He Xiaopeng in Conversation with Xiaojun Zhang: Physical AI Is 100 Times Harder Than Digital AI | Xingxin PORTFOLIO

"So last year we made a massive transformation, a huge bet. We shut down our previous system entirely. That system had cost us several billion yuan."

— He Xiaopeng

In March or April 2025, He Xiaopeng made a decision in his head: stop the old system that had cost "several billion yuan" — what he called an "AI Frankenstein" — and pivot completely to a physical AI R&D path.

This wasn't an ordinary strategic adjustment. It was a CEO's "self-revolution." He gradually realized that what they'd been stitching together with software methodology plus an AI toolkit had hit a hard ceiling.

WeChat Channels link: https://weixin.qq.com/sph/AWdlq4bFbm

Below is a curated and condensed version by Bijixia of Xiaojun Zhang's business interview with He Xiaopeng:

Philosophy on Using AI

Don't Let the #1 Person Become Too Deep a User of AI Products

Lately I haven't used that many AI products — still the very traditional Qwen and Doubao. But for coding, our team uses them heavily internally. Personally, I refuse to.

Why? I once gave a very interesting example. Back when we were building internet products, if you used the product every day, you'd quickly get lost in the details. You'd think this doesn't work well, that's a problem.

But if you keep your distance and learn about it from afar, you think about how its strengths will remain powerful and plentiful in the future. Once you dive in, you get trapped in flaws and problems. You focus on fixing them, and you stop looking toward the horizon.

So for rapidly evolving tech products and capabilities, my view is: use them, but don't use them too deeply — especially if you're the #1 person.

Of course, for the rank-and-file, or for product lines strongly tied to AI, you should encourage them. Let them try everything, judge by results, then gradually standardize. That's the better approach.

AI Coding Is Still Just an Assistive Tool

AI coding is a very good assistive tool for junior programmers. But maybe in two or three years, it'll force all junior programmers to level up to senior.

For intelligent assisted driving or other strong-AI capabilities, though, I think its help is relatively limited. It's just one tool among many.

To truly build out the whole infrastructure, the whole system — you could say it helps everyone at the application layer. But at the kernel layer, say if you want to write an operating system, what matters most is the infrastructure itself, not the coding.

Tokens Don't Matter; Compute and Data Do

Physical-World AI Token Consumption Is a Different Beast Than Digital-World AI

Large digital companies should pay moderate attention to tokens, but not obsessively, because some of their digitally supported operations may not actually be digital themselves.

Here's something very interesting: if our cars are only used three or four hours a day, how many tokens does that consume? I don't remember the exact figure anymore.

But you can think of it this way: the number of tokens used by digital AI is far, far lower than what physical-world AI needs to consume on its own.

An autonomous driving car is essentially an automatic machine that consumes tokens by itself.

So in the physical world, future token usage will be about how much the machine itself consumes and how much value that generates for machine and human alike. Human token usage generating value for people and enterprises — that's a completely different dimension.

I Don't Cap My Team's Token Use, But I Watch Data Costs

People say, if you don't control tokens, they'll burn through a year's budget in one quarter.

If someone can actually do that, what matters most to me is managing the top 10 most anomalous cases. Everything else I keep open. Because you can't know whether 1,000 RMB or 10,000 RMB per person per month is more valuable.

Most of our people's monthly salaries probably far exceed those numbers. If someone truly has the ability to spend more and generate greater value, why would you restrict them?

Right now our compute allocation goes mainly to the autonomous driving and cockpit teams. What we calculate is: give you 30,000 H100s, or 50,000 — what business can you run that uses them with high efficiency?

Many companies talk about engineer token usage numbers. I think those numbers are tiny compared to machine token consumption and what physical-world AI model training requires.

But we recently put special controls on data. Everyone talks about data's value, but very few companies see data's enormous cost.

In digital AI, data volume is small — dozens of terabytes can train a model. For us, one training run runs from dozens to hundreds of terabytes.

So how you manage data, use it, store it — that's massive money. Our rigid annual data costs probably approach over 1 billion yuan.

Which data has value? Which is only temporarily valuable? Which needs very fast access? Every usage decision costs tens of millions. You can analyze, optimize — there's huge business logic for improving efficiency, cutting costs, and raising effectiveness.

The "AI Frankenstein" Era Should End

Past Autonomous Driving Was a "Frankenstein"

Over the past several years, intelligentization developed very fast, but I actually think very unsatisfyingly. From earlier years — Toyota, Google, Baidu in China, Tesla overseas, and Xpeng Motors in China — everyone did a lot in assisted autonomous driving.

In some sense, it achieved results, but not enough height. At that time, I believe it was still AI algorithms plus software rules combined — I jokingly called it a "Frankenstein."

It wasn't actually designed and integrated with full AI driving, nor was the whole system built with full AI model capabilities. It was still fundamentally software logic.

The Ceiling Was Too Low; True Driverless Would Never Be Possible

Xpeng hit a massive inflection point last year. We were simultaneously developing two new generations of assisted autonomous driving, internally called VLA Generation 1 and VLA Generation 2.

Generation 1 was about expanding within the existing end-to-end domain — scaling up the model, reducing software rules, broadening usage scenarios and scope, strengthening backend capabilities like reinforcement and post-training.

That was one direction. But the other direction was: how do I discard that end-to-end logic entirely, use a larger foundation model, and focus more on first opening up the ceiling of autonomous driving, then converging the floor.

Converging the floor means fewer errors, acting and moving according to your specifications and expectations. But actually its generalization capability was very poor in many places.

For example, today no autonomous driving software from any company can drive smoothly in underground parking garages. All parking garage driving is memory-based assisted driving — it drives once, learns where your parking spot is, roughly what the route looks like.

Actually its understanding of the physical world is extremely low. The core problem is insufficient intelligence; the ceiling is too low. You could say you need to open up to 10,000-point possibility to do this well, but they were probably all around 1,000 points.

Around this time last year, our other VLA generation opened something new for me. I thought its ceiling might reach 100,000 to 1,000,000 points, but its floor was also disastrous then.

Normally you'd expect a product with 1,000-point ceiling and 900-point floor — that's decent capability. But then the floor might be only 100 points, lower than our other products' floors, with massive engineering problems.

Digital AI and Physical AI Are Completely Different Worlds

Why only realize this last year? I think most people weren't looking at this from a physical AI perspective. Digital AI and physical AI are two completely different methodologies and paths.

Digital AI, in some sense, uses human language. Language is the world.

Human language is highly summarized and condensed. But in the physical world, the data each of us sees every day simply cannot be summarized, described, restored, or replicated by language. It's too massive.

From another angle, in the digital AI market, many models just benchmark and score-compare. In the physical world that's laughable, because you don't just compare ceilings — you compare floors, you compete on breadth.

Quality, cost, materials, details, regulatory permissibility — these are all narrow boards, short boards, and long boards. Narrow boards must be widened, short boards lengthened, long boards made longer still.

In the digital world, you mainly look at long boards, barely at the other two. So physical-world CEOs either don't dare to bet, or feel overwhelmed by how many boards they still need to fix.

A Multi-Billion-Yuan Bet

We Shut Down the Old System — It Had Cost Several Billion Yuan

We finally made an enormous bet. Previously we used software engineering, software processes, plus AI algorithms and AI tools.

You could say it was still within a software business flow, using AI at some important nodes. But we believed it was wrong. What it produced was still software.

Because using software methodology with an AI toolkit produces a more powerful piece of software — what I call an "AI Frankenstein."

So we made a massive transformation and bet. We shut down our previous system entirely. That system had cost several billion yuan. Why?

I believed it could never achieve driverless. I believed it could never enable true robot generalization.

For example, today if we go to an unfamiliar venue, a robot could walk over and say "Hello, have a seat," and you say "I won't offer you a glass of water," and it says "Right, I don't need it."

With that old approach of strong rules plus limited algorithms using AI, you'd never achieve this. The intelligent electric vehicle we envision is a very smart car, but the method used would never make it infinitely smart.

Decided in My Head, All In

Last March I was thinking: why do all companies think they can make a passable autonomous driving system with simpler methodology and faster efficiency?

But after one or two years on this path, you discover that to solve many short-board problems, you limit your long board. You never reach Level 4 or Level 5.

This path may be a shortcut, but it's not a great road — it's a small path. We need to find a true great road.

There wasn't a landmark meeting. Basically decided in my head, all in. Around the end of Q3 last year, we made a massive move, completely restructuring the core organization of our autonomous driving center.

Every time window has many excellent people, but everyone has their own inertia. They habitually use past methods, want to use the latest tools and tech to make something better.

Most of the time this is right. But many times your entire working methodology must change. Adjust methodology, adjust mindset.

People Who Voted With Their Feet Left

Most non-AI managers thought whether you do A or B, you're probably wrong — because their AI awareness wasn't strong enough then.

Most AI-related managers: some said yes, some said no, but most were in between, unsure whether this was the right rhythm, whether it could achieve real results. In fact even I was the same then — but this is the fun of entrepreneurship.

The biggest opposition was many people voting with their feet. They didn't believe this could be done, so they left. In an organization, when thinking about organizational change, never use a small knife to chop a big tree, slowly. Once you're clear, chop it.

In certain areas you dare to bet — from organization to process to direction, everything changes. What people see may be the top business layer or application layer, but internally it goes all the way to the root.

How to use AI to drive certain businesses is extremely important, rather than using hardware to drive the whole system.

Why Is Physical AI So Hard?

Still Renovating the Old House

If you just make a more complex Frankenstein within software logic, it's like renovating a house — you have more materials, more craftsmanship, you're using AI, but it's still the methodology of renovating an old house. What comes out is still the old house, just maybe built faster.

We want something different. For something different at the most basic level, people need to see it before they believe.

But as a CEO, sometimes you believe first, let people see at certain nodes along the way, because that builds confidence, builds leadership.

Many times you lay things out because you believe. But Xpeng, though a startup, already has tens of thousands of people in scale. How to stage this out presents completely different challenges.

Waymo Has Worked 17 Years and Is Still a "Frankenstein"

Google bet on autonomous driving in 2009 — 17 years now. I think Waymo is a solution that's both good and very not good. Good in that its technical capability is decent.

Not good in that it's very hard to globalize, and it's inherently a more advanced Frankenstein — very difficult to achieve extremely high-intensity generalization within AI.

Even just making driving easier, safer, and faster for drivers — this one thing for 17 years, no company has done amazingly well.

Many very small companies, a few people, a dozen people, say they want to use AI to change the world. I think they completely underestimate the diversity and complexity of changing the physical world.

Like when we first charged from mobile internet into the auto industry, everyone thought if I just take one thing to the extreme, more extreme, it'll be great.

No — because you're not considering the right dimensions. You're in a digital dimension; taking digital to the extreme may have minimal value. The physical dimension has completely different requirements.

Applying Digital AI Methodology to Physical AI Is Insufficient

Most discussions I hear are basically about digital AI. Very few can think about AI from the perspective of a tens-of-thousands-people organization with massive real-world interaction.

Copying their methodology to physical AI — many parts aren't suitable. Some business logic may be right, but it's not comprehensive or sufficient enough. That's the biggest problem.

How to Do Physical AI? I'm Still Exploring

Today AI has four main directions.

First, how digital AI evolves — many people are thinking and acting on this.

Second, physical AI. It may be 100 times harder than digital AI, but I think starting 2027–2028, people will begin seeing physical AI results emerge — similar to the massive changes ChatGPT and Sora brought to digital AI, what jobs they massively help, even in the long term somewhat replace.

Third, beyond Google coupling AI with the human body — the human body has not just internal circulation but external circulation. Pharma is an extremely complex industry.

Fourth, how enterprises couple with AI. Department-AI coupling is relatively simple; enterprise-level is harder. I see some sub-1,000-person companies exploring more now; we're observing and learning.

The "Dao" of Robots: Build a Person, Not a Robot

From Quadruped to Biped, From "No Brain" to "Brain-Driven"

Xpeng's robots have three stages. First, 2018–2020, an independent team doing quadrupeds like many other Chinese robot teams.

2020–2023 was stage two. We spent three and a half to four years trying robots built with robot methodology, robots built with car methodology, and some stitched-together approaches in between — each with different successes and failures.

After 2023, when we saw new models in 2022, we felt that the robot brain, which we previously thought could never succeed, finally had a possibility of success.

Today many people say the robot cerebellum is already done. I say that's nowhere near a cerebellum? Walking forward slowly with the same monotonous gait — that's not cerebellum, that's spinal cord or brain stem, just maintaining balance. You're not even close to cerebellum.

So after 2023, Xpeng's robots entered a new inflection point: firmly moving from quadruped to biped, from not believing in the brain to firmly using the brain to drive a completely new robot design.

Plus from the auto field we learned: good tech doesn't mean good product, good product doesn't mean good commercial product, good commercial product doesn't mean you can scale up. By year-end we expect to enter SOP (mass production) for robots similar to cars.

2027 will very likely be the first year of high-grade robots entering commercial mass production, both China and the U.S. will be training. The past era of motion-remote-controlled robots will gradually decline as high-grade robots emerge.

Why Must It Be Humanoid?

We chose the most human-like humanoid robot. I think there are many solutions in this world, but humanoid is the only form that can truly integrate into human living environments.

Two examples. We previously made quadruped robots, dogs and horses. Whatever quadruped, entering rooms doesn't work. A 1.1–1.2 meter robot dog entering a room can't turn around at a nightstand.

A golden retriever raises its tail when turning, might scrape the wall and your bed — you don't think it'll get hurt, nor the bed.

But a robot dog — 100% it'll make both of you feel hurt. Make it very small, its capability level becomes very poor, just light companionship, with very short battery life.

Now biped. If a robot is fully armored, very imposing, 1.8 meters — even as its designer, you wouldn't want to walk just 1 centimeter from it.

You think it might be dangerous, hot, electric, dirty. If even adults feel this, what about elderly and children? What about safety and regulations in society?

Industrial robots can be built that way because they're not in homes or commercial spaces to begin with.

But if robots are to enter human society and create value for basic blue-collar and basic white-collar workers, they must come among us. So we chose a very hard path.

Today's Robot Is Just an Intermediate Version

Last year's explosive discussion — we completely didn't expect it. Last year's robot we started building in Q1 the year before; to us it was just an intermediate version.

We only wanted to verify: if a robot has this many joints, its walking effects would change and reduce dramatically. That's a completely different thing.

This generation is around 1.69–1.70 meters, a height both men and women find comfortable. It should be able to wear clothes, even have hair, but it cannot have its own face.

Partly uncanny valley, partly legal and sociological effects. What people see of Xpeng's robot today and what Xpeng is actually building still differ quite a bit.

Wait until second half of this year. I expect with very deep thinking, using the simplest intermediate form, with full company effort, we'll produce it.

How Was the Leader Chosen?

In 2023, we took our 300-person robot team down to under 60 and disbanded it. I believed we needed to reconstruct the entire robot logic — couldn't use pure car people, couldn't use pure robot people, but couldn't know nothing either.

So I chose a completely new team. Someone who knows some AI, some cars, some engineering, some robots — trying to build a completely newly thought robot logic.

Why him? Many times it's fate. His quadrant aligned with my thinking quadrant, and he also had to choose me. First you choose him, he also has to choose you.

So today many robot companies very quickly make a demo — to me that's nothing. Like the many Level 4 autonomous driving companies in China in 2017, same as robots today — doesn't mean these technologies will ultimately see real value.

Seeing My Weaknesses Through Skill

If even my capabilities can now be clearly logicized and skills-ified, what does that mean? It means not just basic blue- and white-collar, but higher-end blue- and white-collar face massive risk. CEOs can be replaced. Maybe in decades or a century, capabilities like mine can indeed be skills-ified, but by then every CEO will probably have stronger, more comprehensive capabilities — lose ABC, gain new DEF.

Never Admit Defeat, But Accept the Bet

Never admit defeat, but accept the bet

At the end of 2022, when Xpeng faced many challenges, I gave myself two thoughts, both related to betting. First: never admit defeat. Second: accept the bet.

If you balance these two well, it means even facing massive difficulties, you persist — maybe one more push and you're through.

But mentally prepare too. Like when I believed Generation 2 VLA should be scaled up, and finally firmly stopped Generation 1 VLA completely — from business to tech to organization — that was massive psychological and material pressure.

The more you hesitate, the more you wait, the more you want to observe, think about waiting 6 months, the harder success becomes.

Of Course Anxious in an Era of Radical Technological Change

Seeing problems, solving problems, and building systems that have both high ceilings and plugged floors to prevent problems — these are three completely different capability levels.

Before I started my company, I also saw countless problems. Once you become an entrepreneur, you discover many problems naturally should exist, you fundamentally can't solve them — you're a CEO, still can't solve them. You constantly reconstruct your theorems, even axioms.

After doing autonomous driving for many years, the more you did, the more it felt like Level 5 would never arrive. Because using software across all global scenarios, regulations, human-world interactions — it's infinite loops.

But when you truly reconstruct this with AI, you feel it might be possible. When it becomes possible, you develop new thinking about many original so-called moat logics.

Use Super Smart People for Super Hard Things

From late last year to first half this year, we recruited close to 80 Tsinghua-caliber PhD graduates for just one department. They're all expensive, but we're willing.

You need the ability to gather such people, support their long-term exploration, believe these young people can create miracles.

I call this talent potential. Use super smart people to do super hard things, rather than using very clear direction, process, or tools to forge them. At certain stages, this is very right.

20% Odds, But Sustained Investment

Today Xpeng already has tens of thousands of people. Tech change might complete in a month; organizational change, for a globalized mid-size organization, three years would already be terrifyingly fast — I even think 5 to 10 years counts as quick.

In sustained innovation, Xpeng has both patience and courage for long-term investment. There may be countless solutions in this world; we chose the most human-like humanoid robot. What are the odds? We probably have about 20%.

---

Heart Capital was founded in 2022 as a venture capital fund focused on investing in China's early-stage technology startups.

Heart Capital's team consists mainly of Lightspeed's founding partners, core investors, and senior investors from industry. The team's past investments include MetaX (688802.SH), Xpeng Motors (NYSE: XPEV, 09868.HK), Full Truck Alliance (NYSE: YMM), Sunmi Technology (06810.HK), RoboSense (02948.HK), Ambiq Micro (NYSE: AMBQ), Hanshow Technology (301275.SZ), FinVolution (NYSE: FINV), Qimengdao Group (NASDAQ: HERE), as well as LandSpace, Microsat, Baichuan, Yunmanman Cold Chain, World Logistics, Fan Deng Reading, and Lanhu, among others.

Rooted in China with a global vision, Heart Capital is committed to early-stage companionship and support for entrepreneurial teams with the potential to become future world-class companies in China's technology sector. Heart Capital advocates the value of "heart," believing technology can be a bridge connecting hearts. Heart Capital looks forward to accompanying more young Chinese entrepreneurs onto the world stage.