Former NVIDIA VP Takes On Jensen Huang: A New Challenge in AI and Robotics

After leaving NVIDIA for 15 years, he aims to challenge Jensen Huang.

Published on February 11, 2026 by China Entrepreneur Magazine

The most crucial factors in determining the success or failure of AI are talent and organization.

Zhang Jianzhong, the founder of Moore Threads, exemplifies the trend of former NVIDIA executives returning to China to start their own businesses and compete directly with their old boss, Jensen Huang. Previously, Zhang served as NVIDIA’s Vice President and General Manager for Greater China before founding Moore Threads in 2020, which has been dubbed the “Chinese version of NVIDIA.” By December 2025, Moore Threads made its debut on the Sci-Tech Innovation Board, reaching a market valuation of over 300 billion yuan.

Another notable figure is Huang Xiaohuang, the founder of Qunke Technology. While pursuing his master’s degree in the United States, he joined NVIDIA as a software engineer and contributed to CUDA development. In 2011, he returned to Hangzhou and co-founded Qunke Technology with classmates. The company’s flagship product, “Cool Home,” has become China’s largest spatial design platform. However, the journey towards Qunke Technology’s IPO has been more complicated. In April 2021, Qunke applied for a listing on NASDAQ, alongside over 30 other Chinese companies. By the second half of the year, the landscape shifted dramatically, leading Qunke to withdraw its application. By February 2025, Qunke submitted a listing application to the Hong Kong Stock Exchange, with J.P. Morgan and CCB International serving as joint sponsors, coinciding with the rise of the “Hangzhou Six Little Dragons.” Unfortunately, the application was not approved within six months, causing the prospectus to become invalid. On August 22, 2025, Qunke applied for a listing in Hong Kong again and is currently still awaiting approval.

During his time at IDG Capital, Mao Chengyu, the founding managing partner of Yunqi Capital, led Qunke’s first institutional investment and continued to provide additional funding. After the popularity of the “Hangzhou Six Little Dragons,” he expressed to Huang Xiaohuang, “Luck and fate are quite mysterious. If Qunke had gone public back in 2021, the ‘Hangzhou Six Little Dragons’ (all non-listed companies) wouldn’t have existed.” Huang Xiaohuang reflected that for the past 15 years of entrepreneurship, he has been wielding the “hammer of GPU cloud computing” while searching for the right opportunities. Initially, he targeted the home furnishing industry, launching the popular product “Cool Home.” As the real estate market declined, Huang shifted focus to physical AI applications like robot simulation training and industrial simulation, directly entering NVIDIA’s territory.

By the end of 2025, Qunke further advanced its spatial intelligence open platform, Aholo, continuously enhancing its foundational capabilities in spatial reconstruction, generation, editing, and understanding.

At the 2026 CES conference, Jensen Huang predicted that the moment for applying physical AI in autonomous driving and robotics is approaching rapidly. NVIDIA updated its digital twin simulation platform, Omniverse, and the world model, Cosmos, while also launching proprietary robot models, Groot and Alpamayo. Huang Xiaohuang also made predictions about physical AI, stating, “In three to five years, the breakthrough point for spatial intelligence will arrive. This depends not only on algorithms and data but also on computing power and hardware. A clever individual might suddenly solve a key algorithm, overcoming the bottleneck.” In November 2024, Qunke officially launched the spatial intelligence training platform, SpatialVerse, which provides synthetic training data for robots. In March 2025, Qunke released and open-sourced the spatial language model, SpatialLM; in August, the open-sourced spatial generation model, SpatialGen; and in November, the industrial AI twin platform, SpatialTwin, was launched, serving as the “brain of the factory,” capable of simulating real industrial environments in real-time and supporting the large-scale operation of intelligent agents. Currently, leading robotics companies like Galaxy General and Zhiyuan have become users of Qunke’s technology.

Since its establishment in 2011, Qunke has successfully secured hundreds of millions of dollars in financing from various institutions, including IDG Capital, GGV Capital, Shunwei Capital, Matrix Partners, Hillhouse Capital, and Yunqi Capital. According to the prospectus, in the first half of 2025, Qunke’s revenue reached 399 million yuan, marking a year-on-year increase of 9.4%, with corporate client subscription revenue being the primary contributor. The company has achieved profitability, with an adjusted net profit of 17.825 million yuan. Thanks to its subscription-based SaaS model, Qunke maintained a gross margin of over 80% from 2024 to the first half of 2025.

Unlike NVIDIA’s expansive innovation approach, Qunke’s strategy involves small steps and learning from mistakes. In 2018, Qunke open-sourced the InteriorNet dataset, which was the largest indoor spatial cognition deep learning dataset globally at that time, attracting interest from major companies in Silicon Valley. However, as Qunke attempted to explore larger spatial models, it encountered bottlenecks. Although the Transformer architecture was introduced by Google’s team in 2017, it had not yet become mainstream technology. Qunke had long aimed to convert spatial data into scripts, but the output was challenging. Huang Xiaohuang mentioned, “At that time, we had a small team of 5 to 10 people working purely out of interest. It was like enrolling a child in a hobby class without seeking commercial returns.” Eventually, the technical challenges that troubled Qunke were resolved only with the rise of open-source large language models. Nevertheless, the path of creating world and spatial models remains relatively “unorthodox.” Outside of Jensen Huang and Fei-Fei Li, few have fully committed to this area. The data engineering requirements for world models are significantly higher, with greater technical barriers, potentially taking 5 to 10 years or longer to yield returns, while capital is often focused on immediate concerns.

Huang Xiaohuang’s entrepreneurial partners also share a similar commitment to their mission. He, along with Qunke CEO Chen Hang and CTO Zhu Hao, have supported each other for 15 years in their entrepreneurial journey. All three have backgrounds in computer science, with Huang Xiaohuang and Chen Hang both hailing from Zhejiang University and Zhu Hao from Tsinghua University. The three of them also studied together as master’s students at UIUC (University of Illinois Urbana-Champaign).

Mao Chengyu once remarked during a conversation with Huang Xiaohuang, “I find it curious that the three of you have managed to work together without major conflicts, which is rare among team entrepreneurs.” Huang Xiaohuang replied, “Perhaps it’s because our backgrounds and judgments are quite similar, leaving little room for significant disagreement.” Mao Chengyu cautioned, “You should also prevent the emergence of cognitive blind spots. While it’s beneficial that you three think alike, it is also essential to have external perspectives to trigger new ideas.” In earlier years, Huang Xiaohuang did bring in external executives, but they discovered that external input was beneficial during stable business phases but could serve as a significant obstacle during transformations. “Any forward-looking decision must be made by a small group, not through discussions involving ten people,” Huang Xiaohuang emphasized.

Huang Xiaohuang believes that the most critical factors in determining the success of AI are talent and organization. Since 2023, companies like ByteDance, Alibaba, and Tencent have been competing globally for top talent, making it challenging for startups. However, ultimately, “financial capability” does not determine everything, as some large model startups have risen in the shadows of larger corporations.

In 2025, Huang Xiaohuang participated in two interviews with China Entrepreneur, following the release of the spatial generation model, SpatialGen, and the industrial digital twin platform, SpatialTwin. Currently, Qunke’s status resembles NVIDIA in 2006 when Jensen Huang was promoting CUDA and offering it for free to university professors and the meteorology industry.

At the GTC conference in March 2025, Huang Xiaohuang met Jensen Huang, where they exchanged pleasantries and took photos together. The public often jokes about Huang Xiaohuang’s departure from NVIDIA in 2011 and his decision to sell his stocks. In response, Huang Xiaohuang posted on social media, stating, “From GPU high-performance computing to today’s embodied intelligent training, the influence of NVIDIA and Jensen Huang on my entrepreneurial journey is invaluable. Discussing money diminishes the essence.” Now, while Huang Xiaohuang continues to open-source spatial language datasets and models, he is also providing tools and foundational capabilities for various scenarios, including factories, robotics, and e-commerce. He awaits a breakout moment akin to what ChatGPT brought to NVIDIA.

Full Dialogue (edited):

Great companies start small, and entrepreneurship always presents opportunities.

China Entrepreneur: The product most familiar to the public is “Cool Home,” but now you’re shifting towards spatial intelligence, which feels like a second startup.

Huang Xiaohuang: This isn’t a second startup; we are upgrading our existing business system while continuously strengthening our original business. The large models we train are actively participating in our services. Our initial focus was on using GPU high-performance computing for 3D spatial design, which has allowed us to accumulate the currently scarce 3D data, laying the foundation for our move towards spatial intelligence. Everything is actually progressing naturally.

China Entrepreneur: You open-sourced the world’s largest indoor spatial cognition deep learning dataset, InteriorNet, back in 2018. Were there any controversies regarding this? Did the technical bottleneck only get resolved after the emergence of ChatGPT in 2022?

Huang Xiaohuang: There were definitely debates. We were quite excited in 2018. After open-sourcing, many partners approached us, and we even expanded our team. Later, we hit a technical bottleneck and had to downsize the team. At that time, training a large language model to solve the problem was impossible. Everyone was building layer by layer, but without anyone open-sourcing foundational large language models, it was hard to produce a complete code. Perhaps later, other models will use ours for training, which is the essence of open-sourcing.

China Entrepreneur: You mentioned you didn’t acquire enough GPUs back then.

Huang Xiaohuang: In 2021, I believed AI was on the rise because many papers were being published. We hired an AI professor from the U.S. and thought that having dozens of GPUs per person would suffice, with just a two or three-person team. Who would have known that now, you need thousands of GPUs? If you were in my position back then, would you have approved that? It would have been impossible.

China Entrepreneur: Given your background at NVIDIA, you must be very familiar with GPU physical rendering.

Huang Xiaohuang: Initially, we developed a physically accurate rendering simulation. At that time, we were also searching for the right opportunity and coincidentally found a booming home furnishing industry, leading us to create Cool Home.

China Entrepreneur: Now, that opportunity has expanded to include robotics.

Huang Xiaohuang: We have explored various areas, including VR and AR, but those industries quickly declined. When you’re looking for the right opportunity, you must find the fastest-growing sectors.

China Entrepreneur: The process of finding these opportunities also tests your abilities as a founder and leader.

Huang Xiaohuang: It requires foresight and keeping a pulse on the latest technology. Over the past 14 years of our entrepreneurial journey, we have seen various industries rise and fall. Initially, home furnishing industry leaders were thriving and making profits easily. When our product was not yet released, we created a demo, and everyone came to place orders, overwhelming our POS systems. However, no industry remains perpetually good. Just like the automotive industry, which was thriving two years ago, it is now entering a phase of stock competition. Yet the pursuit of technology and efficiency by people is everlasting, so it’s essential to follow trends; going against the tide can be exhausting.

China Entrepreneur: Currently, everyone is developing language models, while you are focusing on spatial intelligence. Does it feel like you are going against the grain?

Huang Xiaohuang: When you are going against the trend, you need to minimize your team and focus on core technical explorations and accumulations. When opportunities arise, you can then expand. I believe in maintaining curiosity and faith in technology, but without being impulsive. When we ventured into spatial intelligence in 2018, there were several instances where we felt opportunities were coming, only to later realize we were mistaken. This time feels like a significant opportunity, with the core factor being whether embodied intelligence can truly take root. Personally, I believe it can. This wave consists of several overlapping trends, including large models and embodied intelligence, presenting vast opportunities. Thus, we are increasing our investment.

China Entrepreneur: Therefore, whether it’s models or tools, it’s crucial to establish an iterative rhythm?

Huang Xiaohuang: In the ten years leading up to 2022, when we began working on Cool Home, I felt there had been little significant technological change. I once thought that entrepreneurship was about competing in business and sales capabilities, as everyone’s offerings were relatively similar. While I knew I could perform better than others, it didn’t lead to any significant differentiation. However, after 2022, there has been a clear distinction. Now, products are abundant and growing explosively. When others stack GPUs, the resulting capabilities exceed all previous algorithms. We also try using simple algorithms combined with multiple GPUs and vast amounts of data, which is much more powerful than painstakingly developing algorithms.

China Entrepreneur: The rapid iteration of technology can also lead to failure; what are you particularly cautious about?

Huang Xiaohuang: Entrepreneurship is inherently fraught with risks. I have explored various new businesses and modules, some succeeded, while others did not. Human cognition is limited, and most attempts are likely to fail. The core principle is to quickly identify failures and minimize the associated losses. Startups have limited resources, so founders must learn and adapt quickly. Startups can act more swiftly than larger companies. As long as you remain agile and adjust quickly, I believe there are always opportunities for growth. Today, all great companies started small; there are always opportunities.

China Entrepreneur: How do you convince robotics companies to use your models and datasets for training? Galaxy General and Zhiyuan are already your clients, but some companies may worry about data security.

Huang Xiaohuang: We provide contract data unidirectionally, so there is no risk in that regard.

China Entrepreneur: What does this process look like? How do robotics companies customize training data with you?

Huang Xiaohuang: Each company is unique. Our engine can quickly configure synthetic data to meet different needs. Currently, the quality of physical and manually collected data is high, but it is slow and expensive. Synthetic data is faster to produce and can be generated in large quantities.

China Entrepreneur: Is the entire process from spatial to data, and then to robotics usage seamless? Are there any difficulties?

Huang Xiaohuang: If you conduct interactive training for robots, you must verify the results in the physical world. You will discover many anomalies due to missing data. For instance, if we have a client with a cat and dog mess in their room, you must supplement and re-measure the physical parameters. This process can be painful; colleagues often spend time in front of screens, reviewing numerous images and videos to resolve edge cases.

China Entrepreneur: In November 2025, you launched the industrial AI twin platform, SpatialTwin, which combines data, models, and tools. How did your internal deliberations lead to this decision?

Huang Xiaohuang: While advancing the old business (Cool Home), we discovered that providing just raw data is manageable for only a few companies. Therefore, we aim to offer a more comprehensive toolchain to help various companies address these challenges. Tools, data, and models are all indispensable, creating a data flywheel. Moving forward, we will focus on strengthening these three aspects to build a closed-loop of “data—model—application capabilities.”

Qunke Technology’s data flywheel.

China Entrepreneur: The CPU era has long discussed twin factories. What changes have occurred in digital twin factories during the GPU era?

Huang Xiaohuang: Previously, it was the automation era, but the future will transition to the intelligence era. The difference is that when a robot drops a cup, it can easily determine whether to pick it up and continue. However, for an automated device, if something drops, it will still follow the original program and perform actions as programmed. There are many challenges that require a series of toolchain systems to solve.

China Entrepreneur: When you initially chose new business areas, such as industrial scenarios, why did you not develop an Agent? That seems to be more popular currently.

Huang Xiaohuang: We believe Agents are too “thin.” If you are creating a consumer product, an Agent could suffice. However, for serious industrial applications, it requires a richer set of capabilities.

China Entrepreneur: What pressing issues still need to be resolved with the industrial AI twin platform?

Huang Xiaohuang: We believe that to enable robots to genuinely perform tasks, three issues need resolution: first, spatial cognition; second, spatial reasoning; and third, action decision-making. We previously open-sourced the spatial cognition model. However, that isn’t enough; you also need to address reasoning and action control, which SpatialTwin addresses.

China Entrepreneur: How many resources have you internally allocated for this?

Huang Xiaohuang: About twenty to thirty people have worked on it for over a year.

China Entrepreneur: SpatialTwin seems quite similar to NVIDIA’s Omniverse, both being factory simulation systems. Are you in competition with one another?

Huang Xiaohuang: We can be compatible with Omniverse. Omniverse is a local simulation system, while we focus more on providing data and simulating interactions between data.

China Entrepreneur: What role does Qunke aspire to play in the factory transformation?

Huang Xiaohuang: Similar to a “water vendor.” We don’t produce these devices ourselves; instead, we research what is required for devices to become intelligent and which aspects we need to address. Currently, the easiest transformation is in factory planning. Transitioning from a traditional automated factory to an intelligent factory begins with planning, followed by training intelligent agents. You need to envision each machine as an Agent running on a twin platform, which creates a mirror of the physical world.

Qunke Technology’s SpatialTwin digital twin imagery.

China Entrepreneur: At what stage does spatial intelligence presently stand compared to the capabilities of ChatGPT?

Huang Xiaohuang: It is roughly at the stage of GPT 2.0 (2019).

China Entrepreneur: Do you consider this progress fast or slow?

Huang Xiaohuang: Quite fast. Many view AI development as a gradual ascent; in reality, it resembles a series of jumps. Someone suddenly discovers a new algorithm, causing AI to leap forward. The development of AI algorithms relies on brilliant individuals having sudden insights. The gap between GPT 2.0 (2019) and GPT 3.0 (2020) wasn’t long.

China Entrepreneur: What factors could accelerate this progress?

Huang Xiaohuang: Algorithms, computing power, data, and hardware all have the potential to drive progress. In 2025, hardware iteration is rapid, and everyone is seeking to address algorithm, computing power, and data issues. These elements are interdependent. The amount of data required depends on your algorithm, while the type of computing power dictates what can be achieved with the algorithm; everyone is striving to optimize this.

China Entrepreneur: Is there a scaling law in spatial intelligence, similar to that in language models?

Huang Xiaohuang: Current large model frameworks do exhibit scaling laws. However, spatial data is relatively challenging to obtain, and the associated costs are high, making it difficult. For instance, when training our own large model, SpatialLM, to recognize blueprints, labeling those blueprints is quite labor-intensive. Labeling a single blueprint can cost around 100 yuan, and for one million blueprints, that amounts to 100 million yuan. If the data volume increases tenfold, the model’s performance may only double. Finding appropriate methodologies relies heavily on tools, making this a complex engineering challenge.

China Entrepreneur: Do investors support your pivot towards spatial intelligence?

Huang Xiaohuang: Many investors, like ordinary people, have their opinions. Most investors recognize the potential of spatial intelligence, while a minority may not agree, and those individuals may choose to withdraw.

China Entrepreneur: Your company currently has over 1,300 employees. How do you persuade everyone to embrace the idea of physical AI and spatial intelligence, which seems like something only a large company like NVIDIA could accomplish?

Huang Xiaohuang: Every week, we communicate with all employees, sharing our latest progress and demonstrating various technical demos. Every time we embrace change and new trends, some people believe, and some do not. You can only strive to convince as many people as possible to believe and recognize the vision. If some remain unconvinced, then parting ways may be necessary; ultimately, a company must find like-minded individuals to journey together. The toughest aspect of entrepreneurship isn’t the technology itself. Just like in investing, changing people’s minds is difficult. There are no guarantees that embodied intelligence will succeed, but if you believe in its potential, you will find joy in the process. As long as you keep the company afloat and profitable, that’s a significant success. Whether things succeed or fail involves various phases, timing, and luck. No one can predict this, but as long as you engage without regrets, you will be fine. The greatest fear in entrepreneurship is doing something solely for profit; if it fails, you may end up disillusioned and resentful.

China Entrepreneur: With large companies aggressively recruiting talent with high salaries, how do you ensure that the people you cultivate are not poached?

Huang Xiaohuang: We have experienced several cycles in the industry and have built a solid organizational culture to attract exceptional talent. We build our capabilities on the organization rather than relying on a few individuals. Today, even without me, the company continues to operate. We have been in business for over a decade, and our entire system is established, so the loss of a few individuals does not cause problems. What matters most to us is attracting the best talent from around the world. As long as our hiring rate exceeds our attrition rate, it is a positive outcome. Employee turnover is entirely normal.

China Entrepreneur: Early in 2025, you also launched the “Star Core Talent Program,” offering salaries as high as one million yuan annually.

Huang Xiaohuang: Some truly talented individuals care less about earning a few thousand yuan more and would prefer not to work at high-paying companies; otherwise, it would be impossible to compete with large corporations. When recognizing talent, differentiation is essential. Just because large companies recognize someone as talented doesn’t mean they are strong; many excellent talents struggle to thrive within large organizations. In our medium-sized firm, we can work closely with them daily, and they can reach out to me directly. In a large company, you cannot directly contact someone like Jack Ma to report issues. Like Meta, which spends a billion dollars to recruit a single person, that’s beyond our means. However, the work environment and culture are more important than any individual. We strive to provide better benefits to outstanding talents within our capabilities. Competing purely on salary won’t solve problems; in the AI era, it’s not just about offering enough money to accomplish everything. As a founder, if I do not engage in AI research and read papers, I cannot make reasonable decisions. Moreover, the best algorithms may not necessarily come from the highest-paid individuals; it’s a nuanced matter. This is why many companies are inclined to hire fresh graduates; they may not be stronger than experienced individuals, but many experienced professionals are already nurtured by companies like ByteDance and Alibaba, making it challenging to recruit them.

China Entrepreneur: Therefore, it’s about having faith in young talent.

Huang Xiaohuang: You have to have faith. There are also top-notch talents among them, sometimes even outperforming those with experience. Recently, we interviewed a recent graduate from MIT, and he performed better than a senior executive with over ten years of experience at a major company.

China Entrepreneur: Has the rise of the “Hangzhou Six Little Dragons” helped your talent recruitment?

Huang Xiaohuang: Attracting talent is always beneficial. We usually prefer to stay low-key, focusing on making contributions that are worthy of our reputation.

China Entrepreneur: Why did you choose to open-source your spatial models?

Huang Xiaohuang: One of the reasons for open-sourcing is to attract talent. After open-sourcing models, many talented individuals will come across your models or try them during their research, leading them to join your company. Once you open-source, competitors cannot threaten you with your models, as only a few truly compete with you; if they were to use your models, they would still fall short of your capabilities. Overall, the benefits outweigh the drawbacks.

China Entrepreneur: What is your most important task right now?

Huang Xiaohuang: Leading the company into the next phase. We have experienced several cycles, and every time a new cycle begins, it can be tumultuous. If we mishandle it, we risk complete failure. The decision to either upgrade, re-enter the entrepreneurial space, or stick with the existing track is a very careful and challenging one.