何小鹏 | 张小珺Jùn｜商业访谈录 Summary

Podcast: 张小珺Jùn｜商业访谈录 | Episode 143 Participants: He Xiaopeng (Chairman & CEO, XPENG/Xiaopeng Group), Host Xiao Jun

1. Key Themes

The "Stitched Monster" Problem: Why Existing Autonomous Driving Approaches Hit a Ceiling

He Xiaopeng made a pivotal internal decision to scrap a multi-hundred-million yuan autonomous driving system because it was architecturally incapable of reaching true autonomy. The old approach combined software engineering rules with AI tools — what he calls a "AI stitched monster" (AI缝合怪) — which could be optimized but never fundamentally transcended.

"You're using software methodology, using the AI toolbox, and what you produce is still software. I call it the 'AI stitched monster'... So last year we made a massive transformation and bet — we completely stopped the old system. That system cost us several tens of billions of RMB." 00:17:06

The new paradigm (VLA - Vision-Language-Action model) has a theoretical ceiling of 1,000,000 points vs. 1,000 for the old system, though it initially had a much lower floor. He made the bet to invest in raising the floor on the higher-ceiling approach.

"I believe VLA's upper limit can possibly reach 100,000 to 1,000,000 points. But at that time its lower limit was also terrible — where you hoped to build a product with an upper limit of 1,000 and lower limit of 900, its lower limit was maybe only 100." 00:16:20

Physical AI vs. Digital AI: A Completely Different Paradigm Most Companies Are Getting Wrong

He draws a sharp distinction between digital AI (language, text-based) and physical AI (robots, autonomous vehicles). He argues that most companies trying to enter physical AI are applying digital AI thinking — which is fundamentally mismatched.

"Digital AI is human language, which is highly summarized and condensed. Language equals the world. But in the physical world, the amount of data each of us perceives every day cannot possibly be described, restored, or replicated in language. It's too massive." 00:18:54

"Many people talk about Physical AI and I just laugh. I ask: have you ever actually run a business in the physical world? If you haven't successfully built a business in the physical world and you're only analyzing physical world definitions from a digital world perspective to build physical AI — I think that's a very shallow physical world." 00:29:33

The 50/50 Hardware-Software Value Split: The North Star for the Next Decade

He articulates a specific, measurable thesis: in the current era, hardware represents >50% of a car's perceived value. The goal of XPeng's entire physical AI strategy is to create products where software accounts for 50% of what a customer is willing to pay for.

"I believe that in the next era, hardware and software together might each account for 50%. So you must use AI methods to re-drive your organization, design your hardware-software integration, and capture greater AI value. Only then can software's combined performance — the value users are willing to pay for — reach 50% in new eras." 00:21:12

"Has there ever been a car where a customer buys a 300,000 RMB vehicle and considers 150,000 RMB to be the hardware and 150,000 RMB to be the software's combined capabilities? Not yet. But I believe this will change within this decade." 00:23:52

2. Contrarian Perspectives

The Token Metric Is Largely Irrelevant for Physical AI Companies

While the tech industry obsesses over token consumption as a measure of AI adoption, He argues this metric is essentially meaningless for companies operating in the physical world. Machine-to-machine token usage in autonomous systems dwarfs human-facing token usage, and the two shouldn't even be compared.

"The amount of tokens usable by digital AI is far less than the tokens that physical world AI itself needs to use. But this is actually meaningless, because it's not using humans — it's using AI to satisfy its own needs. An autonomous vehicle is an automated machine. How many tokens it uses to operate — this belongs to a completely different dimension." 00:05:28

"I see many companies talking about how many tokens their engineers use. I think that number, compared to the tokens machines use and the tokens needed to train physical AI models, is a very small number." 00:08:49

Most Humanoid Robot Companies Will Fail — But the Category Has Higher Survival Odds Than Cars

He gives humanoid robots a survival rate among companies of 0.01% for those pursuing general-purpose humanoid robots specifically. But he argues the robot category broadly has better odds than passenger EVs because the segmentation is far more diverse.

"If you pursue general-purpose humanoid robots, 99.99% will die. But for various differentiated robots, I think there are many possible solutions in this world." 01:00:01

"I believe robot difficulty is approximately 20 to 100 times harder than starting a car company. I still gave a minimum of 20x. And I already think XPeng's success probability in robotics is about 20% — and I consider that the highest probability I've seen among Chinese companies." [00:58:20 & 01:08:23]

AI-Coding Is Not Strategically Important for Physical AI Companies

Contrary to the dominant narrative that AI coding tools are transforming every company, He argues that for physical AI companies, coding tools help at the application layer but are irrelevant at the infrastructure level where the real battles are fought.

"For autonomous driving or other strong AI capabilities, I think the help [from AI coding] is relatively small. It's just one of the tools. What really needs to be built well is the entire infra, the entire system. If you want to build the core system... the most important thing is still the entire infra, not coding." 00:04:07

The CEO Should Deliberately Not Use AI Products Too Deeply

He makes a counterintuitive argument that the #1 leader should avoid deep personal usage of AI products — because deep usage pulls you into solving near-term defects rather than thinking about long-term direction.

"When we were building internet products, if you used the product every day, you'd very quickly get pulled into the details. You'd notice where the scale isn't working well. Once you're in it, you focus on solving the problems — like multi-person collaboration in coding — which actually prevents you from looking toward the distance." 00:02:37

Level 4 Autonomous Driving Within 18–24 Months — But Its Value Won't Be Immediately Obvious

He predicts L4 for XPeng within 18–24 months, but immediately cautions that even achieving it won't translate into proportionate business value right away. This challenges both the skeptics who say L4 is decades away AND the optimists who think it will be an instant commercial windfall.

"I roughly believe 18 to 24 months [for L4 for us]... Even if sales increase more, that doesn't mean long-term value. Even if sales grew, by how much — 1x, 5x, 10x — I find it difficult to judge whether that's good or not." [01:19:19 & 01:20:27]

3. Companies Identified

XPeng / Xiaopeng Group (小鹏集团) Formerly XPENG Motors, recently rebranded. Chinese EV and now physical AI company developing autonomous vehicles, flying cars (eVTOL), and humanoid robots. He highlighted that the company scrapped tens of billions of RMB in prior autonomous driving systems to pivot to a VLA (Vision-Language-Action) architecture, achieved 50–70% sales growth in April 2024 when the broader market fell ~20%, and is pursuing full hardware vertical integration for robotics (including chips and joints).

"In April, China's auto sales declined about 20% month-over-month and year-over-year. We grew roughly 50% to 70%. A significant portion of this is related to the second-generation VLA." 01:19:58

Waymo (Wéimò / 威摩) Google's autonomous driving subsidiary, operating robotaxis in the US.

"I think its technical capability is decent, but I think it will be very difficult to globalize. In the time it's taken, their technology is inherently a higher-level stitched monster. It's very difficult to achieve extreme generalization in AI. I don't think it's an extremely intelligent methodology." 00:33:03

Horizon Robotics (地平线) Chinese automotive AI chip and solution company. He acknowledges them as a well-run company with good products but questions whether their third-party supplier model has a sufficient strategic moat as the industry consolidates.

"Yu Kai [CEO of Horizon] and I have a very good relationship. I think he's walking a very interesting, very promising, but perhaps very challenging road. It depends on whether the companies chasing automotive and robot AI are increasing or decreasing. The more there are, the broader Yu Kai's road; the fewer, the more painful." 01:20:40

Tesla Referenced as the global benchmark for physical AI integration, particularly regarding Elon Musk's philosophy of hiring generalist talent.

"I think in Tesla's Silicon Valley, it's easier to find more generalist candidates. I think in most countries, cities, or industries, it's very difficult." 00:47:05

NIO (蔚来) He visited NIO's Li Bin and viewed the ES9 before the Beijing Auto Show, acknowledging the competitive landscape among China's top three EV makers.

"Yesterday I went to Li Bin's place and looked at the ES9. I think NIO, XPeng, and Li Auto (蔚小理) in the large SUV segment will all offer different perspectives and insights this time." 01:15:19

Li Auto (理想) Direct competitor in the large SUV segment with the L9. He visited their booth and acknowledged the competitive quality.

"Today I went to Li Auto's booth to look at the new L9 — it's also a very good car, I believe." 01:14:56

4. People Identified

He Xiaopeng (何小鹏) Chairman and CEO of XPENG Group. Former co-founder of UC Browser (acquired by Alibaba). Known for making massive, costly bets — scrapping tens of billions in prior systems — and for a long-term physical AI vision. He self-assesses his humanoid robot success probability at ~20%, which he considers the highest among Chinese companies in the space.

"I may have a bigger risk appetite. We made a massive bet. We stopped the entire previous system. That system had cost us several tens of billions of RMB." 00:00:59

LC (Robot Division Head, unnamed) The head of XPeng's humanoid robot division. He Xiaopeng selected him despite (or because of) him being neither a pure robotics expert nor a pure autonomous driving expert — he had cross-domain exposure. He describes LC as "even crazier" than himself.

"LC is even crazier than me. He always tells me he wants to 'create humans,' not robots. The vast majority of people think of a robot as a commercial product, but I think he considered the robot's ultimate participation in society, in the universe, and its emotional connection with each of us." 00:48:21 "His vector of thinking was fairly aligned with mine. He was one of the few people I selected within my thinking framework." 00:45:36

Yu Kai (余凯) / Robin Yu CEO and co-founder of Horizon Robotics. Described as a friend and thoughtful industry player, but He questions the long-term strategic ceiling of the third-party automotive AI supplier model.

"Yu Kai is walking a very interesting, very promising, but perhaps very challenging road." 01:20:40

5. Operating Insights

Don't Manage Employees by Token Limits — Manage Only the Top 10 Anomalies

Rather than setting token budgets per employee (a common practice He dismisses), He's approach is to leave usage open and only intervene on the top 10 anomalous cases. The underlying logic: if an employee's monthly token cost is less than their monthly salary, restricting usage is irrational if it creates value.

"Internally I try not to control everyone's token usage. Many people ask: if you don't control tokens, one quarter's spending could equal a full year's budget. I think if it truly delivers value, what matters most is managing the top 10 most anomalous cases. Everything else I think should be open. You don't know if a person spending 1,000 RMB or 10,000 RMB per month in tokens is more valuable — and each person's monthly salary very likely far exceeds that amount. If they truly have the ability to spend more and generate greater value, why restrict them?" 00:06:36

When Making a Major Pivot, Cut Fast and Deep — "Don't Use a Small Knife to Chop a Big Tree"

He describes his internal philosophy for organizational transformation: once you've decided on a direction, restructure completely — organization, process, and people — rather than incrementally. Prolonged indecision is more costly than a decisive wrong call.

"In an organization, never use a small knife to slowly chip away at a big tree. Think it through, then cut it. So in certain areas, dare to bet — from the organization, to the process, to the direction. Change everything." 00:35:31 "The more you hesitate, the more you wait, the more you want to observe, the more you say 'let's revisit in six months' — the harder it becomes to succeed." 00:28:33

For Data in Physical AI, the Cost Side Is Massively Underestimated — Build a Dedicated Data Management Team

He revealed that XPeng spends close to 1 billion RMB per year on data storage and management alone for autonomous driving training, with separate dedicated teams for data and compute optimization. The insight: data management in physical AI is not an IT function — it's a strategic cost center requiring specialized teams.

"Very few companies today see the enormous cost of data. In digital AI, data volumes are small — a few tens of terabytes can be used for training. When we train once, it's tens to hundreds of terabytes. So how you manage, use, and store data is enormous money. We may spend close to 1 billion RMB per year in direct hard costs on data alone. Which data is valuable, which is temporarily valuable, which needs to be accessed very fast, which can have some warm-up — each category costs tens of millions of RMB." 00:07:50

6. Overlooked Insights

Humanoid Robots Have a Potential Explosive Adoption Curve That Could Exceed Automotive — and Almost No One Is Modeling For It

He briefly mentioned that once humanoid robot software capability reaches a certain threshold, the hardware scaling speed could dramatically outpace the automotive industry because robots don't require road infrastructure, regulatory buildout, or the slow factory ramp-up constraints that cars have. He compared it to the speed of Cursor/Copilot's adoption in coding — from near-zero to dominant in under 18 months.

This is a sleeper insight with enormous investment implications: the standard modeling for humanoid robot adoption assumes a gradual automotive-like S-curve, but He is suggesting it could be a near-vertical adoption spike once a quality threshold is crossed — creating a winner-take-most dynamic that investors may be severely underweighting.

"Robots might scale much faster than cars once the software capability is there. Cars took over 100 years to gradually scale because roads needed to be built, traffic regulations needed to be established, and mass production difficulty is very high... But robots are different. If done particularly well, at a certain single point — remember, a point not a plane — it's possible there could be a very large explosive effect. Just like what I think was the biggest change in digital AI last year: Cursor's coding. In a very short period, roughly 18 months... a massive change. I think robots have a similar possibility." 01:05:58

XPeng's GX Integrates eVTOL Redundancy Systems Into a Car — A Regulatory and Safety Moat Nobody Is Talking About

In passing, He mentioned that the GX SUV features eight full-safety redundant systems (冗余) derived from their flying car division — meaning the car continues operating even if the power source, drivetrain, or wiring harness fails. He positioned this as an industry first for a production vehicle. This is not just a feature — it's a potential regulatory moat as autonomous driving regulations tighten globally, since redundancy requirements for L3/L4 systems will likely mandate exactly this kind of architecture.

"We brought the core redundancy components from our flying car into this vehicle. This GX is, I believe, the first OEM in all of China to offer factory-installed full redundancy — eight full-safety redundancy systems. So if you drive into the wilderness and the power fails, you can still drive. If the drivetrain fails, you can still drive. If a mouse chews through the wiring, you can still drive. It works like an aircraft — it allows you to have redundancy." 01:10:22