黄晓煌 on 15 Years to Qunhe's IPO: GPUs, Kujiale, Spatial Intelligence
Summary
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.
Deep dive
1. When nobody cared about CUDA, he bet on general-purpose computing—then made putting GPUs in the cloud his startup wedge
During his PhD at UIUC, 黄晓煌 entered startup competitions and built a device to prevent drivers from falling asleep. After joining Nvidia’s CUDA development effort in 2009-2010, he saw internal applications accelerate CPU workloads by dozens or even hundreds of times and concluded that the technology would “change the world”—although almost nobody was using it.
Nvidia was losing roughly $300M a year, its GPU business was seen as having hit a ceiling, and the shift toward general-purpose computing was not going smoothly. 黄晓煌 proposed putting GPUs in the cloud, but the team worried that letting 10 people share a cloud card would cannibalize hardware revenue. He later admitted that he was junior at the time; the team thought the idea was “not very credible,” and the project stalled.
The prototype was first built in the US, where the team also tried to raise money, but the lack of a green card created practical obstacles. Around 2011, Silicon Valley was obsessed with cloud computing while GPUs had barely moved into the cloud, so he returned to China to start a company around the simple idea of “putting GPUs in the cloud.”
2. The first check did not come from calling the GPU—it came from an acquaintance giving the technical team room to experiment
Early investors were unimpressed by the 3 founders’ resumes. One co-founder had a respectable background at Amazon, but 陈航 had only an internship, while Nvidia was viewed by some as little more than a traditional manufacturer. One prominent fund even suggested that they work at Google or Facebook for 1 or 2 years before coming back to start a company.
The bigger mismatch was with the era: infrastructure R&D was slow and capital-intensive, while O2O dominated the market narrative. 黄晓煌 said that when they talked about GPUs to most investors, “80% of them didn’t even know what a GPU was.” The team eventually learned to tell an O2O story, while still using GPU rendering technology to search for applications.
After other partners declined, alumnus 王淮 invested roughly RMB500K as a personal angel investor. Before securing IDG funding, Qunhe took on a large volume of B2B projects and outsourcing work. IDG’s 毛丞宇 was more interested in the home-furnishing market and the possibility of eventually selling furniture.
3. Kujiale was not a planned wedge; users selected it
Qunhe initially turned the same technology into demos for film and television, home furnishing and architecture, simulation, and other fields: “Whichever industry was willing to pay, that was the one we would pursue.” 黄晓煌 saw this as a classic case of “taking a hammer to look for a nail,” similar to Nvidia’s approach of building one CUDA demo after another and using customers to validate demand.
Kujiale launched at the end of 2013, broke out in 2014 and began making money in 2015. The initial product targeted homeowners, but homeowners might churn within 6 months of use; the users who truly stayed were designers and renovation companies. The move toward professional users was not decided in a single strategy meeting—it emerged from retention data.
The team initially continued pitching homeowner traffic, O2O and furniture sales, and even cited a US company called House. But the furniture transaction never really got off the ground before the O2O narrative collapsed. 黄晓煌 already preferred selling software, so the company moved naturally toward designer subscriptions.
4. The GPU made subscription software structurally awkward
After the 2015 stock-market crash, it became harder for loss-making companies to raise money, so Qunhe moved quickly to SaaS pricing and reached cash-flow breakeven by the end of that year. The shift later became its main revenue engine, but also exposed a structural conflict between SaaS pricing and GPU costs.
黄晓煌’s explanation was that once a CPU product gains users, marginal costs can be spread down substantially. GPU compute, by contrast, is difficult to optimize by multiples; even the underlying architecture could deliver only improvements in the tens of percent. Qunhe had to cap rendering-image sizes and push high-compute jobs into the middle of the night, leaving users to wait until morning.
He therefore sees today’s widespread move to token-based and usage-based pricing as a change in the pricing model itself. Every ChatGPT computation carries a meaningful cost, so it cannot offer unlimited use in the way Google or Facebook historically did. Once the AI boom made token or usage pricing acceptable, “a shackle that was almost impossible to unlock” finally came loose.
5. Tens of millions in ad revenue still lost money because every interaction burned compute
Qunhe once had home-furnishing brands pay for exposure of cabinets, refrigerators and other products inside design renderings. Ad revenue reached tens of millions of yuan, but the business was ultimately cut. The problem was not a lack of advertisers: each operation in the design tool could consume roughly RMB1 of compute, while a single impression generated only a few cents.
The incremental cost of another display in web advertising is almost zero. A 3D-design ad, by contrast, requires the company to provide expensive compute for free before it can monetize exposure or a transaction. 黄晓煌 concluded that advertising, e-commerce and SaaS may be standard monetization models, but they do not fit every underlying cost structure.
Qunhe grew fastest from roughly 2015 to 2018. Its most profitable customers at the time were O2O platforms, some of which paid more than RMB1M a year for software and compute. The subsequent collapse of internet-finance platforms dragged the business down: 8 of its top 10 customers went bankrupt, and growth came under pressure.
6. Against Big Tech, Qunhe first planned for a scenario in which it lost everything
Around 2018, Big Tech entered the home-renovation market. The most direct damage was not to the product but to the organization: entire teams were poached in batches, some leaving overnight without handover. Customers could not even find anyone to continue serving them, and parts of the business were forced to shut down.
Qunhe did not respond with a price war in the same market. Instead, it moved the industrial-software business to the front: the project was launched in 2016, largely built by 2018 and deployed in 2020. 黄晓煌 viewed industrial software as long-cycle “dirty work” that required patience and had little to do with e-commerce—exactly the kind of work Big Tech was unlikely to stick with.
The strategy established 2 lines of defense. If the online business disappeared entirely, Qunhe would still have industrial 4.0 and flexible manufacturing; if the China business disappeared entirely, it would still have overseas operations. 黄晓煌 called this “bottom-line thinking under any circumstances.”
BAT companies had all expressed interest in investing, but Qunhe did not want to be pulled into an internet war. Baidu offered limited synergies, Alibaba might have pushed the company toward e-commerce, and taking Tencent money in Hangzhou could have alienated Alibaba. Investors became so alarmed that one directly suggested: “Why don’t you sell to Alibaba?”
7. Taking on Big Tech was a coming-of-age ceremony—and first shattered fake moats
Big Tech achieved short-term dominance in customers, capital and talent, but copying the product did not solve commercialization. After the rival business head spent 1 or 2 years without making meaningful progress and was eventually replaced, Qunhe concluded that the competitive episode was largely over.
黄晓煌 and 陈航 called the experience a unicorn’s “coming-of-age ceremony.” Only after competing directly can a company distinguish between areas where it has merely grown larger and those where it has grown stronger. “Any area that collapses as soon as competition arrives clearly has a problem.”
The most painful case involved the rendering-engine team: almost everyone apart from the manager was poached, and the product was copied “exactly.” 黄晓煌 concluded that treating pure algorithms as the sole moat in China was dangerous: “An algorithm is like a layer of window paper—one poke and it breaks.”
Qunhe responded by strengthening data, network effects and customer service, but 黄晓煌 then overcorrected and cut back on algorithm hiring. The breakthrough in large models forced another revision: “The algorithm is the engine, data is the fuel—you need both; you cannot be missing either one.”
8. Three- to five-year sector reversals made survival the priority
Early in Qunhe’s life, investors told the company to talk about O2O every day. 2 years later, the same people said: “Whatever you do, don’t mention O2O—not even a single O.” 黄晓煌 came to believe that Chinese sectors can undergo a complete reversal in reputation and capital preference every 3 to 5 years.
His conclusion was not to avoid hot industries forever, but that “staying alive matters more than looking impressive.” Inside Qunhe, steady execution is described as “fortifying and clearing the land”: once a business is secured, it must build accumulated advantage rather than be wiped out by a single controversy or counterfeit-goods incident.
黄晓煌 sees industrial software as a cyclical opportunity because real, hard demand does not disappear. O2O and the metaverse look more like concept-driven waves, although the metaverse could reemerge after a new breakthrough. AI replacing workflow-management SaaS, by contrast, is a productivity revolution rather than an ordinary cycle.
9. The 2020-2021 boom also made Qunhe admit it had gotten carried away
Funding was easy, SaaS sold well and property customers paid extraordinary sums. 黄晓煌 recalls that contracts worth several million yuan, even RMB10M-plus, were sometimes not yet signed when the money had already arrived. Qunhe responded by hiring aggressively and expanding into real estate, commercial fit-outs and industrial-4.0 production.
The problem was that much of the workflow management around the core 3D capability was not a true technical moat. 黄晓煌 estimates that roughly 80% of the people hired during the SaaS boom worked on these front-end, back-end and enterprise-process layers—exactly the roles and products generative AI made easier to replace.
Qunhe’s eventual adjustment framework was to first determine whether a change was cyclical or an irreversible trend. A cycle can be endured; a trend requires the top executive to drive a revolutionary overhaul, even if that means shutting down profitable businesses that do not belong to the future.
10. “The closer software is to the physical world, the safer it is” is Qunhe’s long-cycle thesis
黄晓煌 says the specific algorithm he worked on during Qunhe’s first day has already been overturned by multiple technology cycles, but the direction of connecting the digital and physical worlds can be measured in centuries. The physical world will not disappear, and people will not want to wear VR headsets forever, so Qunhe has continued searching for technologies that connect the 2.
His broader judgment is that software closer to the physical world is safer, while pure-virtual software that merely packages information workflows is easier for AI to replace. Qunhe’s progression from physically correct rendering to spatial understanding, reconstruction and robot training is, in his view, an upgrade in capability—not a complete change of track.
Around 2016, renovation loans and internet-finance revenue looked attractive. Related companies had teams roughly one-tenth the size of Qunhe’s but several times its revenue; employees also suggested using idle GPUs to mine cryptocurrency. 黄晓煌 rejected both ideas because they departed from the core technology and did not create the productivity value he cared about.
His project filter has 3 tests: use the core technology, create social value, and compound through data or network effects. Looking back at 2018, he thinks the bigger opportunity may have been “putting hardware around the algorithm.” Pure software remains difficult in China, but what form of hardware Qunhe might pursue remains an open question.
11. Technical exploration can defer market sizing; business expansion cannot
Qunhe first used the robot-vacuum industry to validate the synthesis of training data for spatial-model development. The market itself was too small to become a major business, but it was large enough to prove the technology. The company could then move the same “hammer” into embodied intelligence, factories and other larger markets.
黄晓煌 distinguishes between 2 types of exploration. Technical-product exploration should first validate capability and can temporarily set market size aside; pure market exploration must confirm the number of customers and revenue opportunity. Model and core-algorithm investment is enormous, so the eventual user base must still be broad enough.
Qunhe encourages teams to build demos from the bottom up, then lets a review committee decide whether to allocate resources, but the rules have tightened. If a product merely wraps Gemini or OpenAI without its own core algorithms or models, 黄晓煌’s instruction is explicit: “Don’t build that product.”
12. Qunhe saw the AI opportunity in 2018, then the real economy forced it to doubt itself
In 2018, the team began studying papers on using spatial images to identify walls, columns and other structures, and released an open-source dataset containing floor plans, appliances and furniture. In 2021, it brought in a professor from the US and established a lab dedicated to spatial models and large-model training.
But Big Tech’s poaching emptied out one algorithm team. One manager told 黄晓煌, “We can’t hire people as fast as we’re being poached.” Beyond GPUs, talent was the most important input into model development; Qunhe lacked people and had no way to know whether the breakthrough would come in 5 years, 10 years or 30 years.
The counterweight was the “money raining from the sky” in real estate and SaaS during 2020-2021. Qunhe had several hundred people working on property and construction-related businesses, while only a few remained on AI. Colleagues thought 黄晓煌 was using policy and headlines to steer the company—“running the company by reading Toutiao”—and he too began to wonder whether he had fallen into a technology echo chamber.
13. After the first IPO was derailed, real estate, COVID and internationalization turned in succession
Qunhe had planned to list on July 15, 2021. The Didi incident was followed by a property-market correction and the collapse of major customers, sending sentiment from a peak to the bottom almost immediately. 黄晓煌 described the experience as being “extremely dazed.”
In 2022, Qunhe brought employees from different countries to Shanghai to build an international team, only to see everyone in Shanghai leave. Before ChatGPT and Midjourney broke out, 黄晓煌 started each day by reading the news to find out “what happened this time.” The excitement returned only when AI took off in 2023.
14. Copilot convinced Qunhe that workflow SaaS was not in a cyclical downturn—it was being rewritten
After using Copilot to write code, 黄晓煌 concluded that workflow-based SaaS such as ERP and CRM had suffered a “disruptive hit.” Enterprise processes that once required an entire team to code could now be generated quickly through natural-language instructions. Qunhe therefore insisted on building an “engine,” rather than treating a UI shell as the core product.
At the beginning of 2023, the team judged that large language models and image models would be battlegrounds for Big Tech, where Qunhe had no chance to compete head-on. It therefore pursued 2 paths in parallel: training its own spatial model, and building adjacent capabilities such as LoRA and image processing around foundation models.
The trade-off became clear quickly. An in-house model is slow and expensive but can form a moat; adjacent businesses must move fast, because a new model can make months of prior work worthless and strong teams can easily jump to platforms such as Alibaba and ByteDance. Before the model matured, Qunhe used humans to patch synthetic data manually, with the human share declining as the model improved.
15. Strategy in 2024, talent replenished by the Six Little Dragons in 2025
By 2024, the route was clear: own the spatial models and algorithms, while building the supporting data-engineering and GPU infrastructure. Execution remained bottlenecked by talent and algorithms; 黄晓煌 says he lost “a lot of hair” reading papers.
While preparing for the Hong Kong listing, Qunhe used AI to reduce the cost of traditional front-end and back-end work and pushed SaaS pricing away from annual and monthly packages toward tokens, rendering volume and compute volume. Customer acceptance and the risk to existing business made the organization broadly reluctant to change, so the transition had to be forced through step by step.
After the Six Little Dragons went viral from late 2024 into early 2025, the company launched its “Little Dragon Plan.” The goal was not to monetize the label, but to use it to attract talent. In 2025, 黄晓煌 nearly stepped away from all operating work, personally reading papers to find candidates and designing coding tests that AI could not answer directly, with a focus on reproduction, optimization and learning ability.
Resumes rose roughly 9x from 2024, while overseas resumes increased about 20x. The 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. 黄晓煌 believes the talent Big Tech has misjudged is the main source of Qunhe’s differentiated recruiting.
16. Spatial intelligence is split into 3 layers: models, capability services and human-machine products
The foundation consists of 3 core capabilities: Qunhe’s existing physically correct GPU rendering; Spatial LM, which handles spatial understanding and reasoning; and Spatial Gen, which generates 3D spaces. The middle layer exposes APIs for spatial understanding, reasoning, reconstruction, editing and rendering.
The top layer serves humans through the Kujiale design tool, e-commerce studio photography and the short-drama production product LuxReal. It also serves machines through Spatial Twin for robotics and factory planning.
17. Qunhe is betting on 3D, not prettier video
黄晓煌 divides the competing approaches into 2 camps: Google, Veo 3 and Seedance 2.0 are oriented toward video generation, while 李飞飞’s Marble and Qunhe focus on 3D generation and reconstruction. He concedes that “we definitely can’t beat companies like Google at video,” leaving Qunhe to bet on a different technical path.
Video is polished and flexible. 3D content offers real dimensions, interactivity, better memory, and explicit distance, position and occlusion relationships. In a video, whether a person appears 1 meter or 5 meters tall may not disrupt continuity; for a system that must act in physical space, the difference is fundamental.
His view is clear: video models optimize for visual continuity, not physical correctness, so tasks that genuinely approach the physical world still require 3D. He retains 2 core uncertainties, however—how intelligent Qunhe’s model will ultimately become, and whether the video route will evolve enough to replace 3D.
18. Robotics is a certain future—but not yet a large enough customer pool
黄晓煌 compares today’s robotics industry with autonomous driving 10 years ago. The future is clearly imaginative, and he believes the embodied-intelligence brain will eventually be solved, but the timing of mass deployment is unknown. Batteries, range and compute are also problems Qunhe cannot solve on its own.
Spatial intelligence therefore cannot serve humanoid robots alone. Qunhe is also applying it to factory planning, AGV, conventional automated equipment, film and television, and cultural tourism. Even with real customers, each segment remains small. The host summarized the dilemma as having “a dragon-slaying hammer, but all the nails are tiny,” and 黄晓煌 agreed.
Incremental compute purchases currently account for roughly 5% of company revenue. 黄晓煌 wants compute-based product revenue to reach a 1:1 ratio with monthly and annual SaaS revenue within 3 years. Investment can be large and profit limited, but Qunhe is past the stage of pure cash burn: “You still need profits.”
19. China favors real-robot data; Europe and the US are more willing to trust synthetic data
From serving customers, Qunhe has observed that Chinese robotics companies generally prefer real-world, in-situ and real-robot data, while European and US companies are more willing to use simulation and synthetic data. 黄晓煌 attributes part of China’s preference to the technical habits of autonomous-driving teams, but stresses that “you cannot change a technical team’s beliefs.”
Qunhe therefore does both. Multi-angle images and video combined with LiDAR can reconstruct a space close to reality; friction coefficients, materials and interaction feedback cannot be replicated through vision alone, but can be collected separately and transferred across more virtual environments. For a robot, the ultimate input is still a digital signal.
Real-world scanning overseas also raises privacy and regulatory issues, limiting data usage. When the host asked whether Nvidia promotes synthetic data mainly to sell chips, 黄晓煌 rejected the premise, saying the more important factors were the US ecosystem and regulatory constraints.
20. Robot companies, 黄晓煌 believes, will vastly outnumber humanoid-robot companies
In his view, the current 100 or 200 related companies are nowhere near enough: “Until there are 10,000, it doesn’t count as many.” Intelligent robots will not all be humanoid. A crane that lifts shipping containers cannot be humanoid, and a miniature device that clears a blocked blood vessel is also a robot.
曼祺 countered that many machines are already adapted to existing production processes and may not need more intelligence, while intelligence adds compute, power and energy costs. 黄晓煌’s response was that every machine needs some degree of intelligence; the difference is whether it operates fully automatically, semi-automatically, or lets one person operate more machines at once.
On energy costs, he added that automation can also reduce the cost of producing and installing energy infrastructure. If solar power were automated from manufacturing through installation, the main remaining inputs would be minerals and land. “Efficiency improvements are never a negative.”
21. Spatial LM aims to add the physical world’s missing translation layer to LLMs
Qunhe does not intend to build an entire system for any single robot. Spatial LM focuses on translating the physical world into scripts that already powerful large language models can understand, reason over and use to interact with the physical environment. If the capability is general-purpose, it can be reused across different robots and all of Qunhe’s business lines.
黄晓煌 compares it with animals: creatures can move through space without running into walls—“apart from the occasional goofy husky.” The first step for spatial intelligence need not be complex labor; simply moving freely through different spaces already has value. A more mature system would be able to sweep the floor today and wipe a blackboard tomorrow.
He believes spatial intelligence will eventually reach the envisioned state, but does not know how long it will take. His reasoning is not based on short-term product data: human capabilities fundamentally consist of communication and activity in the physical world. Since biological systems can achieve those capabilities, machine systems should be able to achieve them in principle.
22. Platformization was not a personality transplant; foundation models changed the conditions for horizontal capabilities
Before large models, the industry playbook was to go deep in a single vertical. When Qunhe studied AI companies such as Megvii and SenseTime in 2017-2018, it mainly saw government projects, large enterprises, facial recognition and security applications. The company therefore also believed in vertical deployment rather than a technology platform.
The premise is different today. Large language models are already powerful, but they lack a shared layer for understanding the physical world. 黄晓煌 therefore describes Qunhe’s move from the Kujiale-style vertical product toward Spatial LM, APIs and multi-industry applications as “going with the tide,” not abandoning its original technology line.
His most familiar reference remains Nvidia: it moved from the niche gaming-GPU market into general-purpose computing and infrastructure, while the CUDA ecosystem took nearly 10 years to form. The lesson is that “a company’s transformation is a very long process, partly luck and partly endurance”—with cash flow and survival still the prerequisites.
23. The Six Little Dragons turned Qunhe’s tech-company identity from self-definition into organizational consensus
Qunhe listed in Hong Kong on April 17, becoming the first listed company among the host’s Six Little Dragons. 黄晓煌 says that from day 1 the company had worked on physically correct rendering and the digital-physical bridge, but outsiders, some employees and shareholders had long viewed it simply as a real-estate or home-furnishing company.
The Six Little Dragons label did not change his business judgment, but it materially reduced execution resistance. Shareholders had previously questioned whether Qunhe should train its own models—“Do you really know your own limits?”—and urged it to cut R&D and focus on profit. After the label emerged, similar objections declined, and business colleagues became more willing to accept investments that might take years to monetize.
Qunhe’s internal rule is: “Do not monetize the fame, but use it to recruit more people.” The cost is a sharp increase in visits and events. 黄晓煌 worries that these activities consume time needed for model and technology work, especially when a single training run can cost several million yuan. Conversations with the founders of the other Little Dragons often center on which events to attend and who has launched a new product.
His general standard for R&D payback is that some form of commercial return should appear within 3 years; beyond that, the opportunity must be reassessed. But if the opportunity is large enough, Qunhe can allow a small team to research it for years. Spatial intelligence, which began in 2018 and still generates limited revenue, is the exception he has permitted.
24. The spatial-intelligence pivot rewrote both the organization and the founder’s role
After 2023, 黄晓煌 handed commercial and user-facing work to the other partners and became more of a research leader connecting science and product. 陈航 now leads the B2B business, another co-founder is responsible for R&D and PLG, and 黄晓煌 leads the spatial-intelligence pivot. The 3 founders have rarely had strategic disputes since 2015-2016.
AI is also compressing the product organization. Work that once passed through requirements gathering, research, project approval, committees and scheduling can now begin with a model; 1 or 2 people can quickly build a prototype and put it in front of the market. 黄晓煌 wants to eliminate unnecessary coordination and let core capabilities move ahead of process.
The hardest part has been switching the old team over. When long-tenured employees resisted, he admits that he was slow to act, repeatedly hoping to train them and wait for them to change. He eventually handed the relevant teams to a partner and accepted the criticism that he was “avoiding the problem.” Musk’s rapid restructuring in Breaking Twitter struck a nerve precisely because it was the part 黄晓煌 found hardest to do himself.
One co-founder once ran the numbers beneath the Statue of Liberty in New York and concluded that the returns from years of entrepreneurship were lower than what he would have earned by staying at Amazon. 黄晓煌 does not measure the outcome against a salary: money is only a tool for advancing an ideal, while the real objective is to build a product with global impact. “If the understanding is wrong, everything is wrong.” For the current phase, his definition of success is for spatial intelligence to contribute half of Qunhe’s revenue.