comfortably between North and South during the Civil War, got caught speculating and landed in prison, then had the prisoners playing cards with him every day. He was living very comfortably.
This aesthetic becomes one word in business: margin. “My name is Yu, after all—Yu Kai, Horizon, and margin. Always leave yourself room in how you conduct yourself and your affairs. We never do anything that makes cash flow unhealthy. We never believe that the braver side wins when two armies clash. Strategy is fundamentally about not betting—just as Buffett says the essence of investing is not losing money. It sounds incredibly boring, but he can tap-dance to work.”
He rejects all identity labels: “Among Chinese entrepreneurs, I admire Duan Yongping most—forgetting each other in the jianghu. My favorite investor is Buffett. I would rather not become any kind of ‘master’ at all.”

37. AI for the Physical World: Let Machines Be Machines and Humans Be Human

The farthest future he can see is not optimistic: “Humans will inevitably move toward a direction I do not want to see—being raised in captivity by AI and algorithms.” Food-delivery couriers are controlled by algorithms physically; Douyin and other digital content control people at the level of the mind.
Horizon has chosen the opposite path: “We are firmly building AI for the physical world, not the digital world. It is cleaner. It truly frees people from boring, exhausting, and dangerous physical labor.” Beijing commutes and Foxconn production-line workers who are timed even when they go to the bathroom are examples. “Let machines be machines and humans be human.” Let people chat, drink, do scientific research, and pursue curiosity.
His next major bet follows from this: fully solve autonomous driving. People spend 2 hours a day on the road—one-eighth of their 16 waking hours. “If our technology gives every person the equivalent of extending their life by one-eighth, I want to get this done first, with total satisfaction.”

38. DeepSeek and Globalization: China Lacks Idealistic Companies

On this generation of young founders at DeepSeek: “They are remarkable. First, they are idealistic. Second, they have a very clear strategy and can connect reality with ideals rather than treating the two as contradictions.” China does not lack companies that know how to make money; it lacks idealistic companies.
His globalization view is simple: Deglobalization is local and temporary; globalization is the larger trend. Horizon is committed to globalization, aiming for the position occupied by Microsoft, Intel, Nvidia, and Apple. The test is: “Has the world changed because your company existed, compared with if it had never existed?” He believes China still has not produced a company that advances the technology and civilization of all humanity.

39. The Program Is Written, but You Can Change the Script by Moving Up a Dimension

His most out-there rapid-fire answer: “This world is a written program. I’m becoming more and more convinced of it. Everyone is acting according to a script.” But the script can change. At one dimension, life looks like a maze with dead ends everywhere. After moving up a dimension, as Shakyamuni and Zhuangzi did, you see that everything connects. The script is predetermined, but in a multidimensional world there are multiple scripts waiting for you. Xiaojun summarized that the only way to change the script is to change yourself. Yu replied, “Change what you do, right?” His own dimensional shift was entrepreneurship: “Because it was so difficult.”
His life details echo the same idea. He has no Beijing household registration and still rents his home. He does not play Texas hold’em, golf, or guandan. “I’d rather live a simpler life and throw myself completely into one thing.” Required reading: the Tao Te Ching and the Diamond Sutra. His ideal day: “If I hear the Way in the morning, I can die in the evening. If one day I suddenly understand some truths and come closer to the truth of the world, I’ll be happy all day.” Best food: stir-fried pork with chili peppers.

40. Bonus Material: Andrew Ng the Action Man, Brain’s Thousandfold Miss, and the Guan Yu Avatar

During their Silicon Valley years, he and Andrew Ng seriously considered starting a company, around 2008 or 2009. Yu casually suggested a training school for Chinese engineers that could “probably make a lot of money.” “I just moved my mouth.” The next day, Ng went to Stanford to ask about classroom rental prices and began running the numbers. “Andrew really is a man of action.” The background was the poverty of elite professors: The more prestigious the school, the more it exploited young faculty. Annual pay was $70K, with a beat-up car and an apartment. Yu once also asked whether Ng had heard of Mobileye. He had not. “Last year our market share surpassed it.”
Brain, the cuDNN author, a Mormon from Utah with many children and little money, was recruited by Andrew Ng to Baidu at 3 times his salary. He later returned to Nvidia. The book says he gave up stock worth 1,000 times that salary by taking the deal. “I thought, wow, this has something to do with me.” Another talent footnote: MiniMax’s Yan Junjie was once their intern; an investor backed him partly because “his hairstyle looks like mine.” Jia Yangqing was his intern at NEC, whose team also produced Li Mu and professors at CMU and COTECH.
The reality of early management: “Internally, we were a brotherhood, not an organization. We ran on personal relationships.” To transfer one intern to another department, Yu had to drink a bottle of Moutai with the department head. The contrast was Li Xiang: “He said, do I really need to ask for this budget? I’ve run a listed company.” Yu replied, “Okay, I have a lot of lessons to catch up on.”
2 final Easter eggs: The Guan Yu avatar was generated from his photo by a female Baidu intern using a neural network. “Look closely and you can see he’s wearing glasses”—an early deepfake. For a period, Yu called himself chief scientist rather than CEO: “When I disguised myself as a chief scientist to go out and sell, would people feel a little more sympathy for me?”

Market Overview — July 8, 2025 Is Crypto Still The Best Trade?