Manchi | 晚点聊 LateTalk Summary

Podcast: 晚点聊 LateTalk, Episode 160 Guest: Huang Xiaohuang (黄晓煌), Co-founder & Chairman of Qunhe Technology (群核科技) Host: Manchi (曼琪)

1. Key Themes

The Survival-First Doctrine: Longevity Over Brilliance

Huang articulates a philosophy forged by watching peers fail spectacularly. The company's north star was simply staying alive through multiple industry cycles, not maximizing growth or fundraising optics. This wasn't timidity—it was pattern recognition from firsthand observation of China's brutal 3–5 year industry rotation cycles.

"I think living long is more important than living brilliantly. And many companies that look brilliant—they're burning money to look brilliant, or their foundations aren't solid—they may just be a flash in the pan." [00:18:41]

"I've been told throughout my entrepreneurial career that 90% of the star tech companies I've known have disappeared or become obscure. But we're still here. We grow every year. I think that's already quite remarkable." [01:13:17]

The Physical World as the Enduring Anchor

Huang has maintained a single strategic thread since founding: bridging the physical and digital worlds using GPU-based spatial computing. Every pivot, product line, and AI investment traces back to this core. He makes the non-obvious argument that proximity to physical reality is the moat against AI disruption.

"Personally, I believe that the closer to the physical world, the safer. The further from the physical world—purely virtual—that's very dangerous." [00:31:51]

"No matter how information technology develops, our physical world will not disappear. You live in the physical world. You don't want to be wearing a VR headset all day." [00:30:23]

China's 3–5 Year Industry Cycle Problem

Huang observes a uniquely Chinese phenomenon: investment themes rotate completely every 3–5 years, and companies that fail to adapt are not just forgotten—they become actively despised. This shapes his entire operating philosophy, including a constant readiness to abandon the current hot narrative.

"China's industry track switching is too fast. On average, it cycles every three to five years. A once-golden track or star company—if you don't adjust in time within three to five years—immediately becomes a company everyone despises." [00:18:11]

"At one point, investors were teaching us every day how to tell an O2O story. Then two years later, they'd say 'don't even mention O2O—not a single O.' That kind of massive switching process—I felt it deeply." [00:19:10]

2. Contrarian Perspectives

Algorithms Are Not a Moat in China—Data and Network Effects Are

Most tech founders treat proprietary algorithms as their core IP and competitive advantage. Huang argues the opposite: algorithms in China are dangerously fragile because they walk out the door with your engineers. Only data (which you can't copy by hiring people away) and network effects constitute real barriers.

"I think in China, building a company on a core algorithm is too dangerous. The algorithm is like a layer of window paper—one poke and it's through. Data is a real barrier. Algorithms are very hard to form into a moat." [00:23:33]

"When your algorithm team gets poached, it's copied wholesale. The rendering engine development team—except for the lead, everyone was poached and the same thing was replicated elsewhere." [00:23:04]

Raising Too Much Money Early Can Kill a Tech Company

Counterintuitively, Huang suggests that starting lean and scrappy during an unfashionable period for tech may have saved Qunhe. Well-funded companies in hot markets tend to over-expand, lose discipline, and collapse when the cycle turns.

"If you founded your company when tech was extremely popular, and you could raise a lot of money, and you had abundant resources from the start—do you think you would necessarily have developed better? Possibly you'd have actually died, because you might have become arrogant. We became somewhat arrogant too—in 2020 and 2021 we hired way too many people." [00:25:59]

2018 Was the Golden Window for Hardware + AI in China—Not Software

This is perhaps the most striking investment insight in the interview. Huang believes the biggest missed opportunity was not investing in hardware infrastructure during the 2018 trade war era, not in AI software. He implies that the real structural opportunity in China post-trade war is hardware, not software.

"If you invested in hardware—I think 2018 should have been the golden time to invest in hardware. After the trade war, the biggest opportunity is not in software—it's in hardware." [00:50:24]

"I think doing pure software in China is still quite difficult. You have to do some form of hardware. But the form of hardware now is definitely different from 2018's hardware." [00:51:50]

Video Generation Models Can Never Achieve Physical Correctness—3D Wins Long-Term

Against the prevailing consensus that video generation (Sora, Veo, SeaDance) represents the path to world models, Huang bets on 3D spatial reconstruction as the superior technical path for any application that requires real-world physical accuracy—robotics, manufacturing, simulation.

"Video generation models are fundamentally not physically correct things. They're more about visual continuity of video. When you need a scene close to the physical world, video models cannot do that. A person walking down the street—a video model can make them look five meters tall or one meter tall, and it doesn't really matter. But for a 3D scene, that's a huge difference." [00:56:09]

3. Companies Identified

Qunhe Technology / Coohom (群核科技 / 酷家乐) Description: Hangzhou-based spatial intelligence company, listed on Hong Kong Stock Exchange April 17. Operates Coohom (home design SaaS for professional designers), industrial manufacturing software, and is building spatial AI models (SpatialLM, SpatialGen) for robots and physical-world AI. Why Mentioned: The subject of the interview; first of Hangzhou's "Six Little Dragons" to IPO. Notable for 15-year survival across multiple Chinese tech cycles; pivoting to spatial intelligence as foundational AI infrastructure.

"We want to become the glasses for large language models—able to understand our physical world." [01:06:54]

NVIDIA Description: Global GPU and AI computing company; Huang's former employer. Why Mentioned: Cited as the primary organizational and strategic inspiration. Huang draws direct parallels between NVIDIA's decade-long CUDA ecosystem buildout (which he worked on) and Qunhe's long-term spatial intelligence bet. Also noted for promoting synthetic data in robotics.

"NVIDIA also sponsored many people to use CUDA for general computing research during its hardest times. That whole process took nearly ten years. A company's transformation is a very long process—part luck, part endurance." [01:12:17]

DeepSeek Description: Hangzhou-based AI lab, one of the "Six Little Dragons." Why Mentioned: Referenced as an example of a company that has achieved genuine world-level influence—the benchmark Huang holds for what he aspires to build.

"Compared to DeepSeek, [our past work] certainly doesn't have that kind of influence." [01:34:34] "These companies, including DeepSeek, have been doing this for many years." [01:37:00]

Li Fei-Fei's World Labs (implicitly referenced as "Marble") Description: Spatial intelligence startup founded by AI pioneer Li Fei-Fei. Why Mentioned: Cited as aligned with Qunhe's technical bet on 3D spatial generation over video generation—validation that this technical path is taken seriously by top global AI researchers.

"Li Fei-Fei's Marble—our thinking is very much about generating 3D content." [00:55:12]

SpaceX Description: Elon Musk's aerospace company. Why Mentioned: Huang's go-to example of a company with genuine world-level impact—his personal benchmark for what constitutes a truly meaningful enterprise.

"I think something like SpaceX has real world-level influence." [01:33:34]

4. People Identified

Huang Xiaohuang (黄晓煌) Description: Co-founder and Chairman of Qunhe Technology. Former NVIDIA CUDA engineer and UIUC PhD student. Founded Qunhe in 2011–12. Why Mentioned: Subject of interview. Notable for extreme long-termism, technical depth (personally reads AI papers, writes code, designs interview questions that AI cannot solve), emotional stability under pressure, and a consistent 15-year bet on physical-world computing.

"I'm currently functioning more as a research leader—connecting scientific research outcomes with products. That's my main work now." [01:21:27]

Jensen Huang / 黄仁勋 Description: CEO and co-founder of NVIDIA; Huang Xiaohuang's former boss. Why Mentioned: Cited as an example of a founder who maintained conviction in an unpopular technical direction (CUDA/general computing) for a decade before it paid off. Direct parallel to Qunhe's own trajectory.

"At meetings, he was very persistent—saying 'the next S-curve growth.' Nobody believed him. He had a very strong belief in the opportunities he saw, and he persisted and was willing to invest long-term." [01:35:31]

Cheng Hang (程航) Description: Entrepreneur who competed against a major tech giant (unnamed), referenced in context of surviving big tech competition. Why Mentioned: Cited as someone who "survived competing against a major internet company" and calls that experience a "coming-of-age ceremony" for startups. Story illustrates how big tech eventually abandons ventures when business heads get fired for lack of progress.

"Cheng Hang told me: 'Only those who have fought against a major tech company are truly unicorns.' We all treat competing against big tech as a coming-of-age ceremony." [00:22:34]

Wang Huaisu (王怀所效用 — appears to be Wang Huaisu) Description: First angel investor in Qunhe. Why Mentioned: Invested RMB 500,000 personally as an angel when all other partners at his fund rejected the deal—solely because of a long-term personal relationship with the founder. Represents the importance of relationship capital in early-stage tech.

"Other partners all rejected it. He thought, we've known each other for so many years, it would be embarrassing not to. So he invested about RMB 500,000." [00:06:52]

Mao Chenyu (毛成语 — likely IDG's Mao Shengbo) Description: Partner at IDG Capital; first institutional investor in Qunhe. Why Mentioned: Made the bet on Qunhe based on conviction in the home decoration market (via connections to Qijia.com founder), not on GPU technology or O2O narrative. Represents backing market thesis over technology story.

"He was very bullish on the home decoration industry at the time—he knew the founder of Qijia.com well, and saw the industry exploding in growth." [00:09:21]

5. Operating Insights

Designing Interview Questions That AI Cannot Solve

As talent competition intensifies, Huang developed a differentiated hiring filter: questions that take ~1 hour, that AI cannot answer in 5–10 minutes, specifically designed to test learning speed and hands-on ability (e.g., reproduce and improve a research paper). This bypasses résumé screening and finds people who can actually build, not just credential-match.

"I write my own questions now—ensuring AI can't solve them. Not something done in five or ten minutes, probably about an hour. The main thing being tested is the interviewee's learning ability and hands-on ability. I don't look at their résumé." [00:48:25] "Learning ability is expressed through hands-on ability. For example, if you read a paper, did you actually understand it? You need to get your hands dirty and implement it—reproduce it, even optimize it." [00:48:55]

Use External Reputation Events to Solve Internal Alignment Problems

When Qunhe became one of Hangzhou's "Six Little Dragons," Huang deliberately used the external credibility boost to break internal resistance to long-term R&D investments that had previously seemed unjustifiable to skeptical employees and shareholders—without changing his actual strategy at all.

"Before the Six Little Dragons label, if I pushed for training our own spatial intelligence models, many people would object. After the Six Little Dragons, those voices became much fewer. Cohesion became much stronger." [01:17:35] "Even some shareholders who had been criticizing me—saying 'cut R&D, make more profit, stop burning money on models'—after that, they went quiet." [01:19:01]

Separate Strategic Decision-Making Circles by Expertise Domain

Huang identifies that he is highly susceptible to being swayed by people around him—and turns this self-awareness into a deliberate organizational design: when making technical bets, he only includes technical experts in the decision room. Including sales or business people in strategy discussions about deep tech produces systematically wrong outputs.

"If you're discussing spatial intelligence but you've pulled in a bunch of salespeople to discuss it, they'll definitely tell you 'launch tomorrow, pivot the day after.' But when we discuss spatial intelligence, we pull in the Chief Scientist and people like that to discuss strategy." [01:25:19]

Build Redundancy Into Business Lines Before You Need It

Huang pre-built two independent survival businesses (international + industrial manufacturing software) as insurance against losing their core home design SaaS to big tech competition—before the threat materialized. This optionality saved the company.

"We made a few assumptions: one, if all our online business is completely lost, we still have the manufacturing business. Another: if all China business is completely lost, we still have overseas business. I think in any situation you must do bottom-line thinking." [00:20:08]

6. Overlooked Insights

The GPU Cost Structure Problem Predicted the Token Economy—And Nobody Noticed

This was mentioned briefly and technically but is enormously significant. Huang observed in ~2015 that GPU-based SaaS fundamentally cannot follow the subscription model economics of CPU-based SaaS (where marginal cost approaches zero with scale). GPU cost is roughly fixed per computation regardless of user count. This means GPU-native products must eventually be priced per computation—exactly what happened with OpenAI's token pricing. He connected this structural observation directly to why AI companies today all use consumption-based pricing, and why it vindicates his original architecture. This is a deep, non-obvious insight about AI business model structure that most people have not traced back to hardware economics.

"The SaaS model had an inherent defect. GPU is different from CPU. The more people use CPU, the lower the per-unit cost. With GPU, no matter how many people use it, the cost is roughly the same. So you can only sacrifice quality to lower costs." [00:11:17] "In the CPU era, products were all annual/monthly subscriptions—use a lot or a little, the cost is about the same. But in the GPU era, everything is per-token, pay-per-use—like ChatGPT. Each computation is very expensive. If you let users use it unlimitedly, it just doesn't work." [00:12:16] "After this wave of AI, once everyone accepted token/consumption-based pricing—for us, it was like a shackle being unlocked." [00:13:07]

Domestic vs. International Robots Have Fundamentally Different Data Religion—and It's an Addressable Market Wedge

Huang briefly mentions that Chinese robotics companies overwhelmingly prefer real-world (真机) data for training while Western/overseas companies prefer synthetic data—and he traces this to the autonomous driving lineage of Chinese robotics teams (who inherited real-scan data practices from AV development). This is a rarely-discussed structural difference with major implications: it means a spatial data company needs a bifurcated product strategy by geography, and the winner in each market may be different. Qunhe is uniquely positioned to serve both because they have deep capabilities in both real-world spatial reconstruction AND synthetic generation.

"We found that typical Western companies prefer synthetic data. Domestic Chinese companies almost universally prefer real-world data. I think this is because many Chinese robotics companies come from the autonomous driving industry, which naturally used real vehicle scan data—so there's an inertia." [01:00:33] "You can't change the belief system of a technical team. If you believe in real data, I'll get you real data. If you believe in synthetic data, I'll get you synthetic data." [01:01:01]