Vol.172 Interview with Qunhe Technology’s Huang Xiaohuang: 14 Years as a Technical Entrepreneur

Summary

• Core judgment: AI has eliminated SaaS’s know-how moat. Huang Xiaohuang: “A crucial moat in software used to be turning industry know-how into code… After this wave of AI, computers can directly understand that know-how and generate code from it, so people no longer need to do that work.” From 2023, Qunhe shifted its strategy wholesale from architecture and home-furnishing information-services SaaS to infrastructure for an intelligent world—synthetic-data training, digital-twin modeling, and robot management. Workflow-management software is difficult; workflow-based projects such as construction-site management “definitely have no future.”
• “GPUs surpassing CPUs is a law of physics,” and it will hold for 20-30 years. After being “brainwashed every day” by Jensen Huang and David Kirk as a student, he never doubted it: CPU clock speeds have stayed at 2-5GHz for 20 years, so Moore’s Law can continue only through horizontal scaling to thousands or tens of thousands of cores. The contrast is hype: “This year it’s one hot trend, next year another—one changes every year. There’s no point chasing it.” The same logic implies that every increase in compute will produce new applications: “It’s like genetic mutation—you never know what will emerge, but you need the compute.”
• “The robotics industry is definitely not a bubble.” Humanoid robots are the “high ground,” analogous to L4 autonomous driving, but “I’m not sure whether humanoids will work in the short term.” What is certain is the intelligence upgrade of every machine and device, shifting from instruction-based to task-based operation. Synthetic data went from being widely challenged by robotics companies—“How can you train on something synthetic and therefore fake?”—to gradual acceptance. The inflection point came during the pandemic, when Silicon Valley giants saw the papers and the open-source InteriorNet training set and came knocking with money: “They taught us step by step how to train.”
• Open source is a geopolitical hedge. One motivation for open-sourcing SpatialLM was the lesson from DeepSeek: “The wisdom of all humanity is still more important than that of a single company.” The other was a sanctions contingency plan: “If China were ever cut off from AI, I could train models through an overseas subsidiary and open-source them so people in China could keep using them—a perfect way around the small yard and high walls.” Open-source models do not factor in a business model. The legacy business is moving from selling accounts to selling compute, while robotics partnerships charge by the number of synthetic 3D scenes.
• His AI ecosystem view: competing on foundation models is pointless; too few companies are building applications. “So many companies are competing on foundation models, and none of them have made money. That’s not a healthy ecosystem”—like the mobile-internet era, when everyone competed to build app stores and nobody developed applications. The applications that can win are Cursor-like: “solving problems that couldn’t previously be solved, and making expensive tasks cheap and fast.” Management applications built on process know-how will be easily replaced by AI-generated code.
• His startup filter has only 2 criteria: real social value that he personally believes in, and something he genuinely likes. That led him to reject renovation loans and supply-chain finance—“I can’t hire someone to sit outside people’s homes and collect debts”—and avoid the traps of internet finance and other sectors. He also admits he cannot build a content community: “I can’t build Xiaohongshu.” Content quality cannot be measured; it felt “like managing a group of people who don’t speak the same language.”
• His talent philosophy: smart and fast at execution—“one person can do the work of 10.” In interviews, he throws candidates a paper and asks them to improve it on the spot. Experience is not required; “I can even accept an undergraduate dropout.” DeepSeek became an inspirational story: “They used to say you hadn’t invested enough money. Now it’s that your people aren’t good enough—smart people are more useful than more cards.” He exchanges ideas with the DeepSeek team. The best anecdote: “We always thought they were stock traders, while they always thought we were in home renovation.”
• Cash flow funded the transformation. Cash flow broke even in 2016 and reached roughly RMB40M-50M in 2017. Buying GPUs was the largest expense after labor, but operating cash flow covered it. The cost of the transformation was a brutal fight with shareholders over valuation: “Is an intelligent system worth 2x PS or 20x PS? How would I know?” The eventual solution relied partly on economic concessions: “Any problem you can solve with benefits is a small problem.”

Deep dive

1. Zhukezhen College: A group of smart people together matters more than class

• In 2003, he ranked around 60th or 70th across Zhejiang in the gaokao and entered Zhejiang University’s Zhukezhen College directly, 5 classes behind Huang Zheng. “At school, I’d already heard about his exploits—his grades were exceptionally good.” The college had no majors; after freshman year, students joined labs for research. He chose Zhejiang University’s CAD lab to study rendering.
• His study method: “I’d go to the first class and the last class of a standard course, and skip everything in between”—sleeping during the day and coding at night. The real gain was his peers: “It’s very important to have a group of smart people together.” As a small-town exam grinder, he saw the wider world through senior students’ experiences and ideas. “That may have influenced me even more.”
• In a class of roughly 30, about two-thirds went abroad and more than half ended up in finance or IT. His reflection: “When you’re young, a high salary feels exciting. Once you get a little older, the ceiling feels painfully low.”

2. Nvidia’s ceiling: He studied the entire company directory

• At Nvidia, he once used differences between mainland and Hong Kong-Taiwan romanization systems to study the entire company directory and count mid-level managers with mainland backgrounds. “The proportion was extremely low. Do you think these people weren’t smart? They were all exceptionally smart. That means there was a ceiling”—a “frog-boiled-slowly process.”
• The structural reason was that high-performance-computing projects required US government approval: “By the time approval came through, the project was almost over.” For key businesses involving the Department of Defense, “you couldn’t even attend meetings; all information was kept away from you.” He therefore “gave up on building a career there.” The situation can only get stricter: “Once they see you’re from the mainland, they’ll just tell you not to participate.”

3. UIUC and the CUDA lab: A PhD student funded by Nvidia

• He applied for a PhD, not a master’s—“There were no scholarships for master’s programs; everyone applied for PhDs.” UIUC was strongest in supercomputing and needed someone with a graphics background. His adviser ran a joint program with Nvidia CTO David Kirk. Nvidia provided roughly $1M plus large quantities of GPUs to study CUDA, with the scholarship wired to the university and then paid to students.
• At UIUC, he stopped skipping class. The coursework involved chip design: “Holy shit, that software was so expensive. We didn’t even have it, and you couldn’t use pirated software.” Access to the supercomputer was also limited to class time. Scarce equipment dictated the way he learned.
• Nvidia wrote to borrow him for a year to develop the CUDA system. His adviser warned that “most people who went wouldn’t want to come back.” He never returned to finish his PhD: “The work at the company was the same as at school, but the company paid more and had compute clusters everywhere.”

4. The first GTC epiphany: Jensen used 128 GPUs; I needed 4

• He volunteered at the first GTC and watched Jensen use 128 GPUs for local real-time rendering. “I was thinking, why does this thing need so many GPUs? That doesn’t seem right.” He ran the experiment himself and found that “about 4 GPUs should be enough.” That became the startup idea: physically accurate rendering had taken 30-60 minutes per image; moving it onto CUDA made it “particularly effective.”
• He articulated the startup paradox himself. Nvidia was struggling at the time: “If it had been doing well, I would either have kept working there or gone back to finish my PhD. If it was doing badly and I still kept studying this field, I’d have been crazy. But I never expected it to do so well later. That was outside the plan.”
• He built the demo in a month. His UIUC classmates became co-founders—Zhu Hao was at Amazon, Chen Hang was doing a PhD, “big tech plus PhD.” After seeing it, they felt they had “found a treasure.” “If you waited another year or two, there’d be nothing left for you, so they rushed back to start a company.”

5. GPUs surpassing CPUs is a law of physics; hype is meaningless

• The core conviction took shape while he was still a student: “CPU clock speeds are now stuck at 2-5GHz. They were roughly at the same level 20 years ago. We’ve hit the physical limit.” Only architectures with thousands or tens of thousands of cores, combined with a new programming paradigm such as CUDA, could extend Moore’s Law. “GPUs surpassing CPUs is a law of physics… It’s a law that will hold for 20-30 years. That part is deterministic.”
• The contrast: “Today’s hot trend is one thing, tomorrow’s is another; it changes every year. There’s no point chasing it. You can chase stories, but what you build takes 5 or 10 years of investment.”
• Before 2018, few people believed it. “Once everyone believes it, there’s no point talking about it anymore.” Which company succeeds is a matter of probability; at the time, he even thought Intel had the better odds.

6. His adviser’s thought experiment: With 1,000x more compute, choose physics simulation over the human brain

• A student project assumed compute would eventually increase 1,000x, with 2 possible directions: simulating the human brain or physically accurate rendering. He chose the latter: “I really had no idea how to simulate the human brain.” The CNN methodology “was probably introduced in 2009 or 2010. At least I hadn’t seen it yet.”
• At Nvidia’s lab, his first reaction to a Stanford deep-learning paper was: “How can something this slow be useful? I was still researching how to accelerate it.” That led to his meta-conclusion: every increase in compute will produce something new. “It’s like genetic mutation. You never know who will suddenly have a flash of insight and come up with something. But you need the compute. Once you have the compute, someone will figure it out. It’s only a question of how long.”

7. Silicon Valley financing failed; he returned to China through a “group tour”

• He originally wanted to start up in Silicon Valley. In 2011, after the financial crisis, “financing wasn’t easy.” When he returned to China, friends asked whether Nvidia was about to go bankrupt and whether he planned to short Nvidia. “I said Nvidia wasn’t doing well, but that definitely wasn’t why I was returning to start a company.”
• Zhejiang’s investment-promotion delegation in Silicon Valley rented out a Chinese restaurant and “bought everyone a good meal.” “Basically, everyone went for the free food and drinks, and as they listened, they became interested.” Around 30 Silicon Valley companies returned to China as a group to look around. “I think the Suzhou government paid for the flights,” while Hangzhou may have offered free accommodation. “It was still a group tour.” He added: “That would be unimaginable today.”
• They started in an empty apartment owned by a co-founder and used it for a year. “During an interview, you’d take candidates into a bedroom. It seemed completely unreliable.” Candidates suspected it was a scam company, “especially female employees.”

8. Chinese VC in 2012: Half of investors did not know what a GPU was

• VCs were investing in apps, group-buying, and O2O during the height of the “thousand-group war.” “Almost nobody looked at technology.” RMB funds would not even consider them. The 3 questions he heard most often were: “What is a GPU? Why isn’t Nvidia doing this? Nvidia itself isn’t doing very well—can you build on top of it?” “I realized that half of the investors didn’t know what a GPU was.”
• Some also asked: “What if Nvidia goes bankrupt? How would you keep your company going?” “I found that very hard to answer.” 2012 was difficult. In 2013, China’s capital-market bubble arrived and “one financing round followed another.”

9. Learning to tailor the pitch: The home-renovation platform story and FA-written lines

• Talking to investors about technology and Moore’s Law was fatal; they thought he was “talking nonsense.” The packaging that worked was a Houzz-like pain point: photographing a home cost $3,000, or roughly RMB20,000, while GPU rendering cost less than RMB200. “The ROI was so high. It could overturn the platform’s model.” Financing suddenly became easy.
• Once he learned the game: “When you meet a consumer investor, tell a consumer story.” The FA also prepared a stack of lines for him to memorize: how huge China’s home-renovation market was, how the middle class was rising. “I had never renovated a home.” But the underlying business never changed: “Whatever industry we were in, we stayed focused on GPU acceleration. I’ve lived through many cycles. It didn’t matter.”

10. The first check and the breakout: Wang Huai’s RMB500,000 and the data in 2013

• The first money came from Zhejiang University alumnus Wang Huai, a connection from volunteering together at a Silicon Valley alumni association: “We’re friends, so I don’t care what you’re doing. Here’s RMB500,000. Don’t come ask me again.” He later raised more through Wang’s introductions.
• The product took off in the industry in 2013. “The data was especially good and spread wildly on Weibo.” From 2014 onward, “a pile of investors wanted to invest every day,” while early investors blocked them. Some investors fell out with him over it. “It was a completely different kind of headache.” The issue was not price; existing shareholders believed the company did not need to raise and would command a better price later.

11. Investors urged him to build a community; he could not build Xiaohongshu

• After the breakout, investors urged him to build a Houzz-like home-renovation community. The engineering team’s frustration was concrete: one colleague stayed up until 2 or 3 a.m. writing an article that received zero shares, while another received tens of thousands. “I studied it for ages and still didn’t know how to analyze why one worked and the other didn’t. My head was a mess. I still haven’t figured it out.” It felt “like managing a group of people who don’t speak the same language”—you understand every word they say, but not what they mean.
• He brought in consultant Feng Shu to explain “what makes something beautiful.” He asked, “How do you quantify beauty?” and then had no answer. They considered using an AI model to score designs, but “that’s still pretty difficult today.” Finally, he told investors: “I’ve assembled a very good team. Why don’t they start a company themselves, and you invest in them? Let me go. I really can’t do this.” The conclusion: “I can’t build Xiaohongshu.”
• The management principle that remained: do not build something whose quality you cannot judge yourself. “You still have to give colleagues bonuses, but you can’t explain at all whether their work was good or bad. The people doing it become frustrated too.”

12. Rejecting renovation loans: He could not make a business out of debt collection

• Shareholders pushed renovation loans every day: “You can make tens of thousands on a single deal. A designer pays you only a few hundred yuan a year for software. Why bother?” He investigated and saw former colleagues starting businesses under heavy debt. “There was a debt collector sitting outside his home every day. I can’t hire someone to sit outside people’s homes and collect debts. I can’t do that.” Not being able to do it meant absorbing large bad debts. “I’d rather not make that money. Why make money outside my understanding or on something I don’t believe in at all?”
• The era was telling: every company in the building except theirs was an internet-finance company. “If you didn’t do that business, you couldn’t even raise money.” Li Xiang responded, “Just like doing AI now.” Huang agreed: “Exactly. He said what I was thinking.” Many of those lucrative sectors later blew up, including wealth-management businesses funded by delaying supply-chain payments. “I saw a lot of them.”
• That established 2 hard rules: “First, the thing has to make money through real social value, and social value that you personally believe in. Second, you have to genuinely like it. If it doesn’t meet both conditions, I won’t do it even if it makes money. The risk is too high.”

13. The starting point of the business model: RMB50,000 in pre-orders; follow whoever pays fastest

• Right after the demo, before the product even existed, he asked companies in different industries: “Would you pay RMB50,000 upfront for this?” A real group of home-furnishing companies did pay. “So we went deeper into that industry.” The selection rule was simple: “Go wherever people pay more.” The technology was general-purpose from day one. “But if you target every industry and every customer from the beginning, you can’t handle it. You need an entry point.”
• The counterexample was film. They built a post-production system, but a project from 2012 did not pay its final balance until 2016 because payment waited for box-office revenue. “If you do business in that industry, you go bankrupt. You get dragged to death.”
• Investors had no interest in selling software or APIs. They wanted the company to sell furniture sooner or get into renovation loans. He accepted one summary: “In China, the things that bring in money quickly are finance and retail. Everything else is hard work.” At least software had high gross margins. “Most Chinese industries have low margins and long payment terms. It’s painful.”

14. Cash flow and forced expansion: RMB40M-50M in 2017

• Selling APIs and cloud services kept the team alive from 2015 to 2017. “That period was quite happy—hard work mixed with enjoyment.” Cash flow broke even in 2016 and reached roughly RMB40M-50M in 2017. “Then investors forced you to expand in all kinds of directions. The pressure became enormous.” You would say growth was good, and investors would reply: “Other companies are growing 10x or 20x a year. You’re grunting away selling software, and in a year you still haven’t made as much as they make on one deal.”
• His own software-sales experience was both a joke and a fact: “You eat, drink, and talk for half the evening, then they say, ‘About your requirements…’ If you want a discount, just say so. I’ll reimburse you for dinner.” The close rate was low; many deals disappeared during negotiations. “I’m mainly a researcher. I didn’t know all those tricks, so it was easy to get played.”

15. Why Intel lost: The founder stopped running the company, and Wall Street took over

• He once thought Intel had a higher probability of success than Nvidia and temporarily ported the system to Intel’s new chips. “Intel’s chips were more stable,” and it had a CUDA-like competitor. The reason it lost: “Intel’s product line was too broad”—motherboards, displays, laptops, everything—while Nvidia had spent years doing only GPUs. “Once you get distracted, it’s easy to fall behind.”
• The deeper issue: “The founder stopped running the company, and it became a company managed by Wall Street. Today Wall Street reads a research report and says you should make embedded chips; tomorrow it says vision chips; the day after, ARM chips. How many lines can one company run?” The lesson: “You have to stay focused, find a direction you believe in, and sometimes just bet on it and keep going.”
• Li Xiang mentioned Morris Chang’s autobiography: when Chang retired, he seriously asked Jensen Huang whether he wanted to become TSMC CEO. “At the time, TSMC’s market cap was many times Nvidia’s. That’s true.” Huang’s interpretation: Nvidia and TSMC created each other’s success. Once process technology reached its physical limits, Nvidia expanded horizontally, while Intel’s in-house fabs took it down another path. He joked that if he were starting up in China today, he might build chips. “But doing that back then would have made people think you were crazy. Investors wouldn’t even want to talk to you.”

16. The confusion of 2022: AI directly eliminated the know-how moat

• The original strategy had been working smoothly: interiors, then small buildings, then large buildings. After succeeding in Industry 4.0 home furnishings, they moved into real-estate construction. In 2021-22, “the country’s entire strategy was being sharply reset.” Construction weakened as AI and autonomous driving took off.
• The key insight is worth preserving in full: “An important moat in software used to be turning industry know-how into code and building a powerful barrier over time. After this wave of AI, computers can directly understand that know-how and generate code from it. That means people no longer need to do that work. What you thought was a moat is no longer a moat.”
• “I was very confused for about a year.” That was 2022. Several things hit at once: project-management companies’ results deteriorated, while experiments with large models showed that know-how could be converted directly from text into code. “Holy shit, the world had changed. You had to adjust. If you didn’t, you could become a casualty of history. I saw companies fighting the era head-on. There was no point.”

17. Turning toward an intelligent world: Informationization was replaced halfway through

• An internal summary set the direction: “We used to think the future would be an informationized world. In reality, it will be an intelligent world. Informationization may be replaced by intelligence halfway through the process.” “You have to embrace it. Resisting or fighting it is pointless.”
• The new blueprint is a world in which intelligent robots enter factories, offices, and homes over the next 5-10 years, and devices shift “from instruction-based to task-based.” Qunhe wants to build training systems using synthetic data, multimodal rapid modeling and digital twins—feed in drawings and directly generate a 3D structure so a robot “immediately knows what every room looks like”—as well as robot training and management. “All work used to be designed for people. Now it has to serve people and robots.” He accepted Li Xiang’s summary: Qunhe aims to become an infrastructure company for an intelligent society.
• The legacy business will not disappear, but it will be downgraded: “Use one-tenth of the team, working with AI, to keep it going.” It is like customer service: “You should provide it, but you don’t think customer service is your company’s moat. The real moat returns to your algorithms, your results, and your compute.”

18. The 3 parts of transformation: Technology, team, shareholders—the hardest is shareholders

• Technologically, the center of gravity shifted to algorithms. “That year I basically stopped seeing customers. I studied papers and read code every day.” After 2022, the field moved so quickly that “you couldn’t finish reading the papers.”
• The team transition was the most painful. They had hired many people who both understood the industry and could code—previously the company’s most important employees. “Now you tell them, ‘Asking AI is better than asking you.’ That’s a huge blow.” Internally, consensus was easier to reach. A colleague joked that after more than a decade of doing work the boss did not enjoy, “he can finally do what he likes. ERP is more interesting than selling furniture, and selling furniture is a little better than making loans.”
• Shareholder communication was the hardest. Some investors thought construction was reliable, had cash flow, and that project management was still highly profitable. They accused him of chasing trends. “You explain why the technology works, but they aren’t technologists—they’re finance people. It’s exhausting.” There was little consensus in 2023. Most investors came around in 2024; a small minority still said, “If you stop doing construction, I’m pulling my investment.”

19. Fighting over valuation: “Any problem you can solve with benefits is a small problem”

• The transformation collided with a valuation reset. He reproduced the argument: “You say the future is an intelligent world, so how do you value this? Is it 10x PS, 20x PS, or 2x PS? How do you prove it to me? …I can’t explain it clearly either. I can only say higher is better, right?” Without a resolution, shareholders could vote against everything and leave him unable to act.
• His candid conclusion: “I couldn’t persuade them. Sometimes you have to sacrifice some benefits, and the benefits did the persuading. I won’t discuss the details—any problem you can solve with benefits is a small problem. It’s still better than being forced into a business you have no interest in.”
• His organizational lesson: once you realize something has already become historical dust, continuing to fight it is pointless. “You have to move with the larger era to attract people who share your goals, ambition, and ideas. Otherwise your team will age, the people joining will get worse, and you’ll slowly be淘汰 by the era.”

20. His AI application view: Foundation-model competition is pointless; workflow applications are dead ends

• His ecosystem judgment: “A huge number of companies are competing on foundation models. Investors have put a lot of money into technical-model companies, and they haven’t made any money. There are too few application companies. That isn’t a healthy ecosystem.” It is like the mobile-internet era, when everyone competed to build app stores and nobody developed applications, so the ecosystem never got off the ground. The push toward AI applications is, in his view, investors’ normal attempt to develop the ecosystem.
• His test for applications that can win: “Workflow-based things definitely aren’t what AI Agents should do.” He uses Cursor himself. “It isn’t workflow-based. It solves problems that couldn’t previously be solved. If a task used to cost a lot and now can be done cheaply or quickly, there will definitely be a market.” Management software for construction sites, files, or accounts—software built on know-how—is difficult. Construction-site and project-management software, he says, “definitely has no future,” because AI can generate the code too easily to justify spending so much on software.

21. SpatialLM open source: DeepSeek’s lesson and another route around the firewall

• One motivation was DeepSeek. “Before it came out, we never considered open-sourcing. We thought, why should something we worked so hard to build be free for others to use, especially under the MIT license, which lets anyone commercialize it?” The impact was significant: “The wisdom of all humanity is still more important than that of a single company.”
• The second motivation was geopolitical hedging. “If the US imposes an AI blockade on China, models trained in the US can’t be brought to China, and Chinese models can’t go to the US. If Chinese companies are ever blocked from training models, I can train through an overseas subsidiary and open-source the result so China can continue using it. Isn’t that a perfect way around the small yard and high walls? They can’t exactly ban open source.” He compared it with the data-separation problem facing Tesla FSD in China and the US.
• Commercially, he is relaxed about it: the open-source model “doesn’t consider a business model.” The legacy business is “gradually moving from selling accounts to selling compute.” It also has a natural defense: Qunhe’s internal system is a complex system made up of countless large and small models, so “open-sourcing one model doesn’t let you replicate it.” That differs from DeepSeek, where “the whole website is just one model.”

22. Silicon Valley giants came knocking: The pandemic turned synthetic data into a strategic priority

• The seed of the transformation was planted by customers. During the pandemic, people could not enter offices or users’ homes to run experiments. “If a stranger came to your home, would you be willing?” Silicon Valley giants saw their papers and the open-source InteriorNet training set—created after seeing Fei-Fei Li build ImageNet at UIUC—and came looking to collaborate and pay. “At first we were completely confused. They basically taught us step by step how to train.” He is clear-eyed about the context: “Given the political environment at the time, if they were willing to work with a Chinese company, they must have had no choice.” This was still only a side project.
• The industry resisted in earnest. Robotics companies initially asked: “You always feel that synthetic things are fake. How can you use fake things for training?” He expects acceptance to come gradually over the next 1-2 years, driven by academia. The initial business model is to charge by the number of synthetic 3D scenes, similar to selling compute. “Making money isn’t the priority yet. The goal is to move the whole industry to a new level.”
• The defensive logic is equally hard-nosed. Part of Qunhe’s revenue comes from flexible customization for unmanned factories. As machines upgrade from traditional robotic arms to intelligent equipment, “if you don’t secure this track, that revenue will be hit significantly. So you have to position yourself early.”

23. Humanoids are the high ground, but “the robotics industry is definitely not a bubble”

• The case for humanoids: “A humanoid robot is the only truly general-purpose robot you can imagine.” Existing infrastructure is built for people, so other forms will ultimately “converge toward a humanoid form.” But the path resembles autonomous driving: “L4 is the high ground, but you don’t necessarily start with L4. You can move up gradually from L2 and L3. Some companies will move toward humanoids step by step; others will start with humanoids. Eventually they converge at one point.”
• His boundary is clear: “I’m not sure whether humanoid robots will work in the short term, but the robotics industry is definitely not a bubble. Many robots are already working—robot vacuums, AGVs, delivery robots.” Even if humanoids fail in the short term, the accumulated supply chain, talent, and experience will be highly valuable for other, less advanced forms of robotics. He also rejects the Tesla-style vision that everything becomes humanoid: “Not every piece of factory equipment will turn into a person. Maybe it’s just an arm without legs, or a leg without an arm.”
• At the national level: “China is having a hard time in the foundation-model race because pure compute is constrained, but China definitely has an advantage over the US in robotics. The national team is clearly watching this closely.”

24. Hiring comes down to 2 things: Smart and fast at execution—“one person can do the work of 10”

• The chief scientist’s hiring criteria: “First, intelligence. Second, strong execution. Nothing else is required.” His interview method: throw the candidate a paper, have them read it, propose improvements immediately, and start building immediately. AI experience is not required: “Algorithms change too fast. One or 2 years of relevant experience means nothing to us. You can learn it here. New graduates are fine. I can even accept an undergraduate dropout. They just have to be smart.” With the same paper, one person understands it in an hour and another in a week. “The gap between people is enormous. Someone like that may be worth 10 people in this era.”
• Competing with big tech for talent, Li Xiang relayed that one major company was offering new graduates RMB3M a year and still had to beg them to join. Huang’s answer is differentiation: “Someone I consider a genius may not be considered one by ByteDance. Every company has its own criteria. You also don’t need many algorithm people.” DeepSeek became an inspirational story: “They used to say you hadn’t invested enough money. Now it’s that your people aren’t good enough—smart people are more useful than more cards.”
• The best joke of the episode: he exchanges ideas with the DeepSeek team. “We always thought they were stock traders… They always thought we were in home renovation. We’d occasionally have a meal together.” Li Xiang joked that the “home renovation” label had left him with a psychological scar. Huang replied: “People doing renovation are still better than people doing stock trading. If I had started out targeting robotics, I don’t know how many times I would have died.”

25. Open source lowers the barriers: “There are more opportunities in this era, not fewer”

• He rejects the big-company threat: “We’ve also been targeted by big companies—AT-level companies, ‘much bigger than Beike.’ It’s nothing to fear. As long as you focus on what you’re good at, unless you’re on a big company’s core track, you don’t need to worry too much.” Large models are not limited to language and image generation. Spatial cognition, speech, and digital humans are also large models. “A model isn’t a large model only because OpenAI is building it.”
• Open source is leverage. SpatialLM was trained on the open-source foundation of Alibaba’s Qwen. “It suddenly became much simpler. Script generation isn’t our core capability. There’s no point spending hundreds of millions training a model just to do this.” Stable Diffusion solved the longstanding problem of rendering plants and people. “Things that used to be very hard for me suddenly became easy. This era is still very exciting.”
• His remaining anxiety is about unknown unknowns: “Problems you can imagine can all be solved. What worries me are problems outside my understanding. That makes me uncomfortable or anxious.” He has also rehearsed the extreme case: “If humanoid robots suddenly advance dramatically and everything really becomes humanoid as Musk says, the original assumptions won’t hold. Maybe in the future you won’t ride in a car; you’ll ride on the back of a robot running through the streets. Anything is possible.”

26. Hangzhou is most like Silicon Valley; the Six Little Dragons are just the tip of the iceberg; Nvidia is his methodology

• His location logic: before returning to China, he studied several cities. Beijing and Shanghai had stronger research capabilities, but “the costs were too high.” In the most expensive places, like New York, “you can’t develop many technology companies. Everyone is too restless.” Silicon Valley’s technology companies are not in San Francisco but 1-2 hours away by car. “That’s actually a lot like the relationship between Shanghai and Hangzhou. Hangzhou is the place most like Silicon Valley. Many of us can get to work in 10 minutes.”
• He attributes the attention around Hangzhou’s Six Little Dragons to the spotlight on robotics during the Spring Festival and DeepSeek’s breakout. “These 6 companies are just the tip of the iceberg of Hangzhou’s technology-startup scene.” Who is underestimated? “Before DeepSeek became famous, who would have thought they would take off? They never appeared on the various unicorn lists. That wasn’t surprising.”
• His corporate role model is Nvidia: “Build the technology first, develop your core capabilities, then explore directions. Every product you build doesn’t need to occupy a huge position, but it has to be the strongest thing in that spot.” Secure the strongholds and prevent the company from “getting up early and arriving late.” GPUs are the largest cash drain after labor, but “normal business cash flow can cover them,” so there is no need for additional financing.
• Over the past 2 years, “120% of my energy” has gone into studying new things, while daily operations have been handed to a mature team. His 2 fastest-growth periods were 2015-16, when he first built and sold the product, and the past 2-3 years of the AI boom: “I found that feeling of starting up again and learning every day.” By contrast, 2019-20 brought “nothing new. It was boring. Talking to customers every day was dull.” His closing self-reflection quotes Jack Ma: “‘The kinder your heart, the sharper your blade.’ Researchers have a problem: their blades don’t get sharp enough. I’ve been reflecting on that too.”