黄晓煌

群核’s hardest asset is an accidental goldmine of physical-world data: InteriorNet, an open-source dataset released in 2018, was discovered by Silicon Valley giants after the pandemic shut down offline data collection and has since made the company a major data provider to embodied-AI “brain” companies and multimodal-model players. 黄晓煌 admits, “I had an instinct that it would be useful, but at the time I didn’t know what it was useful for”; inspired by ImageNet, the team open-sourced it anyway. The clearest sign of commercial validation is whether customers keep buying more data, and repeat purchases are “pretty decent.”

The core strategic logic was choosing a lane under China’s compute constraints: large language models showed that more data means more intelligence, but with compute restricted in China, 群核 moved into physical AI, where data is less constrained, data scarcity is greater, and compute requirements are less extreme. New models must also improve the legacy business—using large models to infer the physical parameters of objects in images and replace manual work. “The old business is our cash cow,” and that is the line against a second-startup-style transformation.

The company deliberately avoids red oceans: “I’ve been hit by giants myself—I’ve been running from them ever since. Why would you go looking for a market packed with giants? That’s suicide.” The major tech companies’ physical-AI efforts are still “small teams exploring,” unlike the company-wide push behind large language models. 群核 is entering through 3D data, combining 3D and video models as a complement to video foundation models; 黄晓煌 says he has not seriously studied Meta’s V-JEPA 2, while 群核’s stated edge is spatial consistency.

AI is rebuilding the business model: after seeing OpenAI charge by the token in 2023, 群核’s first move was to revise its annual and monthly SaaS plans, because under the old model “every time someone used it, I lost money,” while compute-heavy features such as video generation were blocked by the CFO. SaaS value is also being measured in reverse—from “users multiplied by the time they spend each day” to “the less time I occupy, the more value I create.”

The company’s closest brush with disaster came in the first half of 2021, when “everything was exceeding expectations,” hiring surged toward 3,000 people, and the combination of property, the pandemic, and capital markets suddenly reversed course by year-end. “If I had been even more reckless then, I probably would have been finished.” Since then, the hard rule has been break-even at the floor and never exhaust the ammunition: even while highly bullish on embodied AI, “I still won’t burn money recklessly,” and when investors push for faster growth, his answer is, “We’ll grow steadily. I’m not burning.”

Organizationally, 群核 separates a “process army” from an “innovation army,” with the latter accounting for roughly one-tenth of the company; the test is not business maturity but whether “you can’t even formulate the KPI, in which case you need an innovation team.” His view of shareholders is equally blunt: “A shareholder just wants your money, not your life.” For shareholders who do not share the company’s direction, “don’t drag it out—find a way, whatever the cost, to buy them out and get them out as fast as possible; any price today is still the smallest price compared with the future.”

The Six Little Dragons effect has materially changed the talent pipeline: resumes from C9 universities are up about 9x this year, overseas-returnee resumes are 20x last year’s level, and after the property downturn many people “barely dared to ask for the salary you were offering,” leaving projects indefinitely delayed. 黄晓煌 has made recruiting his top priority. His interview process is to hand candidates a paper and an open-ended problem; some work from 9 a.m. until 1 or 2 a.m., and “I stay with them.”

The operating philosophy has stayed consistent: “take a hammer and look for nails,” while commercial timing follows academia—public-facing hot topics are “2 or 3 years, even 3 or 4 years behind,” with robotics hot in academia in 2022–23 but pushed by industry only in 2024–25. “If scholars haven’t solved it, don’t go messing around with it yourself.” 群核’s biggest success case is when technology works across industries and then catches 1 or 2 industries in a breakout phase; for now, robotics is the sector he likes best.

🔗 Original source & video: 70. The First Listed Company Among Hangzhou’s Six Little Dragons: An Interview with 群核’s 黄晓煌 on 15 Years of Hard-Tech Evolution

Qunhe’s IPO is not the story of calling one hot trend right over 15 years, but of repeatedly moving the same technology into new scenarios and business models—and surviving every cycle. 黄晓煌 started by putting GPUs in the cloud in 2011, then moved through O2O, SaaS, industrial software and overseas expansion before turning to spatial intelligence; his underlying view is that Chinese sectors can reverse course every 3 to 5 years, so “staying alive matters more than looking impressive.”

Early SaaS charged monthly or annual subscriptions on the surface, while the underlying GPU incurred real per-use costs—giving Qunhe an early taste of the token economics AI faces today. A single operation could consume roughly RMB1 of compute, while advertisers would pay only a few cents for an impression; even tens of millions of yuan in ad revenue still translated into losses. Qunhe limited image sizes and pushed heavy workloads into the night, until usage-based pricing finally removed what 黄晓煌 called “a shackle that was almost impossible to unlock.”

Competition from Big Tech forced 黄晓煌 to rewrite his view that core algorithms alone constituted a moat, replacing it with a combination of algorithms, data, network effects and customer service. Almost the entire Qunhe rendering-engine team, apart from its manager, was once poached, and the technology was copied, leading him to conclude that pure algorithms were “a layer of window paper.” But after underinvesting in algorithms, he corrected himself again: “The algorithm is the engine, data is the fuel—you need both.”

Since 2023, Qunhe has been using the revenue and organization of its legacy SaaS business to build a new foundation centered on spatial models. Its algorithms and model team grew from fewer than 10 people in 2023 to roughly 60, with annual compute spending in the tens of millions of yuan. At the same time, AI is compressing traditional front-end and back-end costs, while pricing shifts from annual and monthly packages toward tokens, rendering volume and compute volume—drawing resistance from customers, business teams and shareholders alike.

The key question for spatial intelligence is not the heat around robotics, but whether 3D reconstruction and generation can become the general-purpose layer through which large models understand the physical world. Qunhe is choosing 3D because it is measurable, interactive and capable of representing distance and occlusion, while the video route represented by Google, Veo 3 and Seedance 2.0 is more flexible and visually polished. 黄晓煌 acknowledges 2 core uncertainties: “How good can the model become,” and whether video will ultimately replace 3D.

New business already accounts for roughly 5% of revenue through incremental compute purchases, but 黄晓煌’s commercialization target remains deliberately modest: within 3 years, compute-based revenue should reach a 1:1 ratio with monthly and annual SaaS revenue. The robot “brain” still resembles autonomous driving 10 years ago: the direction is clear, but the timing of deployment is not. Qunhe is therefore selling the same capability into embodied intelligence, factory planning, AGV, film and television, and e-commerce, because a “dragon-slaying hammer” cannot be aimed at only a handful of tiny nails. The company is past the pure cash-burning phase: “You still need profits.”

The Six Little Dragons’ most tangible contribution to Qunhe is not immediate revenue, but simultaneous improvement in talent supply and organizational permission. Resumes received in 2025 were roughly 9x the 2024 level, with overseas resumes up about 20x; 黄晓煌 nearly stepped away from operating the business to focus on recruiting for the entire year. Shareholders and colleagues who had questioned whether Qunhe should train its own models also became less resistant, while his interim mission is for spatial intelligence to contribute half of company revenue.

🔗 Original source & video: 160: 15 Years with 黄晓煌 After Qunhe’s IPO: The GPU Everyone Dismissed, Champion Kujiale, Spatial Intelligence, and the Six Little Dragons

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.”

🔗 Original source & video: Vol.172 Interview with Qunhe Technology’s Huang Xiaohuang: 14 Years as a Technical Entrepreneur