He Xiaopeng: After ten years, no Chinese automaker has yet earned its seat
He Xiaopeng says the EV shakeout will likely resolve within five years, leaving roughly five survivors, but every company today—including XPeng—hasn't yet earned its place.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
After wealth, the hardest game is the most honest choice
He Xiaopeng says that after hitting financial freedom, it paradoxically became harder to commit to something difficult. He rates every startup at 100 difficulty; carmaking is 200. His reasoning: pick something harder, not easier, to challenge yourself. That's why he left venture capital to build cars.
— He XiaopengFifty billion in annual R&D demands five trillion in revenue to sustain
In the AI era, a car company needs at least 50 billion yuan in annual R&D—roughly 30 billion in AI, 20 billion in hardware and software. Working backward from the industry median of R&D at 10% of revenue, that requires 5 trillion in annual sales—independent brand revenue only, excluding joint ventures. Few Chinese automakers qualify; XPeng did 40 billion last year.
— He XiaopengKill the hardware bet early—before delivery is catastrophic, after is worse
When He joined XPeng, his first move was scrapping the car in development and starting over. He argues that moving the internet playbook—rapid iteration, obsessive polish—into hardware feels doable but deceives. Hardware can't recover from mistakes the way software can with patches. His rule: the moment you know direction is wrong, cut before handover means a big loss; cutting after means catastrophe. Kill what you think is wrong as soon as possible, even if morale tanks.
— He XiaopengHiring great people and stepping back works for software, not for cars
Early on, He ran XPeng as chairman thinking he didn't need to grind—hire great talent, delegate, move on. Natural for someone from venture capital. But that playbook breaks in auto. Information filters up through layers; you end up the last person in the company to know what's true. That's why he later switched to hands-on management of frontline operations.
— He XiaopengWhen performance drops, that's when bold personnel changes feel most earned
About 90% of XPeng's top managers cycled in just over a year, yet stayed calm because most left gracefully, stock in hand. He makes a counterintuitive point: when a company fails, founders usually freeze on personnel moves, but doing them decisively then—the worse the situation, the more you dare move, the more readily they accept it. Bad numbers give you the opening for the conversation.
— He XiaopengAfter G9, he nearly stepped down—then stayed because retreat wasn't an option
After the G9 launch failed in late 2022, He spent months working backward to find the root cause—and found himself. He seriously considered stepping down; the company had grown to over 10,000 people, and staying felt unfair. But swapping leaders might not help, and he had no exit, so he stayed. By his estimate, companies like XPeng need two to three years to climb out of that hole.
— He XiaopengMost companies say AI matters; real commitment shows in deployed capital and results
At an AI industry panel, He asked the room who was doing foundation model pretraining. Three or four hands out of hundreds. Most carmakers haven't even fine-tuned—they license and relabel. Even China's top cloud companies rarely run heavy AI training on their own infrastructure. His conclusion: when companies say AI is a priority, it's theater. Real prioritization shows in capital spent, time deployed, and results delivered.
— He XiaopengAI code cuts output 22% but review adds 20% back—the gain comes at 50%
XPeng uses AI heavily for coding and does see about 22% efficiency gains—but review cycles, tool onboarding, and automated testing add back roughly 20%, netting a loss so far. Not unique to XPeng; across the industry, labor-heavy companies break even or lose. He sees it differently: not a failure, but a threshold. Cross 50% efficiency and the math flips to pure gain. So deploy the tools now; don't wait for the numbers to work.
— He XiaopengIn their own words · checked verbatim
If you want to ruin a brother, tell him to build cars—it's brutal.
我说如果你想害一个哥 你就劝他去造车 因为太苦了
He Xiaopeng35:19
From a project management angle, you should always kill something you think is wrong as early as you can.
我自己从一个项目管理角度来看 你永远应该 尽早把一个你觉得不对的事情干掉
He Xiaopeng55:29
I never stopped to ask why they could stop at a hundred thousand units. I just thought, everyone's saying it, and they're seniors, so it must be right.
我心里当时也没有去分析 嗯 为什么他们做到十万台就可以了 反正就觉得既然大家都这么说 因为他们都是前辈
He Xiaopeng1:01:37
A lot of founders and CEOs, the worse things get, the less you dare move people. But that's backward. Be more decisive when things are bad, and they're actually more likely to accept it.
很多的时候 有很多的创始人或者CEO 你越差他越不敢动 但错了 你越坚决 他有可能越适合
He Xiaopeng2:03:10
Then you think the best solution is soul side. I don't mean soul side in the physical sense—I mean you shouldn't be running this company. You should give up the seat.
然后你觉得解决问题最好方法是soul side。 我讲的soul side不是那种物理意义的soul side,是指这个企业不应该归你管。 这个位置不要坐了。
He Xiaopeng2:25:21
Figures
| XPeng annual R&D target | 50 billion yuan (approximately 30 billion in AI, 20 billion in hardware and software) | 49:27 |
| XPeng custom AI chip compute | 2250 Tops (three to ten times the industry average) | 1:28:54 |
| AI coding efficiency impact | Reduces work by approximately 22%, but review and testing add back approximately 20% | 2:44:41 |
| Time to recover from G9 crisis | Approximately 18 months | 2:27:25 |
| Flying car cumulative funding | Seven to eight hundred million dollars | 1:45:59 |
| Flying car target safety redundancy | Three thousand times higher than civil aviation safety standards | 1:50:03 |
Glossary
- soul side
- He Xiaopeng's self-coined term for stepping back from leadership after realizing you're the core problem—not physical departure, but accepting you shouldn't run the company anymore.
- Tops
- Trillion operations per second; the unit of measurement for AI chip computing capacity.
How to listen
Founders, investors, and product managers tracking where China's EV shakeout stands and the true cost of AI R&D and organizational restructuring.
Skip the first 25 minutes—UC startup memories and brand-naming stories are biographical filler with low information density.