Summary

Last year, Xpeng made its biggest bet yet: it shut down its entire first-generation autonomous-driving system after investing several billion yuan and pivoted to a second-generation VLA rebuilt around a larger foundation model. 何小鹏’s view was that the old approach—“using software methodology and the AI toolkit”—could only produce more powerful software, what he calls an “AI Frankenstein,” and would never reach driverless operation or true robotic generalization: “What we want is a very smart car, but the methods we use will never make it infinitely smart.”
Physical AI is “100x harder” than digital AI, but 何小鹏 expects a ChatGPT-scale shift in physical AI in 2027-2028. Digital AI often amounts to “running models for a score”; in the physical world, that is “laughable.” A system must compete across its strengths, weaknesses and narrowest constraints—quality, cost, regulation and detail. “The CEO of a physical-world company either doesn’t dare to bet, or thinks, ‘What about all my other constraints?’ I may simply be bolder.”
He estimates Xpeng’s odds of winning in general-purpose humanoid robots at 20%—“already the highest odds I see among Chinese companies.” He also says 99.99% of companies pursuing general-purpose humanoids will fail, while differentiated robots may have “many possible solutions.” Starting a robotics company is 20-100x harder than starting an automaker, and the industry’s motion-control capabilities are still at the level of cars in the 1920s or 1930s. The Model T moment has not arrived.
The key commercial timeline is clear: robots should enter an automotive-style SOP phase by the end of this year; 2027 may well be the first year of commercial mass production for high-grade robots; L4 is likely 18-24 months away for Xpeng; and by 2030 China may have only 5 automakers operating at scale. On whether anyone has escaped the industry’s “red ocean,” his answer is blunt: “I don’t think anyone has.” Robotics, however, could move “from red ocean to blue ocean” quickly because most of the participants “don’t really understand AI.”
Iron was born through a radical reset: in 2023, Xpeng dissolved its 300-person robotics team and kept fewer than 60 people; after the launch event, when skeptics claimed a person was inside, 何小鹏 waited a few hours, lost patience, forced the team to come up with an idea overnight, and cut open Iron’s left leg in public.
GX is Xpeng’s first full-size 6-seat flagship SUV, bringing flying-car redundancy into the vehicle and making it “the first factory-installed robotaxi built by a Chinese automaker,” with 8 full safety redundancies. After Xpeng released the first version of its second-generation VLA at the end of March, industry sales fell roughly 20% both year on year and month on month in April, while Xpeng rose 50-70%. The core argument is that software and hardware could each represent 50% of vehicle value over the next decade: users may eventually pay RMB150K for software capability in a RMB300K car. “That doesn’t exist today, but I think it will change within this decade.”

Deep dive

1. The CEO Shouldn’t Use AI Products Too Deeply—“Use Them, But Don’t Go Too Deep”

何小鹏 says he still uses mostly “very traditional AI products,” including Qwen and Doubao. His team uses coding tools extensively, but he personally refuses to use them deeply. The lesson comes from the internet-product era: “If you use a product every day, you quickly get pulled into the details … and that stops you from looking toward the distance.” The CEO needs to look ahead; frontline teams closely tied to the product can “try every method available, judge by the results, and converge gradually.”
His view of AI coding is deliberately restrained: it is “a very good assistive tool for junior programmers.” In 2-3 years, it will push junior programmers toward senior-level work. But for AI-heavy capabilities such as intelligent assisted driving, the help is limited. When writing an operating-system kernel, “the core is still the algorithm as a whole, not the coding.”

2. Tokens Are Not the North Star of the Physical World—Data Is What Really Burns Cash

何小鹏 rejects token volume as a measure of how AI-native a company is: “We don’t really look at that.” In the physical world, the token consumer is the machine itself. A new-generation VLA running for 3-4 hours a day can consume far more tokens in its inner loop than employees using digital AI. “The number of tokens used by people with AI is far lower than the number physical-world AI itself needs.” The internal north-star question is instead: if given 30,000 or 50,000 H100s, which businesses can use them efficiently?
Internally, Xpeng does not cap token spending: “You don’t know whether RMB1,000 or RMB10,000 a month per person creates the most value, and that amount is probably far below the person’s monthly salary. If someone can spend more and generate more value, why limit them?” Management focuses only on the most extreme outliers.
The real outlier is data. “Very few companies today see the enormous cost of data.” Digital AI can be trained on tens of TB; physical AI may need tens to hundreds of TB per training run. Xpeng’s fixed annual data spending is “close to or above RMB1B,” with each tier of storage and usage strategy costing tens of millions of yuan.

3. Can 何小鹏 Be Turned Into a Skill? The CEO May Be Replaceable, but No One Can Judge Right or Wrong

His direct answer: “Perhaps in several decades or 100 years, capabilities like mine really can be turned into skills.” But by then, CEOs will have “lost A, B and C and acquired new D, E and F.” Asked about the weaknesses of a 何小鹏 skill, he responds self-deprecatingly: “Once everyone is turned into a skill, you’ll discover that the person actually has a great many flaws.”
The deeper contradiction emerged during the robotics work. In coding and autonomous driving, “I know what is wrong,” so the system can be strengthened. But “if you turn 何小鹏 into a skill, it is hard to judge how that skill is right or wrong.” That is why Xpeng is “building a system, rather than using one built by someone else.”
His benchmark for corporate AI intensity: companies in the digital world can reach “the low tens of percent,” but for a physical-world company with tens of thousands of employees, general-purpose AI—including autonomous driving and robotics—at 15%-20% is “enough.”

4. From an “AI Car Company” to a “Physical AI Company”—The Verdict on Frankenstein

The renaming of Xpeng Group reflects a reassessment of the past decade’s “smart electric vehicle” formula. His industry verdict is unsparing: from Toyota, Google and Baidu to Tesla and Xpeng, autonomous assisted driving has achieved results but not enough altitude. At its core, it is “a combination of AI algorithms and software rules. I call it an AI Frankenstein.”
The evidence is concrete: “There isn’t a single autonomous-driving company’s software today that can drive smoothly in an underground parking garage.” Parking-lot assistance is essentially memory-based—drive the route once and remember it. Its understanding of the physical world is “extremely limited”; the core problem is insufficient intelligence, which puts a low ceiling on the system.

5. The Big Gamble: Shutting Down a Stack That Cost Several Billion Yuan

Last year Xpeng was building 2 generations of autonomous assisted driving in parallel. The first amplified end-to-end driving, reduced rules and added post-training. The second discarded end-to-end altogether, using a larger foundation model to “open up the ceiling first, then converge on the floor.” His numerical analogy: the old route had a ceiling of 1,000 points and a floor of 900; the new one might have “a ceiling of 100,000 to 1M points, but at that stage a floor of only 100 points.” The engineering problems were numerous, but he bet anyway: “We shut down the entire old system. It had cost several billion yuan.”
There was no landmark conference or formal trigger. He says the decision was made “basically in my head,” followed by total commitment. In March last year, he kept returning to one question: every company believed a simpler methodology could produce “usable enough” autonomous driving, but after 1-2 years, solving weaknesses would constrain its strengths. That meant it “could never reach Level 4 or Level 5. This is a shortcut, but not a broad avenue.”
By the end of Q3 last year, the decision had reached the organization: “We completely changed the core organizational structure of the autonomous-driving center.” Every window has talented people, but everyone carries inertia—the habit of using old methods and the latest tools to build better versions of the old thing. “Most of the time that is right, but sometimes your entire way of working has to change.”

6. Why Physical-World CEOs Don’t Dare to Bet—Strengths, Weaknesses and Narrow Constraints

His core methodological distinction is that digital AI uses human language: “Language is the world.” But in the physical world, the volume of data each person sees every day cannot be summarized, described, reconstructed or replicated in language. It is simply too large. Automotive CEOs can learn digital AI and map it onto their businesses, “but at that point no CEO will say, ‘I want to use it for everything.’”
His recurring framework is a theory of the board: the digital world is mainly about the strongest suit; the physical world requires “broadening the narrowest constraint, extending the weaknesses, and extending the strengths further,” while protecting the floor and competing across quality, cost, materials and regulation. So “the CEO of a physical-world company either doesn’t dare to bet.” He admits that when he placed the bet, “I didn’t know whether it could solve the weaknesses.”
His refusal to disclose the details is itself a signal. The specific transformation steps “cannot be discussed, sorry.” Any digital-AI CEO would do the same: “First, you don’t know whether it will succeed; second, you won’t share it.” The isolation of the decision-maker is equally explicit: “It is hard to find someone who can discuss it with you accurately … most of them are wrong because they are not looking at the whole picture. The whole picture is yours to control.” In the end, he still had to place the bet, just as in entrepreneurship.

7. Software and Hardware 50:50—A Value Reset for the Next Decade

The most painful part of building a smart electric vehicle is that, however you look at it, hardware accounts for far more than 50% of what a customer needs. His view is that software and hardware could each account for 50% of value over the next decade: “A user may say, I bought a RMB300K car—RMB150K is your hardware, and RMB150K is the combined capability of your software. Does such a car exist? Not today, but I think it will change within this decade.”
There is only one path to that outcome: “Use AI to rewire your organization and design the integration of your software and hardware.” He takes a pure-utility approach to digital AI—“Cloud Code is very good, so I will use Cloud Code; if DeepSeek 4 is better in certain areas and the cost is excellent, I will switch those parts over”—with an internal hub that allows rapid switching. But physical-AI models, data and algorithms are different. “I won’t say more.”

8. Cutting Into the Organization’s Roots—Voting With Their Feet and “Using a Small Knife on a Big Tree”

Resistance to the bet was real. Non-AI managers thought “whether we do A or B, it may both be wrong.” Most AI managers were somewhere in the middle, “including me, but that’s the fun of entrepreneurship.” The strongest opposition came from people “voting with their feet”: they did not believe it was possible, so they left.
His organizational philosophy is worth preserving verbatim: “Remember, don’t use a small knife to cut down a big tree”—nibbling away at it slowly. Once the decision is clear, cut it out. “Change everything, from the organization to the processes to the direction.” Outsiders see only the application layer; internally, the change goes all the way to the roots.
His scale-of-bet framework is straightforward. At the start of a company, the bet is 10x to 1,000x: “A 1,000x bet has a one-in-1,000 chance; a 10x bet may have a one-in-10 chance.” As the company grows, it still has to bet, but perhaps between several times and 10x. Two principles he gave himself at the end of 2022 still apply: “never give up” and “accept the outcome of the bet.” “The more you hesitate, wait, and say, ‘Let’s talk again in 6 months,’ the harder it may be to succeed.”

9. AI’s 4 Directions—and the Verdict on Waymo

何小鹏 sees 4 directions for AI exploration: the transformation of digital AI, where many people are already working; physical AI, which is “100x harder than digital AI” but will show results beginning in 2027-2028, with ChatGPT-like disruption at the level of individual jobs; human AI in medicine and eldercare, where nobody except Google has truly integrated the pieces; and the coupling of enterprises with AI, which smaller companies with fewer than 1,000 employees are beginning to explore.
His judgment on Waymo is among the sharpest in the interview. Google bet on autonomous driving in 2009, 17 years ago. Waymo is “both a good and a very bad solution”: technically capable, but difficult to globalize. “Level 4 technology is inherently a higher-order AI Frankenstein.” Even making driving easier and safer is a problem no company has solved exceptionally well after 17 years.
He also pours cold water on small AI startups. Today AI can help with 3 things at most: a basic analyst, a basic programmer and a basic designer. It remains mainly assistive, has not replaced them, and has not produced massive-scale collaboration effects. A few people claiming AI coding will overturn the physical world are “completely underestimating the diversity and complexity of change in the physical world.” “You can take the digital dimension to its extreme, and its value point may still be tiny.”

10. Anxiety in an Era of Upheaval—Even the Axioms Are Being Rewritten

Asked about anxiety, he gives an unguarded answer: “Of course I’m anxious. You discover that many of the logical arguments you relied on before—even the underlying theorems, paradigms and axioms, extending all the way to your values, view of life and worldview—no longer work as well under these changes.”
His example of an axiom being overturned is autonomous driving. The longer you work on it, “the more you feel that Level 5 may never arrive,” because software’s response to global scenarios, regulations and human-machine interaction is an infinite loop. “But once you truly rebuild the problem with AI, you feel it may actually exist.” That has generated new thinking about many former sources of defensibility. “I can only go this far.”
He adds a hierarchy of capabilities: seeing a problem, solving a problem, and building a system that has a ceiling while plugging the floor to prevent problems are “3 completely different levels of ability.” Even as a CEO, you cannot solve everything, because the problem you see is only one angle or link in the world’s problem.

11. 3 Stages of Robotics: Dissolving a 300-Person Team and Choosing a “Stitched-Together” Leader

Xpeng’s robotics effort has 3 stages. From 2018-2020, it acquired an independent quadruped team. From 2020-2023, over 3.5 years, it built 3 different models—integrating them with machine methods, automotive methods and a stitched-together approach—with varying successes and failures. After 2023, it committed to moving from quadrupeds to bipedal robots and from disbelief in the brain to “a new design in which the brain drives the robot.” His jab at the industry’s claim that its “cerebellum” is finished: moving forward with the same monotonous gait is not a cerebellum; “that is the spine, or your brainstem.”
In 2023, management concluded that a team with deep robotics expertise still could not build a good new-generation machine. It kept fewer than 60 of the 300 people. “I hear they went out and started 10 companies, and most have raised money”—one company, whose name sounds approximately like “中情” in the audio, has raised multiple rounds. The new leader, LC, “knows a bit about AI, a bit about cars, a bit about engineering and a bit about robotics.” Why choose him? “A lot of the time it is fate. His quadrant aligns fairly closely with the quadrant I think in.”
The distinction in their talent philosophies is telling: LC talks about “talent density”; 何小鹏 talks about “talent potential.” From the end of last year through the first half of this year, one department alone hired nearly 80 graduates with PhD-level credentials from China’s top universities. “They are expensive, but we are willing. Super-smart people should work on super-difficult problems.” LC is even crazier: “He keeps telling me that he wants to create people, not robots.”

12. Why Robots Have to Look Human—Robot Dogs, 1.8-Meter Armor and the Uncanny Valley

Why take the hardest route of a general-purpose humanoid? He offers 2 examples. First, a robot dog entering a bedroom: a 1.1-1.2-meter quadruped cannot turn around in place beside a bedside table. When a golden retriever turns, its tail may brush the wall and bed, but you would not think anyone could be hurt. “With a robot dog, both you and the dog would feel 100% at risk.” Make it smaller and its capability is extremely limited—little more than light companionship—and its range is short. Second, a 1.8-meter biped covered in armor: “Even if you designed it, you would not want to walk 1 centimeter away from it.” It could be hot, electrically charged, dirty and dangerous. If adults feel that way, what about older people and children?
That is why Iron’s design constraints are so specific: it is 1.69-1.70 meters tall, comfortable for both men and women, can wear clothes and may even have hair. “But it cannot have its own face.” The uncanny-valley effect, along with legal and sociological considerations, means it “must retain a certain difference from humans.” He stresses that this is already a middle ground: “If you want to bring robots into human society in a truly general-purpose way, we are already choosing the easiest middle state.”
He has reordered the ultimate value proposition of robots. It used to be emotional value; in the future it will be “a combination of physical value and emotional value”—the ability to get work done. Survey data points the same way: young people expect robots to enter the home and do chores, while middle-aged men and women around his age or older are often asking what a robot will be useful for when they grow old. “Older people may well use a robot as their only source of support. Very few people think about it that way.”

13. That Accident: He Couldn’t Wait 24 Hours and Cut Open the Left Leg

After the launch event, some people claimed that Iron was a person in disguise. The team split into 2 camps: one argued that “the innocent need not defend themselves”; the other wanted to observe for 24 hours. 何小鹏’s version is blunt: “I waited a few hours, and I couldn’t take it anymore … after 24 hours, the bullet could already be anywhere.” That night he called the team: “By tomorrow morning, come up with a concept. I want to tell everyone this is a real robot.” He forced the idea out of them.
The left leg was chosen for an engineer’s reason: Iron walks from left to right, so “the left leg is easiest for everyone to see.” He also acknowledged that the Iron on stage was only an intermediate milestone, with thermal-management problems—it “was very hot.” The real issue was not public skepticism but “how to get everyone internally aligned,” because many people felt there was no need to explain anything, while “99.99% of people in this world have no way to touch it.”
His postmortem on the episode is that the net effect was acceleration. Many outstanding people in robotics joined Xpeng, “raising our odds of success a little.”

14. Motion Control Is Still Before the Model T—Robotics Is the Automobile Industry of the 1920s

The technical judgment beneath the humanoid form is that most people underestimate robotic motion control. Much of today’s robotics practice traces back to certain open-source MPC approaches from 2018, which is why the systems have so many similarities. Automotive motion control is better, but it is still domain-specific, single-point control; “as a whole, there is no full coordination.” His example: if an autonomous-driving car has snow under its left tire, grass under its right tire, and needs to turn 47.5 degrees, how should it turn? “That is hard even for a person.”
His historical comparison is direct: “Robotics is still perhaps the automobile industry of the 1930s or 1920s—maybe the Model T era. Does today’s Model T exist? In robotics, definitely not.” Humans have more than 200 joints, along with muscles and skin. “If you added muscles and skin to many robots today, their walking performance would change and degrade dramatically.” Current robots either walk or fight. “You need 5 types of motion control? There aren’t any.”
Xpeng is targeting “full-posture, fully AI-combined motion control,” closer to human instinct. When a person sees snow and grass on the ground, they take a probing step to feel the friction. That is why “most robots can only stitch things together with software,” making strong combined capabilities difficult.

15. The Odds in Robotics: 99.99% of General-Purpose Humanoids Will Die; We Have 20%

More than 300 companies once entered auto manufacturing—“I personally think it was around 100”—while robotics companies reportedly number more than 200 already. But the categories are fundamentally different. Robotics has countless segments, from medical care to freight and inspection, and “many robots do not need to be humanoid.” His conclusion: “99.99% of companies pursuing general-purpose humanoids will die. But for differentiated robots, there may be many possible solutions in this world.” 张小珺 believes the overall odds in robotics are higher than in passenger cars.
His own odds: “We have roughly 20%. That is already the highest probability I personally see among Chinese companies.” The difficulty benchmark is stark: “Starting a robotics company is probably 20-100x harder than starting an automaker. Notice that I gave you a floor of 20x.” The pitfalls cannot be avoided through understanding alone; “you only understand the entire logic after you actually step into them.” He is keeping his forecast for the competitive landscape private: wait 3-4 years, after many companies have stepped into those traps, and the shape of competition will become clear.
Who is the competitor in general-purpose humanoids? “There is no competitor right now. Everyone is competing with themselves.” Each company must maximize its own capabilities, from the foundational organization and infrastructure to technology, product, commercialization and engineering.

16. Robotics Across 3 Curves—Scale Will Come Much Faster Than in Cars

Xpeng’s 3 curves for the new decade are turning the car into a fully intelligent agent, building robots—which “are agents in their own right”—and going global. The key distinction is that cars may need 10 years to prove software can account for 50% of value, while “robots don’t need to prove it. If the software is very poor, you basically won’t want the robot.”
One point is widely overlooked: once robots can scale, “the speed of scaling will far exceed that of cars.” Cars are constrained by roads, regulation and production ramps: “If you sell 100,000 units today, sorry, you cannot produce 100,000 units next month.” Robots could explode at a particular point, much as digital AI’s biggest shift last year was “SOP Big coding,” which changed dramatically in less than 18 months. Xpeng’s approach is 80% in-house hardware development—hands, chips and joints included. “We are Tier One and work with more upstream Tier Two suppliers. Only this can deliver quality and scale.”
Before mass production, 3 hurdles remain: hardware reliability and stability; whether multiple high-level models can be fitted together effectively; and whether commercialization can be proven. “Robotics has not yet been clearly defined. The market is waiting for something like an iPhone 4, but I believe the first commercially mass-produced robot will not yet reach the level of the iPhone One.” The analogy is also imperfect: when the iPhone launched, Nokia and Motorola were available to replace; when robots arrive, “many people have not built anything similar.”

17. GX: Stitching Flying-Car and Robotic Capabilities Into a Flagship SUV

GX is Xpeng’s next push upmarket: its first full-size 6-seat flagship SUV, differentiated by “connecting many of the capabilities of cars, flying cars and robots.” Redundant core components from flying cars are being brought into the vehicle. It is “the first factory-installed robotaxi built by a Chinese automaker, with 8 full safety redundancies”: it can still drive if the power supply fails, the drivetrain fails or a mouse chews through the wiring harness—“like an aircraft.” Its steer-by-wire chassis, new EEA and VLA are integrated, raising safety, cutting latency and improving control responsiveness by close to several dozen percent. Robots’ task-oriented reasoning will also flow into the car. The third row folds completely flat; features include Fuyao privacy glass and Midea’s next-generation onboard refrigerator. The formal launch and pricing are set for May 21.
His verdict on the 30 full-size models in the same segment is harsh: “The exterior looks big, the interior is crude, the details are not user-oriented, and the cars are stitched together from many capabilities.” In this war, “your weakest point also has to be an 80, with many more 90s and 95s, before you can win.”
Will GX repeat the G9 misstep? “I’m not worried.” The postmortem produced “too many” changes across organization, product planning, customer understanding and commercial logic. “The stronger your system capabilities, the harder it is to fail easily on one or two points.” On Li Auto and NIO, he is gracious: he had just seen Li Auto’s new L9 and 李斌’s ES9, and “魏小李” will each offer different views and interpretations of the large-9 segment this time. The target customer is “over 30, relatively brave, willing to explore new technology and luxury.”

18. L4 in 18-24 Months, 余凯’s Path and the End State of 5 Automakers

The L4 timetable is “most likely 18-24 months”—for Xpeng—and “anyone who says it can be done this year is bullshitting.” He immediately undercuts the long-term significance: even achieving it “does not represent long-term value.” The current evidence is more concrete: after Xpeng launched the first version of its second-generation VLA at the end of March, industry sales fell roughly 20% both year on year and month on month in April, while Xpeng rose 50-70%. “A considerable part” of that was related to the second-generation VLA.
On Horizon’s 余凯 arguing that automakers should ultimately use third parties, 何小鹏 makes the answer conditional: “The more automotive and robotics companies there are, the broader 余凯’s path; the fewer there are, the more painful it will be. I may believe the industry will become increasingly concentrated.” His test for in-house development is simple: “If it is tactical, you should not develop it yourself; if it is strategic, you should.” Cross-domain integration is why Tier Ones are hard to replace: “Providing one total solution with 3 companies is easy; providing 300 personalized solutions is extremely painful.”
His end-state forecast is unchanged: “By 2030, China may have only 5 automakers with scale.” That does not mean the others will go bankrupt, but shrinking scale will make it difficult for them to enter ultra-high-intensity competition. On the red-ocean question, his answer remains: “I think everyone is still swimming … I don’t think any company has made it out.” Robotics will be different: homogeneous competition may be less intense, software value is substantial, and he does not believe a sufficiently good open-source solution exists to help a robotics company build its software. “Robotics’ red ocean, like the internet competition of that era, will quickly become a blue ocean.”
In the closing personal portrait, he says most of his time goes to “strategy and planning.” He does not read much: “By the time a book is written and printed, the world may have changed by another 10%.” He absorbs knowledge through practice. Within an AI system, “you can’t even do the C in PDCA anymore; you should minimize the consumption of any human step as much as possible.” Did he regret any decisions over the past year? “No. Why regret making mistakes? We think together about why we were wrong now, but there is no need for regret.”

Verification Notes

The raw captions render one startup name approximately as “中情”; its identity cannot be resolved from the captions.
The raw captions render a coding product as “SOP Big”; there is no supported resolution to “Claude Code”.

167: Onion Academy’s 杨临风: Using AI to Manufacture Shortcuts Is Killing Real Learning Finding Alpha in People