152: Qianxun Intelligence’s Han Fengtao: The Embodied Model Knockout Race and the Non-Consensus on Commercialization
Summary
- Han Fengtao’s core non-consensus view is that 2026 is not the “make-or-break year for commercialization” in embodied intelligence, but the first year of explosive model-performance gains. He compares today’s embodied foundation models with GPT-2: they show a little intelligence, but building Agents and delivering at scale still requires heavy engineering; truly large-scale commercialization will have to wait until foundation models approach GPT-3.5, which he dates to the second half of 2027 through 2028. In 2026, investors should focus more on model iterations and scaling curves than on equating orders and revenue with technical leadership.
- The data bottleneck is shifting from “is there a solution?” to “who can collect and train at the highest acceleration,” which is why Qianxun is betting on 1 million hours of real-world data. Its recipe combines video pretraining, wearable real-world work data, teleoperation fine-tuning and reinforcement learning in real-world settings: it accumulated roughly 200,000 hours of video over 2 years, while nearly 100 iterations raised wearable-data usability from 20%-30% of teleoperation data to roughly 95%. Using an open-source model previously trained on about 10,000 hours of teleoperation data as a reference, Han says Qianxun’s 2026 data volume will be 100x larger and “the model’s overall performance will improve dramatically.”
- Qianxun is integrating the model, robot body and commercialization not because the brain is unimportant, but because Chinese startups must rapidly turn technical advantages into customers and market position. Early customers will not buy the brain and body separately and assemble them themselves; the complete robot remains the interface for gathering usage feedback, private-domain data and revenue. But Qianxun ultimately intends to sell a standardized “robot + foundation model + data-management platform,” allowing customers to fine-tune Agents for the physical world themselves. Han sets annual sales of 100,000 units as the survival threshold and says bluntly: “Speed is our only advantage.”
- The commercial ceiling in this cycle comes from general-purpose models expanding the range of work robots can do by an order of magnitude—not from repeating the import-substitution playbook for industrial robots. Roughly 500,000 industrial robots were sold globally in 2024 at an average price of about $20,000, putting the total market at only about $10B—“equivalent to Li Auto selling six months of cars.” China has about 3 million industrial robots in operation, yet penetration against its 100 million manufacturing workers is only about 3%.
- The CATL project showed that model capability, scenario selection and industrial experience all have to work at the same time. From nearly 200 battery-production processes, Qianxun selected end-of-line insertion and removal testing: it requires no major retrofit, needs VLA and force control, runs at a relatively slow takt and allows retries after failure, while also creating a reinforcement-learning loop on a real production line. The full project took about 11 months; based on models from the first half of 2025, deployment would take roughly 2-3 months, while the latest model should reduce that to about 1 month. Han’s conclusion is that the weaker the foundation model, the more the “last 100 meters” resemble a high-cost project—and the less scalable they are.
- The 2026 financing split will revolve around brains, capital and data-collection acceleration, with Qianxun targeting a Top 3 position in global model capability. The company raised only RMB30M in its first round at a RMB300M valuation; it has now completed nearly RMB2B in financing at a valuation above RMB10B. Han believes latecomers with strong AI teams and $100M-$200M can still use public experience to shorten the time spent learning painful lessons. The real ticket to entry also includes ample cash, industrial investors and a commercialization base—not merely a large historical data stockpile.
- The most important asset Han brought into his second startup was not optimism, but the anti-scaling traps and operating discipline left by his first company. Robotic once built a globally leading knife-sharpening workstation combining vision, force control and trajectory planning, only to discover global demand was just 8 machines; financing mistakes and reckless expansion also brought the company to the brink of collapse 3 times, leaving its 3 founders with more than RMB10M in joint and several debt. The principles that followed were: “Focus before solving the bottleneck; pursue speed after solving it.” Qianxun will commit decisively to anything on the main track, while voluntarily abandoning short-term revenue that is overly vertical and impossible to replicate.
Deep dive
1. The Main Battleground in 2026 Is Models, Not Orders
Han Fengtao explicitly rejects the “if embodied intelligence cannot commercialize in 2026, the industry is finished” narrative. In his framework, the core of embodied intelligence remains the foundation model, and 2026 should first be the year when data scale and foundation-model performance begin to explode.
Today’s embodied models are roughly at GPT-2 level in his view: “They look like they have a tiny bit of intelligence,” but building Agents or deploying them at scale on that basis is painful. Only after reaching something like GPT-3.5’s baseline capabilities will it be possible to build genuinely usable, replicable robot products.
That does not mean avoiding commercial validation. Han’s timeline puts large-scale commercialization in the second half of 2027 through 2028; in 2026, the industry needs staged model releases and clearer scaling laws so capital and industry can continue to see performance improving.
2. The Limits of Handwritten Rules Were His Earliest Entry Point Into Embodied Models
As an undergraduate at Huazhong University of Science and Technology, Han competed in robot soccer, using a roughly 15-centimeter-square vehicle and golf balls to simulate a match. Every rule was handwritten. “You can’t cover every corner case by hand,” an experience that became an early source of his later pursuit of general-purpose control.
From 2008 to 2011, he studied pattern recognition and intelligent control at Zhejiang University’s School of Control Science and Engineering. Neural networks were in a trough at the time; some people reduced multilayer networks to a single layer and multiple neurons to one. The running joke was: “Isn’t a single neuron just a kind of PID?”
He missed the deep-learning wave triggered by AlexNet at the end of 2012. With few robotics jobs available in China in 2011, he joined China National Nuclear Power; after the Fukushima accident slowed nuclear construction, he translated US third-generation nuclear-power drawings and adapted them for domestic production, returning to industrial robotics at Sinomach only in 2013.
3. The 2014 Industrial-Robot Boom Gave Rise to His First Startup
China became the world’s largest single industrial-robot market in 2014. Han remembers domestic sales at roughly 80,000 units that year, prompting media to call it “China’s first year of industrial robotics.” In 2015, he founded Robotic with Tuo Hua and Cao Hua, who led software and hardware respectively; Han took responsibility for algorithms.
Their startup mindset was classic “ignorance is bliss”: if you had the technology and built the product, you assumed you could “start a company in 3 years and go public in 5,” without even knowing how financing worked. Their first business plan ran more than 60 pages in Word; an investor told them they should be using PowerPoint.
The result 10 years later was not without achievements: Chinese-made industrial robots’ share rose from less than 3% in 2014 to more than 50% in 2024. But Han acknowledges that import substitution fulfilled an industrial mission; it did not automatically translate into high profits or large market caps.
4. The Biggest Misjudgment in Industrial Robotics Was the Market Ceiling, Not the Technology
Roughly 500,000 industrial robots were sold globally in 2024. At an average price of $20,000 each, the total market was about $10B, a pool that must accommodate ABB, Fanuc and every Chinese company. “$10B is equivalent to Li Auto selling half a year’s cars,” far smaller than the boom-era imagination suggested.
The industry once popularized the idea that 500 units would break even and 1,000 units would generate profits. The threshold then became 2,000 and 3,000 units; today, some companies sell 10,000 units a year and still lose money. The reasons include limited total demand, highly standardized products and price competition as the industry enters its second half.
Robotic initially wanted to build only robot controllers, believing they were the “brain” at the time. It quickly discovered that the controller market was smaller, and software performance could not be separated from the robot body. Han repeatedly came back to one lesson: robotics—and autonomous driving—are combinations of software and hardware; treating hardware companies as mere “blacksmiths” is unrealistic.
5. The Previous Generation of AI Could Perform Tricks, but Could Not Create a General-Purpose Market
Robotic used force control, sensors, 3D vision and deep learning to make robots perform tasks such as sewing and knife sharpening that traditional rule-based systems could not handle. Han compares adding sensors to giving a phone more cameras or a computer a graphics card: raising the hardware ceiling creates room for software to release more capability.
But first-generation deep learning was too weak. Robots could only tackle highly specialized processes one by one, not form a general-purpose brain. Facial recognition and smart speakers demonstrated technical progress in his view, but were not “genuine technological revolutions.”
GPT-3.5 changed his mind. Borrowing Wang Jian’s line that “a genuine technological revolution will always give birth to great companies,” Han defines Qianxun’s mission as bringing the foundation-model revolution from the virtual world into the physical world and building a general-purpose base capable of controlling multiple types of robots and performing many kinds of work.
6. Chinese Startups Must Control the Brain, the Body and the Customer at the Same Time
Manqi argued that some US companies can focus only on the brain and use third-party robot bodies. Han believes the difference between China and the US begins with exit environments. US technology companies may be acquired by Google or Tesla; Chinese internet giants tend to build in-house, so Chinese startups cannot treat acquisition as their primary survival path.
Customers buy a smarter robot, not a brain that they must source, assemble and debug themselves. Owning the hardware also protects performance and onboarding experience, while giving Qianxun a direct connection to customers, usage feedback and real-world scenario data.
Han stresses that full-stack does not mean “doing a little of everything.” AI, robotics and To B commercialization must all be strong. If any link is missing, the end-to-end delivery chain breaks; a startup’s only structural advantage is turning a temporary technical lead into a position in a focused market faster.
7. Annual Sales of 100,000 Units Is the Survival Threshold for Becoming a “Mid-Sized Player”
Han defines “mid-sized” as selling at least 100,000 robots a year and believes a company at that scale “will definitely be profitable.” Using the 6-7-year ramp-up cycle for new-energy vehicles as a reference, he thinks general-purpose robots could scale faster; Qianxun’s internal target is 2030.
His description of the global market in 2030 retains clear boundaries: the market will definitely reach the million-unit range, but not necessarily the 10-million-unit range. The end state may have only 5-8 comprehensive robot makers, with the leaders collectively taking about two-thirds of the market and the rest sharing the remaining third.
Qianxun’s “Double Ten Plan” is to give 10% of the world’s population its own robot within 10 years. Industry alone cannot support that target, so the roadmap must move from early industrial scenarios toward services and households.
8. The Brain Can Push Robot Penetration From 3% Toward Order-of-Magnitude Growth
Han puts China’s manufacturing base at roughly 3 million industrial robots in operation and about 300,000 new units added annually, serving approximately 100 million manufacturing workers. On an installed-base basis, penetration is about 3%; on annual sales, it is about 0.3%. The supposedly mature market remains barely covered.
The core problem is not a shortage of robotic arms, but that “the machines are too stupid.” They can handle only simple, structured work. If embodied models allow robots to take on roughly 30% of tasks, industrial equipment demand could rise by at least an order of magnitude.
The larger ceiling lies outside factories. The same brain can extend to hotels, retail, logistics and households. Mass manufacturing combined with software generalization is why Han believes 100,000 robots can be profitable, while 10,000 industrial robots may still lose money.
9. Before Finding Gao Yang, He Screened More Than 100 Potential Partners
In 2023, Han went through roughly 6,000 WeChat contacts, searching from A to Z for people who could introduce him to top AI talent. Over more than half a year, he met more than 100 candidates online and offline. His biggest takeaway was: “Reliable people are too rare; scammers are too numerous.”
He encountered people who wanted to raise money by riding the sector’s momentum, people committed to old AI approaches who nonetheless believed they were suited to foundation models, and people who demanded to bring their entire old team into the new company. Embodied intelligence is at least a 20-year undertaking; beyond technical track records, the harder task is finding a partner willing to endure the long test of human nature.
Gao was also looking for a CEO with industrial and hardware experience. The two met in June or July 2023 and decided to work together after more than 10 conversations. Han valued Gao’s experience with large-scale internet-video training and his work on end-to-end autonomous driving dating back to 2017-2018.
10. The Foundation of Cofounder Fit Is Shared Values, Not the Absence of Arguments
They discussed more than technical direction: division of labor, research versus commercialization, how long they intended to stay at the company, and what to do when they eventually “slammed the table or even cursed at each other.” Han’s view is that arguments are inevitable in a startup; teams break apart when their values diverge under pressure.
He describes Gao as “an extremely intelligent and purely upright person”: when he does not know something, he says so clearly; when verification or study is needed, he handles it. When colleagues offer suggestions for improvement, he accepts them calmly and explains what he will change next.
Gao initially thought Han spoke bluntly and had sharp views, “a bit like a businessman,” and only later realized that he also understood technology. When early investors said Qianxun’s CEO did not understand AI, Han did not evade the criticism: “I really don’t understand AI as well as Gao,” and then systematically caught up.
11. Trust and Excellence Are Qianxun’s Mechanisms for Lowering the Cost of Error Correction
When Qianxun moved into its first office, it had only 11 people, including Gao’s intern. In his opening speech, Han emphasized just 2 things: trust and the pursuit of excellence. “Only a team with trust has fighting power,” and only an excellent company can survive financing and competition.
His definition of trust is concrete: when colleagues offer opinions or suggestions, first interpret them as good-faith input rather than as a personal attack. It is difficult to discover one’s own mistakes; if others can point them out and decision-makers are willing to accept the criticism, organizational correction becomes much easier.
Integrity means respecting objective facts: whether the technology is truly your own, how well it actually works, and whether assessments of competitors are evidence-based. Passing off open-source work, exaggerating capabilities or fabricating rumors about competitors all qualify as “dishonesty” in his view.
12. Rapidly Rising Valuations Pulled the Sector Into a Narrative War Early
Manqi questioned why competition was already as intense as in a mature industry when embodied intelligence had yet to commercialize at scale. Han’s explanation is that the industry remains in “the darkness before dawn,” but valuations are rising quickly; players disagree about the brain, hardware, data and commercialization pace, while capital has difficulty verifying the claims. Financing competition therefore spills into narrative competition.
He confirms that the industry already has PR firms, paid-for articles and coordinated online attacks. Some people have even spread rumors that he and Gao are at odds. Manqi called it the most absurd smear she had heard, because she had never seen Gao constantly slamming the table and arguing.
For rumors that cannot be disproved directly, Han used the metaphor of “how many bowls of noodles you ate”: “Unless I cut open my stomach, there is no other way.” In his view, that shows competition is no longer confined to the technology itself.
13. By the Second Half of 2025, Capital Consensus Had Rapidly Converged on “The Brain”
From 2024 through the first half of 2025, components, robot bodies, sensors, specific scenarios, pure-model companies and integrated software-hardware companies could all raise money. By the second half of 2025—especially in the few months before the interview—capital rapidly shifted toward the embodied brain, as investors began to believe the primary data bottleneck had been solved.
Qianxun raised only RMB30M in its first round at a valuation of about RMB300M; its latest financing approaches RMB2B, with valuation above RMB10B. Manqi noted that some companies founded in 2025 raised more than RMB1B in their first 1 or 2 rounds, reflecting the shift in capital’s understanding.
Han believes 2026 will resemble 2023 for large language models: if a company cannot secure enough funding or reach the top tier, it may lose the opportunity to “get a seat at the table.” “If we can’t get a seat at the table in 2026, I think it’s over.”
14. Data, Scaling and Commercialization Form Three Consecutive Gates
In Han’s technical roadmap, embodied models are essentially “large models for the physical world.” The first gate is obtaining enough usable data; the second is whether model scaling can accelerate fast enough once the data bottleneck eases; only the third is deploying a foundation model near GPT-3.5 level into specific applications at scale.
This sequence explains why Qianxun is unwilling to chase revenue too early. When basic capabilities are too weak, every project requires extensive dedicated data collection, post-training and engineering. It can pass acceptance tests, but cannot be replicated. Once the foundation model crosses the emergence point, the productization workload drops sharply.
Han acknowledges that this is a conditional judgment. If model capabilities do not improve as expected toward GPT-3 or GPT-3.5 in 2026 and 2027, the technology’s failure to advance will itself become the company’s failure path.
15. Qianxun Chose Real-World Data Over Simulation From the Start
When it set its data roadmap in 2024, the team barely considered simulated synthetic data as a primary path. Han’s technical judgment is that the sim-to-real gap is difficult to close; autonomous driving is a simpler robot application that tries to avoid contact with the environment, yet still relies mainly on real data. Robots performing contact-heavy tasks will find it even harder to rely on simulation alone.
Qianxun’s data recipe has 4 layers: roughly 200,000 hours of internet videos showing human work for pretraining; high-quality but expensive and inefficient teleoperation for fine-tuning; wearable devices for low-cost, high-volume collection of real-world work; and robot rollouts in real scenarios to collect failure cases for reinforcement learning.
Video offers scale, low cost and broad scenario coverage, but lacks precision; teleoperation is precise but does not scale. Wearable data fills the middle gap, while real-world reinforcement learning sends problems exposed in specific environments back into the training loop.
16. The Key to Wearables Is Data Usability, Not Form Factor
UMI can record hand movements relatively well, but cannot fully capture the body, arm angle or null-space motion of a 7-degree-of-freedom robot arm. Exoskeletons can fill in arm configuration, while motion capture can provide more precise full-body data. Qianxun therefore is not betting on a single device, but on a combined “full-body UMI” data recipe.
From May 2024 through November or December 2025, the team iterated through 4 hardware generations and dozens of minor versions, making nearly 100 improvements in total. The work covered device design, segmentation, labeling, quality control, data pipelines and model algorithms for learning from low-quality data.
Initially, wearable data had only about 20%-30% of the usability of teleoperation data, requiring 3-5x as much data to learn the same action. After iteration, usability rose to about 95%. Qianxun began scaling collection in January 2026, expects to reach more than 100,000 hours in the second quarter and plans to collect 1 million hours for the full year.
17. Data-Collection Acceleration Matters More Than Historical Stockpiles
Han believes competitors should be measured not only by how much data they already have, but by whether their collection speed can continue to accelerate. The drivers of acceleration are the accumulated work of the past 1-2 years on devices, data-management platforms, training pipelines, data algorithms and engineering efficiency.
He uses Physical Intelligence and Qianxun’s best previously released open-source model as reference points: about 10,000 hours of teleoperation data collected over nearly 2 years. Qianxun’s 2026 target is 100x that data volume, with usability reaching 90%-95% of teleoperation data.
That does not guarantee a 100x performance increase, but Han is highly confident in the direction: “We will get 100x the data, so the model’s overall performance will improve dramatically.” Whether the model exhibits scaling as expected will be the market’s direct test of his thesis.
18. Differences in Chinese and US Capital Markets Force Chinese Teams to Prove Both Technology and Commercialization
Qianxun only discovered how closely its direction resembled Generalist and Sunday’s after the 2 companies publicly disclosed their wearable data-collection plans in the second half of 2025; Qianxun had already been working independently for more than a year. Han believes US companies can secure capital without a demo or commercialization, while Chinese companies must show a demo or commercial progress to keep raising money.
His extreme hypothetical is: “If you gave me $500M in 2024, I would do nothing else and go all in on this data.” If Qianxun had received $200M-$300M on day 1, it would also have devoted more resources directly to data rather than simultaneously developing demonstration projects.
Under real-world constraints, Qianxun still puts more than 80% of its effort into breaking through the data bottleneck, with the rest going to demos and commercialization exploration. Han sees this as a cost Chinese startups must pay given the financing environment.
19. Commercialization Exploration Is Early Product Selection—and a Bid for Private-Data Access
Large language model capability is close to a product in itself: add a dialogue box and you have a chatbot. A robot model is only one part of the product. Once the model reaches GPT-3.5 level, Qianxun cannot wait until then to decide whether to build battery, automotive, cleaning, massage or retail robots; the form factor and requirements must be researched in advance.
Embodied intelligence also involves private data that large language models encounter less often. Data from a CATL factory or a JD logistics hub cannot be collected by just anyone with site access. Qianxun is partnering early both to validate products and to “claim the data positions first.”
Early scenario data can be used to train a shared foundation model; once the foundation model matures, high-precision fine-tuning data collected by customers belongs to the customers. Qianxun wants to reserve about 90% of its effort for foundational capabilities, with a small team handling market development and early data access.
20. The First Step in the CATL Project Was Choosing the Right Process From Nearly 200 Options
CATL sent its requirements to almost every embodied-intelligence company, especially companies founded in 2024. Gao walked through the factory wearing a helmet and protective gear, covering roughly 20,000 steps in a day. Han required technical leaders to speak both “the customer’s language” and “the technology language,” rather than understanding the task only from an office.
Qianxun ultimately chose end-of-line insertion and removal testing for batteries. It required no custom automation retrofit: remove the human and install the robot. The task required both end-to-end VLA and force control, while its relatively slow line speed made it suitable for an early-stage model.
The process also had a high tolerance for error. If the first insertion failed, the robot could retry, and the large plug was difficult to damage. More importantly, the robot could operate on a real production line and continuously collect reinforcement-learning data, completing the loop from data collection to training, deployment and feedback.
21. Scenario Selection Is Rarer Than Simply Winning a Customer Requirement
Some teams arrive on-site and talk only about their own technical strengths, prompting the customer to ask: “What does this have to do with me?” Other companies can discuss solutions but select scenarios where each factory needs only 3 robots; even a successful model cannot create a market.
Han says a suitable scenario must avoid bottlenecks in the model, hardware and organizational coordination while also offering enough deployment volume. A robot may complete the task but fail on payload, takt time or durability; even if the technology works, demand from a single factory may be too small to have commercial value.
This is why he believes industrial experience cannot be replaced by pure model capability. Only with strong AI, robot bodies and To B sales is it possible to identify, from 200 processes, the one that can be deployed, generate data and scale commercially.
22. Each Foundation-Model Generation Cuts the Cost of the Last 100 Meters
The CATL project took about 11 months from initial contact to acceptance, including extensive non-model work. With model capabilities from the first half of 2025, deployment would have taken roughly 2-3 months; with Qianxun’s latest model, Han estimates about 1 month.
The first project required R&D staff to remain on-site for long periods, writing code and modifying the model in the factory. The Beijing R&D team could not simply send down a model and ask a field FAE to handle the rest. Qianxun even created a dedicated CATL team, with every other department instructed to give it the highest priority when needed.
Han sees this heavy service burden as a consequence of weak models, not as a future organizational template. Once the foundation model reaches GPT-3.5, Qianxun can teach customers to collect data and fine-tune models, avoid taking responsibility for the last 100 meters of deployment, and sell a standardized product instead of endlessly expanding its project teams.
23. The End Product Is a Combination of Robot, Model and Data Platform
Qianxun wants to provide standardized robots, foundation models and data-management capabilities to industrial partners such as Huawei, Xiaomi, JD and Inovance, all of which have their own embodied-intelligence teams. Customers understand their own processes and can fine-tune and deploy on top of Qianxun’s base.
Han compares the model to “building physical Agents on our physical embodied model.” Qianxun will not complete every customization for every enterprise; it will provide a sufficiently strong base and tools so customers can handle downstream adaptation themselves.
This is also the dual purpose of bringing in multiple industrial investors: secure data and demand from closed scenarios while developing delivery partners capable of using the foundation model independently. Commercialization can scale only if deployment capability spreads into customers and the broader ecosystem.
24. The First Vertical Markets Must Meet 5 Conditions at Once
Qianxun evaluates To B markets against 5 criteria: the market must be large, still growing, concentrated on a limited number of customers, supported by strong willingness to pay, and as consistent as possible between China and overseas so products refined in China can be taken abroad directly.
Power batteries are one such scenario; the team is also focused on logistics, hotels and retail. Han stresses that embodied intelligence is not about replacing mature automation, but entering tasks that existing systems cannot do and that require generalization across actions and environments.
Retail robots could retrieve goods when shelves are not perfectly organized, and may eventually handle inventory checks, replenishment and cleaning. Hotel robots could clean sinks or collect and fold towels. Traditional dedicated machines can perform fixed actions at fixed locations; what they lack is generalization across objects and environments.
25. The Best Early Tasks Are Those Where Traditional Technology Has Already Done 90% of the Work
Han offers another screening method: traditional robots and previous-generation AI can already complete roughly 90% of the task, leaving only the final 10% of generalization. Embodied models should solve that critical gap rather than reinventing the entire hardware stack.
The mechanics of folding clothes are already executable. What robots truly do not understand is what counts as a wrinkle or a flattened surface, and where the collar, sleeve or corner is. The same applies to insertion and removal: the mechanical system is mature, and AI only needs to adjust actions in real time when positions change.
His optimism rests on this boundary. The goal is not for a GPT-3.5-level model to create a complete product from scratch, but for it to fill in the unstructured portion that traditional robots have long been unable to handle. Once the model crosses the threshold, commercialization should move quickly.
26. The Model Threshold for Household Robots May Arrive First, but the Safety Threshold Will Not Disappear
Han believes that technically, by 2027 robots may be able to do laundry, basic cleaning, tidy children’s rooms, water plants and arrange shoes at home. The real barrier to general-purpose robots entering households may not be capability, price or privacy, but safety above all.
Existing robots weigh dozens of kilograms, and a fall could injure children or elderly people. Their batteries are typically about 3x the capacity of an electric bicycle, whose battery may not even be allowed inside a home. Until solid-state batteries become more reliable, robots will either need smaller batteries and frequent charging or have to work while tethered.
Households may still see specialized robots with generalization capabilities first—for example, a small arm mounted beside a washing machine that transfers clothes to a dryer and folds them. It would not need a large battery or full-home mobility, and if priced around $1,000, it could reach users earlier than a general-purpose dual-arm robot.
27. Qianxun Would Rather Cap 2026 Revenue at RMB100M Than Chase Low-Quality Orders
Qianxun’s main quantitative target for 2026 is 1 million hours of wearable data and the model iterations built around it; revenue from deployments and platform services for industrial partners is planned at about RMB100M.
Han has voluntarily abandoned revenue such as “vegetable factories” that is unsustainable and does not help the model climb the capability curve. He acknowledges that this plan looks restrained for a company valued above RMB10B, but believes that when the foundation model has not crossed the threshold, higher revenue does not necessarily mean the company is closer to the right product.
He divides the industry’s current “commercialization” into several categories: legacy automation or previous-generation AI, large orders that have not truly been deployed, and standalone niches such as singing and dancing. An embodied model at today’s GPT-2 level “cannot be deployed in batches,” so revenue cannot directly represent foundation-model leadership.
28. The True Revenue Inflection Comes From Zero-Shot Capability, Not Project Count
Han’s envisioned inflection point is when a robot can enter a new environment and achieve a 70%-80% success rate zero-shot, then quickly reach a high success rate after a small amount of teaching. At that point, the model may be equivalent to GPT-3.5 or GPT-4, and standardized robot models can ramp rapidly.
Before then, every scenario requires extensive data collection, fine-tuning and engineering, leaving the revenue curve shallow. After crossing the emergence point, standardized products can enter many scenarios simultaneously, and the curve steepens. This is why he places true large-scale commercialization in the second half of 2027 through 2028.
Manqi asked what could obstruct the curve. Han’s answer was forceful: customer demand, hardware, supply chains and capital are all relatively mature; “there is only one bottleneck—raise basic capability.” But that judgment also concentrates model progress as the company’s largest single-point risk.
29. Getting a Seat at the Table Requires AI, Capital and Industrial Access to Work Together
Han’s first condition is a strong AI background, whether built over the past 2 years or brought in from autonomous-driving and multimodal teams. The second is that valuation and cash on hand must cross a higher threshold; a company still valued at RMB1B-RMB2B may find it difficult.
Chinese companies also need industrial investors, a commercialization plan and a base of real deployments. Even a pure technology team with a good model must enter closed scenarios to obtain data and prove it can deliver the technology as a product.
Qianxun’s internal targets are 1 million hours of data, Top 3 global model capability, and a financing position and company standing among China’s leaders. Gao told Lei Jun that Qianxun’s domestic technical lead was roughly “6 months”; Han attributes those 6 months to 2 years of continuous focus on the core bottleneck.
30. Latecomers Can Still Catch Up, but They Must Use Capital to Buy Time
Han actually believes some companies founded in 2025 may have a higher probability of becoming large, because they have no historical baggage and can focus directly on the technical mainline that has already converged. Public industry experience also means they do not have to repeat all of the year-and-a-half of mistakes made by their predecessors.
But time cannot be compressed for free. He estimates latecomers need at least $100M-$200M to fund the investment required to catch up; even then, they may retain a gap of several months relative to first movers.
Those months may not be immediately visible to the market. Just as autonomous driving and China’s large-model market can accommodate multiple leaders, embodied intelligence will not belong to Qianxun alone; a temporary model gap does not automatically produce a winner-take-all outcome.
31. The Real Entry Point for Big Tech May Come in 2028-2029
Han tells the company that its real competitors are not other startups, but large companies such as Huawei, Xiaomi and Li Auto that can develop software and hardware together. They are more likely to do embodied intelligence well; the market is simply still small and commercialization has not yet scaled.
His judgment is that large companies may begin making major strategic investments in 2028 and 2029. Even if a model breakthrough appears in 2026, large companies may not immediately elevate embodied intelligence to a company-wide priority as long as the market has not generated meaningful revenue.
The difference is that “for them, this is a job; for us, this is our life.” Startups must become “mid-sized players” selling 100,000 units annually before big tech deploys massive resources. Neither talent nor capital can substitute for speed.
32. The Knife-Sharpening Workstation Proved That Technical Leadership Does Not Equal Commercial Viability
Robotic once developed a ceramic-knife sharpening workstation for Kyocera. Ceramic knives and diamond grinding wheels are both hard and brittle, so even slightly inadequate force control could cause chipping. Each knife also had a different shape, requiring 3D vision, force sensing and trajectory planning to work together and achieve machining accuracy of about 0.5 millimeters.
The team invested more than a year and outperformed Mitsubishi, achieving what Han called “world-leading” performance. But global demand was only 8 machines. Even at RMB1M per machine, the total market was just RMB8M—“we had developed an ultimate dragon-slaying technique.”
It became Han’s favorite self-deprecating example in his second startup. When the foundation model is immature, every deployment may require knife-sharpening-level engineering; if the product cannot be reused across industries, even beautiful technology cannot scale.
33. Three Cash-Flow Crises Taught Him Financing, Budgeting and Expansion Discipline
Robotic started in 2015 with more than RMB3M and originally planned to build only software. After switching to complete machines, it quickly discovered that the budget was insufficient; in 2016, the CEO borrowed from relatives to keep the company alive for 2 months. The first lesson was that the team had underestimated the cash required for hardware development.
In 2017, an investor promised to put in RMB60M, so the team rejected other funding. Near signing, the investor changed the deal to RMB30M in equity plus RMB30M in convertible debt, while cutting the valuation and adding restrictive terms. The 3 founders ultimately assumed more than RMB10M in unlimited joint and several liability; Han’s monthly salary at the time was only RMB8,000.
The rule left by that crisis was direct: “Do not reject any investor before the documents are signed and the financing has finally closed.” A financing commitment is not cash; a startup must always preserve its fallback options.
34. Reckless Expansion Creates Fatal Illusions More Easily Than Difficult Financing
After raising RMB160M from Shunwei in 2018, Robotic established 4 business units at once, nearly tripling its headcount in a short period and spending almost all the money in little more than a year. By the end of 2019, the company had only about 4 months of cash left and finally realized the expansion was out of control.
The team cut headcount from roughly 300 to 140-150 and reduced everyone’s pay. The 3 cofounders’ monthly salaries fell from RMB30,000 to RMB15,000. Han’s after-tax income was about RMB11,000, but his monthly mortgage was RMB40,000, so on friends’ advice he borrowed RMB50,000 each from multiple people to maintain household cash flow.
He does not describe the experience as a heroic story. He says only that they “had to find a way to solve it,” quoting: “Don’t worry about what is 8 kilometers away or 2 hours from now.” He also considers the pandemic a blessing in disguise: without the earlier layoffs, the company might not have survived 2020.
35. Moving From Robot Thinking to AI Thinking Was the Biggest Cognitive Shift in His Second Startup
At the beginning of the venture, although Han knew he needed a top AI partner, he still made decisions within a traditional robotics framework. Since inverse kinematics and dynamics were already mature, should they be left as much as possible to the cerebellum, with the foundation model outputting only higher-level instructions?
By the second half of 2024, he gradually accepted that embodied intelligence is primarily an AI track, and that data matters more than existing control modules. Qianxun now compresses the cerebellum into basic motion control—choosing the next point of movement and handling basic kinematics and dynamics—while leaving spatial understanding and trajectory planning as much as possible to end-to-end model learning.
This was not an abstract debate. Before its own robot was finished, Han wanted to wait for identical hardware before collecting data. Gao insisted on spending nearly RMB2M to buy external equipment and iterate the model first. The dispute delayed progress by more than a month; Qianxun ultimately followed Gao’s plan and decoupled data progress from robot-body progress.
36. Main-Track Investment Depends on Professional Intuition and Organizational Error Correction
The 2 founders eventually reached a shared view: whenever an investment sits on the main track and can improve model performance, “invest decisively.” For AI, models and large-scale training, Han relies more heavily on Gao’s judgment because Gao has studied AI continuously for more than a decade since moving to the US in 2014.
He acknowledges that experienced practitioners’ decisions are often not backed by 100% solid evidence, but by intuition formed over more than a decade. That intuition may be right or may carry an old paradigm into a new problem; mistakes cannot be eliminated entirely.
The safeguard remains a culture of trust: when someone discovers that a decision is wrong, they can say so directly, and the decision-maker is willing to accept it. At Qianxun, trust does not mean a harmonious workplace; it means different professional intuitions can clash and the company can still correct course quickly.
37. The Company Repeatedly Discusses How It Could Fail Instead of Treating Optimism as a Conclusion
The team periodically brainstorms failure paths, including choosing the wrong data or model roadmap, performance improving less than expected, commercialization moving more slowly than planned, and large companies elevating embodied intelligence to a strategic priority earlier than expected.
Han’s current conclusion remains optimistic: a weak model has already been deployed at CATL, while hardware and supply chains are relatively mature. If foundation models evolve rapidly in 2026 and 2027, commercialization should accelerate with them. But all of that confidence depends on scaling actually occurring.
His attitude toward entrepreneurship comes from his first experience: “Focus before solving the bottleneck; pursue speed after solving it.” The fact that outsiders cannot yet see the model, orders or financing is not the real crisis; losing control of core progress internally is.
38. The End Point of This Opportunity Is China’s First Chance to Stop Playing Catch-Up
Han summarizes more than a decade of Chinese hard-tech entrepreneurship as import substitution. Robotics, semiconductors, machine tools, cars, drones and home appliances all started by catching up with Europe, the US and Japan. That generation of founders endured enormous hardship, but it also enabled the large-scale localization of hardware brands now surrounding Chinese consumers.
In his view, embodied intelligence is one of the few fields where China and the US are starting almost simultaneously. China also has systemic advantages in talent, data collection, application scenarios and cost. He therefore does not want to build another vertical robotics company, but to compete for a position as a top global platform company.
In the episode’s postscript, Manqi added a risk: one investor’s count put the number of companies in the sector valued above $1B at roughly 15, while the previous generation of intelligent-robotics companies has yet to produce a widely recognized breakthrough. Whether today’s excitement compounds into value that survives the cycle will depend on whether teams can truly tackle both frontier AI and the brutal, fragmented reality of industrial deployment.