He Xiaopeng: Large models are bad for small companies — the stronger you are, the more data, the more cost
The network effect of big AI is: the stronger you are, the more data; the more data, the more cost; the more cost, the more cars you have to sell to support the system.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The dimension where large models can actually make money is autonomous driving
He Xiaopeng judges that language large models have basically failed to make money — apart from NVIDIA or something like Microsoft, OpenAI raised money rather than earned it. He thinks the dimension where large models can genuinely make money is AD, autonomous driving: raise autonomous-driving capability 50x to 100x, turn assisted driving into fully autonomous driving or something close to driverless, and users will be willing to pay. His reasoning: users won't buy a piece of software or a service for tens of thousands of yuan on its own, but spending 100,000-plus, 200,000-plus or 300,000-plus yuan on a car they like that includes this capability is a very high-probability purchase.
— He XiaopengAutonomous-driving large models need brakes and a goalkeeper
He Xiaopeng distinguishes two directions for large models: language large models don't pursue reliability or robustness — if they say something wrong, no way, don't care. Autonomous driving is different: if it gets something wrong you have to preserve safety, so you need a controller, you need brakes. He explains that XPeng could adopt the new-generation technology relatively quickly because its previous-generation autonomous-driving capability was strong enough that the previous generation could be modified into a controller; with a controller and safety brakes, the new generation can reach higher capability in certain areas, and the controller and brakes patch and compensate for its weak spots.
— He XiaopengPutting large models in cars means rewriting the data-flow logic
He Xiaopeng says deployment is not releasing some model; it is the team reading the papers, understanding the architecture, then rewriting the data-flow logic according to the logic of large models, adding data collection, doing pre-training, training an engine, and then testing in a simulation environment. At the very beginning the simulation results were extremely bad, what they wrote was completely unusable, and the most important thing was that the data volume was not large: traditional programming assumes 100 million scenarios for a right turn and you can cover at least 1 million to 10 million of them, but a large model cannot cover the first 1 million scenarios — for a long time it is still 0. The core is data volume and enormous training; he reveals that training costs in the second half of this year are over 100 million US dollars.
— He XiaopengBig AI's network effect is bad for small companies
He Xiaopeng says intelligence will form a very technical network effect: in big AI, the stronger you are, the more data; the more data, the more cost; the more cost, the more cars you have to sell to support the system. When a supplier sells you something, it's a piece of software I made and sell to you, but it doesn't require continuous training — otherwise that cost is a joke. But many future collaborations around large models must last a long period of time, and be very large. So in a sense large models are bad for small companies, because their time cost, business logic, R&D barriers and training costs are all thresholds. He mentions XPeng's 3.5 billion yuan a year of AI investment, which most companies cannot do.
— He XiaopengReplicating FSD takes tens of billions and has only a 5% success rate
Asked how much money, how much time and which foundational elements are needed to replicate Tesla's end-to-end FSD, He Xiaopeng says a few tens of billions. First you need a very large amount of money, then you need super-strong people — it absolutely cannot be done by ordinary programmers relying on algorithms; you need many cars running on the roads, you need the entire execution capability to turn, and the time to turn is also long, and the probability is also very low. He draws an analogy to startup success rates: if you invest a few tens of billions, several years and a thousand people, and the success rate is only 5%, will you invest? That is one of the thresholds. Putting in money only means it is possible to succeed, not that it can succeed.
— He XiaopengThe US-China AI gap is widening, and acceleration is the key
On this trip He Xiaopeng spent two weeks in the US across nine cities, and judges that the US-China gap in AI is widening. On last year's trip he saw many people in Silicon Valley pivoting to AI, but had not yet seen AI's changes; across three trips spanning more than a year, the changes are bigger than imagined. This time he looked at both Tesla and Waymo: Waymo has not yet pivoted to large models and still uses its original algorithm system; previously using Waymo felt far ahead of Tesla, but this time it is obvious that acceleration is the most important point. He says people use a traditional linear speed, using past experience in their heads, to look at the speed of technological development, but in large-model-oriented AI, the pace of technological change is starting to accelerate.
— He XiaopengLarge models have not lowered the autonomous-driving barrier
Someone asked whether large models have lowered the autonomous-driving barrier and whether it is just about piling up chips; He Xiaopeng disagrees. He gives the layers: compute is more foundational, called BOM; above compute there is the model, or OS; above OS there is data; above data there is globalisation and globalisation policy; above that is experience. He says no one can just take an ordinary large model and run autonomous driving with it, because all models today, frankly, look at things from a non-real-time, low-reliability angle, but autonomous driving is different — it needs millisecond level, at least hundred-millisecond or even ten-millisecond level. Data is also very hard: every OEM's master's-level work starts over, and how do you satisfy global scope plus policies and regulations.
— He XiaopengThe chance of FSD doing Robot Taxi within a year is zero
Asked about the logic that once FSD's Robot Taxi lands, a Tesla at 30,000 US dollars taking twenty orders a day pays back in seventy-five days, He Xiaopeng says this logic holds in the long run but not in the next few years, because there will be all kinds of additional problems. He completely does not believe FSD will be able to do anything on Robot Taxi within the next year; it might be implemented in a local scenario, but the larger-scale, better-executed version everyone imagines, he thinks the probability is zero. He cites his own experience this trip riding Waymo and taking over twice — actually it was not him taking over, it was the cloud taking over twice; you need a cloud-based management system, and different countries and different regions are completely different.
— He XiaopengIn their own words · checked verbatim
Today for autonomous driving, if you invest a few tens of billions, several years, a thousand people, and the success rate is only 5%, will you invest? That is one of the thresholds. So putting in money only means it is possible to succeed, not that it can succeed.
今天对自动驾驶，如果你要投几百亿，几年，千人，成功概率只有5%，你会不会投，这就是门槛之一，所以投钱，只是代表有可能可以成功，不代表它可以成功。
He Xiaopeng47:23
People use a traditional linear speed, using past experience in their heads, interesting or unintentional, to look at the speed of technological development, but today in large-model-oriented AI, I think the pace of technological change is starting to accelerate.
大家会用一个传统的线性的速度，用过去的，你在你脑袋中间，有意思或者无意思的经验，去看一个科技的发展速度，但今天在以大模型为导向的AI上面，我觉得技术的变革速度在开始加速。
He Xiaopeng1:15:38
Figures
| XPeng's training cost in the second half of this year | over 100 million US dollars | 31:17 |
| XPeng's annual AI investment | 3.5 billion yuan | 43:22 |
| Capital needed to replicate Tesla's end-to-end FSD | a few tens of billions | 46:32 |
| Success rate of investing a few tens of billions, several years and a thousand people in autonomous driving | 5% | 47:23 |
| He Xiaopeng's US trip | two weeks, nine cities | 1:14:36 |
| Experience gap between Waymo and Tesla | three times | 1:18:42 |
| XPeng's headcount | 16,000 | 1:00:27 |
| XPeng grew from a few thousand employees in one year to | 10,000 | 55:26 |
Glossary
- AD / Autonomous Driving
- He Xiaopeng's shorthand for XPeng's autonomous-driving capability.
- BOM / Bill of Materials
- The list of vehicle hardware costs; He Xiaopeng says it is hard to keep BOM constrained while still building a good car.
- Robot Taxi / driverless taxi
- Using autonomous vehicles to provide mobility services; He Xiaopeng thinks scale is hard to reach within a year.
- OTA / Over-the-Air
- Software updates pushed periodically by automakers; XPeng has about 1,000 new features this time.
How to listen
Founders, investors and automaker practitioners watching autonomous driving and large-model deployment, especially those who want the engineering details and the cost threshold.
The opening discussion of seating and colors at the Xiaomi SU7 launch can be fast-forwarded.