Summary

At 22, 黄一 completed 5 funding rounds totaling more than $100M in roughly a year and a half, sustaining a cadence of one round every 1-2 months. He admits embodied AI “of course” has a bubble, but his strategy is to ride the trend and take as much as he can: embodied AI has entered the 15th Five-Year Plan, while secondary-market capital is flowing through LPs into GPs, driving a wave of follow-on rounds this year; the company defines RP One’s launch and move into mass production as its Series A milestone, with the commercial loop reserved for Series B.
Roboparty is a robotics-platform company with open source as a core component, occupying the “missing layer” between model companies and incumbent hardware makers. Existing manufacturers struggle to let model companies develop at lower layers, including the motor layer and the URDF variables for stators and rotors; open source lowers collaboration friction and creates user compounding—sell one platform for RMB100K while gaining 2 engineers earning RMB1M a year to help with R&D and publish papers. The analogy is DeepSeek: open source does not stop the company from selling API access. Hardware switching costs exceed software switching costs; the moat lies in the company’s fundamentals, not open source itself.
The financing playbook: secure as much early allocation as possible, because the first financing agreement can follow a company all the way to Pre-IPO, making terms more important than valuation. Roboparty has never agreed to a QIP buyback or a performance-based arrangement; deep-pocketed funds should be saved for later rounds—give one 10 percentage points in the angel round and it has less incentive to keep adding later. Dollar-based institutions focus more on technical or product leadership, while RMB funds focus more on commercialization; and “the later the round, the lower investors’ level of expertise,” with ICs often “investing blindly” once they hear who is leading.
His industry view: the US, including Chinese researchers there, leads on the embodied-AI brain, while China is “far ahead” in humanoid platforms and the cerebellum/control stack—essentially a generational lead. He makes a sharp bet that wheeled robots are an intermediate state—a “legged system with infinite leg inertia”—with little long-term significance. Ten years out, he favors model companies such as Zhipu and OpenAI: “robotics cognition is cheap,” because it can be filled in by hiring the right people. Roboparty plans to enter models around 2027, and if OpenAI launches a Robotics business, its scaling capabilities will far exceed those of domestic startups and a whole group of Chinese model companies.
His industry map starts with more than 100 humanoid companies and more than 300 embodied-AI companies today, but only 10-15 may emerge by 2030. The second tier could be led by companies such as NIO, XPeng, Li Auto, Honor and Geely, with valuations or operating scale in the hundreds of billions of yuan; embodied AI could carry roughly 40% gross margins versus about 20% in autos. If BYD, Huawei, CATL, BAT and Xiaomi also enter, it will mark a major new era: “buying me would be enough to get a robotics ticket.”
On operations, RP0 ships roughly 100 units a month against genuine orders, with 30%-40% sold to schools; RP One is positioned for research and education, is expected to launch in Q4, and could sell about 2,000 units next year. A European institution once ordered 500 units at once but received only 20-30 because the supply chain then lagged delivery capacity. 黄一 sees the launch date as pulling the trigger, weighing the back-to-school procurement season, IROS and WRC, the industry’s move toward production and mass manufacturing, and competitor timing.
The organization runs on a honeycomb structure, with a management span of roughly 6 and zero voluntary departures in a year and a half, while executives once earning RMB2.5M-RMB3M at major companies have joined for RMB800K. Pay discipline comes from a review of the EV sector: well-funded companies that later blew up “basically all died from their burn rate.” He believes “arrogance is important,” that the peak of ignorance is a good time to start a company, cites 唐杰’s hierarchy of cognition over vision over technology over management, then adds effort at the bottom—and says, “I ate the hardship of 10 years in one.”

Deep dive

1. 22, 18 Months, 5 Rounds and More Than $100M: A Compressed Life Sketch

The basic profile: born in 2004, educated at High School Affiliated to Renmin University of China and Harbin Institute of Technology. He could graduate early after completing his credits and capstone; his thesis was on humanoid robots and remains posted on Zhihu. The team has roughly 140 people, nearly half of them interns, and completed 5 funding rounds totaling more than $100M over the past year. 黄一’s own assessment: “I ate the hardship of 10 years in one.”
On products, the current RP0 ships roughly 100 units a month, “almost all of them genuine orders.” The next-generation RP One is expected to sell about 2,000 units next year, launch in Q4 and carry an as-yet-undetermined price.
He first puts the fundraising figure in perspective: “That amount of money is actually not much in embodied AI, but it is still a lot for a young founder.”

2. Before the First Check: Funding the Robot Himself

The counterpoint to the “invest in young people” narrative: the prototype was already built when he raised the first money—“basically funded by me, probably several million yuan.” Investors were betting on a physical product. The first VC was Bill at 5Y Capital, who proactively contacted him through the robotics group 黄一 had built.
The company initially ran on 2 tracks, using a cash-flow business to support the humanoid effort. He later judged that to be a flawed premise: “In a major industry that requires high-frequency iteration, by the time your cash-flow business gets running, your humanoid robot has already been left behind.” The conclusion was that the company had to leverage VC capital.

3. In the Seed Round, the Real Negotiation Is Terms, Not Valuation

The core logic is to secure as much early allocation as possible, because the first investment agreement “basically stays with you for a very long time,” often all the way to Pre-IPO. Roboparty has never agreed to a QIP buyback or a performance-based arrangement: “It makes the pressure of starting a company a little lower, but only a little.”
Personal joint-liability buybacks are “basically nonexistent”; even state-owned capital generally will not ask founders to sign one. QIPO buybacks are a “sword of Damocles”—hanging over the company, while potentially difficult to enforce when the time comes. Details such as simple versus compound interest on buybacks require lawyers who have negotiated a large number of investment agreements to help avoid hidden risks.

4. Choosing Institutions: Save the Deep Pockets, Value the Exchange of Judgment

The logic behind saving deep-pocketed investors is worth remembering: funds capable of investing RMB500M-RMB1B become rarer in later rounds. If one takes 10 percentage points in the angel round, it “doesn’t have as strong an incentive to add more later”—the company has effectively given up a future option to raise substantial capital from that investor.
Early-stage institutions bring a different kind of value. Firms such as ZhenFund and Matrix Partners China provide resources and teach founders who initially understand little about the process; “the exchange of judgment can be worth far more than the quality of the money itself.”
Financial advisers can take founders from zero to one in the VC/PE world and explain fundraising and institutional decision-making. The screening questions are whether the adviser can provide a full view of the sector and answer “who is investing in what, and how accurate are their calls.” Of 100 investors approached, perhaps only 10 invest in embodied AI. The market is so flooded with FAs that a slot-allocation system has emerged, which he considers “completely baffling.”

5. Strategic, Dollar and RMB Capital: 3 Different Investor Profiles

VC investors bring networks and a founder community. Industrial capital should be taken “very cautiously,” but its ability to empower a company can be substantial: Xiaomi Robotics’ accumulated know-how and documentation could be shared directly. “If you get the strategic decisions right, you can avoid a great many detours.” When a shareholder is genuinely helpful, giving it a board seat can make the partnership work more smoothly.
The currency split is clear: dollar-based institutions are willing to invest because of technical or product leadership, while RMB institutions care more about whether commercialization and the business model work. “The later the round, the lower investors’ level of expertise.” Making a bet without anyone else’s endorsement requires real judgment, which is why early-stage institutions capture excess returns.
His sharper observation is that many later-stage institutions will “invest blindly” at IC meetings as soon as they hear who the lead investor is. Asked whether that was an isolated case or the norm, he did not retreat: “It actually happens quite a lot. There are not that many truly professional investors, but there is a lot of hot money in the sector.”

6. Living With the Bubble: Learning to Raise a Round Every 1-2 Months

He acknowledges the bubble without hesitation: “Of course.” VC is fundamentally “pulling future cash flows forward for use today”; once RMB2B has been raised, everyone’s lived experience is that the bubble is larger. The response is embedded in shareholder Shunwei’s name: ride the trend. “No company can change an era.” The momentum shows up in suppliers being willing to give the company a second look and talent flooding into embodied AI from other industries.
The cadence follows the milestones. Embodied AI entered the 15th Five-Year Plan, while people with excess money in the secondary market invested it into GPs as LPs, and GPs then invested in companies, producing a wave of follow-on rounds this year. The company’s standard is RP One’s launch and move into mass production for Series A, with Series B reserved for the commercial loop. Pointing to the “Stay Around” gym shirt he was wearing, he accepted the host’s interpretation with a laugh: perhaps the company could keep raising Series A rounds.
The registration philosophy is “register where the talent is”: the company’s Shanghai base is in Pudong, while its Beijing base occupies the 20th floor of Qidi C, with Tsinghua’s Institute for Interdisciplinary Information Sciences on the 19th. Local policy is being saved for a “headquarters relocation round”—moving headquarters from A to B near an IPO and using investment-attraction incentives to secure another financing round.

7. Arrogance Matters: The Peak of Ignorance, Shattered Mental Models and Romain Rolland

He insists: “It’s not that confidence is important; arrogance is important.” Before someone has a clear view of the industry, there is often a period of extreme arrogance that may be the best time to start a company—the so-called “peak of ignorance.” At the time, he looked at robots built by the strongest university labs and thought, “They’re nothing special.”
There was no classic valley of despair, but there was a cycle of having his mental model shattered and then rapidly rebuilt. After speaking with the head of Xiaomi’s robotics division, he realized that questions he had been wrestling with—whether to build the brain, where to put the data—were still “at the stage of asking questions,” while the other side already had a highly complete system.
Knowing too much has a cost. The data ledger includes Tera-level, Ego-level and 5M-level data, full-body human data collection and internet data. Both current and terminal prices for each layer can be estimated—a single data item currently costs about RMB200, with the end-state price visible once players such as JD.com enter—allowing a founder to “calculate when the ROI of entering is highest.” But that knowledge also makes people timid and cautious. So he cites Romain Rolland: “After recognizing the truth of life, continue to love life.” The younger a person is, the more fearless; attempting a startup in midlife may require more courage, “though it may also just be ignorance.”

8. Recruiting in 3 Days, the Highest Form of Selling a Dream and RMB2K-RMB3K Monthly Salaries

The longest recruiting process took 2-3 days. “He could go wherever he wanted,” so 黄一 accompanied him everywhere, even following him directly to Zhejiang University. The person is still with the company. Persuasion comes from delivering on the vision: when recruiting the first employee, 关建文, he promised to build “an extremely capable joint-module division” for him. Today, that division’s integration is strong enough to be spun out as a standalone module company.
His definition of selling a dream: “The highest form is when the other person doesn’t feel it is a dream—or it is something I genuinely believe myself: describing clearly a future they cannot yet see.” It is not, “I’ll raise your salary by this much.”
Harbin offered an early hiring advantage. The city had no serious technology companies, so roommates and classmates joined him for RMB2K-RMB3K a month, relying on the trust in his technical ability built through competitions and other earlier interactions: “My technology can piece together a robot; add one person for controls, one for mechanical design and one for hardware, and you can assemble a better robot.”

9. Open Source Is Not Idealism; It Is a Commercial Design for the Missing Layer

The starting point was simple: after finishing the first robot, he felt it was not very useful, so he open-sourced it and posted it on Zhihu. The positive feedback was seeing “3-5 people actually build the robot from your drawings,” which affected him deeply. The analogy is DeepSeek: “Open source only helps once you have made a good product.”
The business logic is that a threshold—fixed assets, people and a technology stack—is not the same as a moat. The money is made in the moat. New-energy vehicles may not yet be very profitable, but batteries are one of their biggest moats, so battery companies earn first. Roboparty’s moat is “extremely low collaboration friction”: model companies do not have a platform, incumbent manufacturers struggle to open access down to the motor layer, and the stator and rotor URDF variables may not be separable. “The industry is missing an open platform for embodied AI,” and the company sits in a valuable missing layer.
The user-compounding assumption is straightforward: a CMU professor buys one platform and may bring in 2-3 PhDs to conduct research on it—“we earn RMB100K selling one platform, while gaining 2 engineers with RMB1M annual salaries” to help with R&D and publish papers. As with DeepSeek, open source does not prevent API sales. Hardware switching costs are higher than software switching costs, so open source itself is not the moat; the company’s fundamentals are.

10. The US-China Gap and the Bet Against Wheels

On the US-China divide, Kimi K3 “directly set off US equities” on the LLM side. The US still leads in the brain for embodied AI—or, more precisely, Chinese researchers in the US lead Chinese researchers in China. China is “far ahead” in humanoid platforms and the cerebellum/control stack, essentially by a generation. Figure and Tesla are active, but they also rely on Chinese supply chains. The gap in model-talent density is one reason Roboparty is considering building a US base.
His sharp bet against competitors is that wheels “have little significance; they are an intermediate state—equivalent to a legged system with infinite leg inertia.” Embodied AI is fundamentally about generality: “Why cut away part of that generality?” Whether end-effector precision is 0.1 or 0.01 millimeters becomes “less critical” as algorithms improve. To exceed humans in automation, the humanoid route does not need to build another intermediate state.
Will Zhipu or Zhiyuan open source? “Maybe, it’s possible.” The key variable is how tightly each company’s technology is locked up.

11. RP0: 100 Units a Month, and Every Customer in One Group Chat

RP0 ships as a kit and is positioned as an education product for going from zero to one. It has roughly 100 genuine orders a month, with about 30%-40% sold to schools—not only elite schools—and some sold to startups. The analogy is the early computer hobbyists who studied the 8086 and later produced Windows and Mac: “Maybe one of my users will become a Bill Gates.”
The counterintuitive commercial move is one he acknowledges may not work: “I think this doesn’t work commercially, but we may well do it.” In the group, user A asks a question and users B or C help answer it, sharply reducing support costs. It feels “more like a community and an ecosystem.”
A European institution once ordered 500 units at once but received only 20-30. The issue was not unwillingness to sell; the supply chain at the time lagged the company’s delivery capacity. Saying, “Other customers had to wait 2-3 months, and I gave him all 500,” he acknowledged the resulting delivery pressure. The supply chain is now gradually coming together and should improve.

12. RP One: From Engineering Prototype to Product, With the Launch as the “Opening Shot”

The upgrade is fundamentally about productization. The team studied 20-30 competing products and their problems: pinched fingers, overheating shoulder joints, insufficient wrist torque, fragile handles, difficulty fitting into a case and whether the robot could stand up autonomously from inside the case. It added 2 degrees of freedom to the head and upgraded the overall structure, internal cable routing and other details. The backdrop was the realization, while coming down from the peak of arrogance, that “others had already started tooling and pushing production capacity, while we were still making small batches of 50 or 100 robots.”
Demand signals emerged after the Behavioral Foundation Model took off and whole-body intelligence became the next step. ETH, Flexiv, several domestic startups and brain-model companies all approached the team for custom work: could the hand use a 48-to-24 structure to connect to a five-finger hand; could the full body use EtherCAT with hard synchronization; could head data and EtherCAT data be captured after the fact? “It gradually felt like we were beginning to set standards.” Customers’ real needs are the best PRD, while requirements from different customers remain confidential.
The product has been in development for some time but is still being tested; “some people in the industry have released products before even finishing their testing.” Timing depends on the back-to-school procurement season, IROS and WRC, when the industry begins competing on production and mass manufacturing, and whether “a highly threatening product is going to launch at the same time.” Competitors may launch before they are ready to seize attention: “Attention Is All You Need.”

13. The Endgame: Model Companies Are the Favorite, and Roboparty Plans to Enter Models in 2027

Asked who will ultimately build embodied AI in 10 years, he backs model companies. Firms such as Zhipu and OpenAI have enough GPUs, infrastructure, talent and researchers; “robotics cognition is actually cheap—you can fill it in by hiring people who really understand robotics.” OpenAI’s hiring of researchers such as 何泰然 is an example.
The host asked directly, “What about you? Don’t you believe in yourselves?” He did not dodge the answer: the endgame requires building a model, and the plan is to start around 2027. But pretraining scale may rise and then fall, as with video generation; whether to enter in the first wave and ride that cycle or cut directly into the second is a strategic decision. If OpenAI establishes OpenAI Robotics, its scaling ability would be far greater than that of any current Chinese startup and even far beyond a whole group of Chinese model companies.
For now, the company is using elimination. It will not do pretraining in the short term or build a very thick data pipeline; it may not care about small models with 1B, 2B or 3B parameters. The fundamental product is RP One: “Maybe a startup or a brain-model company will take RP One and build the data pipeline; we only need to work with them.” Research is driven by vibes, with researchers free to explore within a defined branch and costs kept manageable. Huawei once said it would not build cars, then explored a different path and is now one of the most profitable companies in the automotive sector. On the cerebellum side, the company has produced the DFM-0 upgrade UFO, Make-A-Light, object interaction based on BFM-0, and reinforcement-learning work on parkour, robot table tennis and basketball, all of which will eventually be open-sourced. On the brain side, InAct compresses actions into latent space, preserves a healthy latent distribution and avoids collapse; it won a like from Yann LeCun, and the team went to his company to give a talk.

14. The Honeycomb Organization and “Too Young” as a Shield

The organizational metaphor is a honeycomb. People are circles, and circles packed together form the most stable structure: a honeycomb. The ideal management span is 6—“nature tells me that one person can cover at most 6 faces.” The inspiration came from What Makes Anthropic Different. The company has only 2-3 layers, and “anyone can come directly to me.”
But the full-machine and supply-chain departments must be process-driven. If those teams were entirely talent-driven, the company might reach a point where “you don’t even know whom to call to buy a component today.” The answer is a combination of Huawei- or DJI-style extreme process discipline and ByteDance-style talent redundancy. Lab meeting systems and know-how exchanges can be reused in the operating company, while people who work across teams carry the culture; one pure-AI hire can lift the AI atmosphere across the company.
His self-assessment: “People think I’m quite friendly, but a friendly CEO is not necessarily a good thing.” The solution is not to change his personality, but to “hold my nose and hire someone with a completely different personality who often plays the bad cop.” The finance team is extremely sharp: “If they even approve your expense reimbursement, you’re finished.” Managing older people comes down to cognition: “If your cognition is broad enough, even a high-school student can manage.” Being “too young” is also a shield—at high-level commercial negotiations, youth gives him more freedom and directness.

15. Zero Voluntary Departures, Pay Cuts to Join and Burn Rate as the Sword of Damocles

There have been zero voluntary departures in a year and a half, excluding dismissals. The mechanism is a sense of gain. He likes hiring people with “many zeroes but missing a one”—people who performed well in another industry but lack foundational embodied-AI knowledge. Once they fill that gap, they could move next door for a 2x pay increase. “But nobody has left yet; maybe they feel they haven’t filled it enough.”
Executives earning RMB2.5M-RMB3M annually at major companies have been willing to join Roboparty for RMB800K, treating the difference as “tuition or a cost of transformation.” That is one reason the company can stay cost-efficient. Why not pay a premium if the company can afford it? “We can afford it, but we need to control it a little.”
The deeper reason is burn rate. Looking back at the early EV sector, well-funded companies that later blew up “basically all died from burn rate; they didn’t survive until new-energy vehicles truly reached scale.” Even the flowers at NIO House were replaced with artificial ones. 黄一 speculates that 李斌, who had cared deeply about brand image, could not have accepted the change. Scrimping on every fraction of a penny points to the key metric: capital utilization.

16. Cognition > Vision > Technology > Management > Effort: A 2030 Industry Map

He cites 唐杰’s hierarchy of “cognition over vision over technology over management,” then adds “over effort.” Management still ranks above effort. The most important form of cognition is industry taste: when the call is right, a founder should “take every bullet and go all in,” while knowing which companies to guard against, cover and respond to—and which can be ignored entirely.
His concrete forecast: more than 100 humanoid companies and 300 embodied-AI companies today may narrow to 10-15 survivors by 2030. The second tier could be formed by carmakers such as NIO, XPeng and Li Auto, plus Honor and Geely, with scale in the hundreds of billions of yuan; embodied AI could generate roughly 40% gross margins versus perhaps 20% in autos. If BYD, Huawei, CATL, BAT and Xiaomi enter next, “it will basically be a major era”: “BYD only needs to fill in a robotics division. Buy me, and it gets a robotics ticket.”
The scaling-law signal he cited was Jim Fan’s Sequoia presentation, including the robotics tracking-loss curve. The data figures in 徐南飞’s interviews were not fully consistent, but all were close to 100M hours: “Without 100M hours of data for training, it is unlikely that scaling laws will ever truly become visible.”

17. Marathon Mileage, Pirate Dreams and What It Means to “Make It”

On the marathon scale, the robotics platform is about one-quarter complete, with most system-level components gradually converging. The cerebellum is about halfway there, combining the VFM framework with object interaction—for example, recovering its original posture and continuing to pick up an object after being kicked while holding it. As for the brain: “I don’t know what the endgame looks like; that is the piece of cognition I’m missing.” If the endgame is a complete replication of human interaction, movement and force feedback, “we may have run only 1 or 2 kilometers today.”
In comparing US and Chinese cultures, he mentions Sunday and Generalist, admiring their focus and willingness to decide what not to do. Generalist, for example, essentially only builds models, not hardware, and is committed to data and model-free materials, scaling its data to 200K hours and potentially adding tens of thousands more each week. “In the US, that is what the era rewards. In China, the culture rewards people who can get a company off the ground.” He chooses the latter: “For now, I want to make Roboparty work.” His standard is concrete: his parents, without knowing he made it, genuinely choose to buy one of his robots.
Outside the company, he would like to become a pirate: rent a boat, drift without a destination and find a fish to name “Huang Yi Fish.” Once commercial spaceflight matures, he may “go up and take a look,” preserving a geek’s hands-on instinct for the day, 10 years later, when he understands the industry completely. He has no loyalty to tools: he codes with Cursor and Claude Code and switches models based on which is better; recently Kimi K3, previously Sonnet. The return on entrepreneurship is a compressed life: “It may not be physically happy, but psychologically and in terms of growth it is extremely happy—before 30, you get to live part of the life of someone who has reached 100, and there is a kind of wisdom in that.”

Robinhood CEO's Shocking Prediction On Stocks, The AI Bubble, & Gambling Controversy | Vlad Tenev Sam Altman on Building OpenAI & Betting on the Impossible