The autonomous driving supplier endgame is down to two or three players, and the window closes at year-end
Momenta's Cao Xudong judges that autonomous driving is like batteries: there is only good and better. The endgame for third-party suppliers is two or three in China and three or four globally; the first tier already holds 90% market share, the data flywheel and mass production are the entry ticket, and the next stop is physical AI and home robots.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The competition window closes at year-end
Cao Xudong still insists that competition in assisted driving will end in 2026: China will ultimately have 2-3 third-party suppliers, and the world 3-4. The mechanism he gives is that this industry has extremely strong first-mover and scale advantages; the leaders roll revenue and gross profit back into R&D, while the second tier cannot win customers and cannot afford core R&D. Currently the first tier (Momenta and Huawei) together hold 90% market share and have clear pricing power; he believes this opportunity window may close by the end of this year, and concentration will continue to increase.
— Cao XudongAutonomous driving goes from toy to necessity
Cao Xudong believes that two years ago autonomous driving was still a "fun toy," and now it has become a "usable product," but it has only just crossed the usability threshold; at a rate of 10x per year, by 2028 the driving capability of L2++ passenger cars will reach human level, thus moving from nice to have to must have. One piece of evidence is the change in customer organizations: the autonomous driving lead becomes the intelligentization lead, then the R&D lead, and even the chairman, making it a number-one engineering priority. Another piece of evidence is that in Nielsen surveys, autonomous driving has already ranked fourth in car purchase decisions, behind only price, appearance, and space.
— Cao XudongAutonomous driving is a battery-style business
Cao Xudong uses batteries as an analogy for autonomous driving: there is only "good and better," not a matter of taste where everyone likes something different; differentiation accounts for only 10%-20%. This kind of industry has extremely strong first-mover and scale effects, and the final players will be two or three in China and three or four globally. He rebuts the common prediction that "there will be many third-party suppliers, price wars will drive gross margins low," pointing out that the number of automakers doing autonomous driving will not decrease because of electrification, but suppliers will become highly concentrated; Momenta's gross margin rose from less than 20% in 2023 to over 70% in 2025.
— Cao XudongWorld model unifies the foundation
Cao Xudong says he will not do robots this year, and plans to invest in 2027. Two judgments support this: first, working backward from China as the starting point, home robots will reach an inflection point in 2030 and begin scaled commercialization; second, R7 validated the world model—pretraining on massive physical world data, then post-training on proprietary scenarios. Autonomous driving and robotics R&D paradigms are highly similar, and the infrastructure is reusable. He observes that Silicon Valley robotics companies were talking about LA a year ago, and this year they are all talking about world models; for example, one company used body-less data for pretraining and may have raised its success rate from 50% to 90%.
— Cao XudongRoboVan ranks first
Asked about new players like JD, Meituan, and Hello entering last-mile delivery, Cao Xudong gives his business ranking for the next 5 years: RoboVan first, RoboTruck second, RoboTaxi third. RoboVan does intra-city delivery, covering 4.2-meter Jinbei vans and small three-wheelers; RoboTruck does bulk logistics (hauling coal and ore), operating in suburbs and on rough roads, and does not touch the most difficult long-haul logistics. He calls long-haul logistics "the last piece of the autonomous driving puzzle," with high technical, commercial, and policy difficulty, to be done last.
— Cao XudongData-driven physical reasoning
Cao Xudong explains why he was convinced in 2016 that autonomous driving must be data-driven: he comes from a physics background and is used to using dimensional analysis to validate hypotheses. Human driving safety is about one major accident per 100 million kilometers; for autonomous driving to be widely adopted, it must be at least 10 times safer, and considering statistical confidence, about 100 validations are needed, so the amount of data required can never be accumulated by an owned fleet—it can only be collected through mass-produced vehicles. This reasoning established from day one the two-legged strategy of "data flywheel + mass production + fully driverless," unchanged for ten years.
— Cao XudongFrom research institute to fighting
Cao Xudong looks back at the company's transformation from the end of 2018 to 2019: before that it was like a loose research institute, over-promising to investors, so severely that outsiders could see it; in 2019 he used philosophy guidance and organizational adjustment (whoever makes products according to customer value gets promoted) and it took half a year to turn around, while now it can be done in a month. The transformation crystallized the principle of "reuse the main line, develop the main line"—when encountering problems, prioritize solving them with the main line architecture, and never open a branch line outside the main line, otherwise you will be dragged down by "chimneys"; the first customer project made him understand that the R&D and delivery systems must be data-driven and highly automated.
— Cao XudongFox and hedgehog decision-making
Cao Xudong says he does not take the path of "making the decision only now" for major decisions; instead, he spends one or two years, or even three to five years, repeatedly brewing and discussing within the organization in advance, so that when the time comes to actually do it, it is no longer a choice of whether to do it, but when and with what entry point. He repeatedly emphasizes returning to the front line to get first-hand information rather than listening to reports; for a CEO, the most important thing is not efficiency, but the correctness of cognition and judgment. Finally, he summarizes his methodology with "the fox knows many things, but the hedgehog knows one big thing": underlying principles matter more than multiple paths.
— Cao XudongIn their own words · checked verbatim
You use first-mover advantage and scale advantage to block a bit, and you get the conclusion we just discussed.
就是你用先发优势和规模优势卡一卡，就可以得到刚才的我们讨论的那个结论了
Cao Xudong5:17
When you encounter difficulties and cannot solve the problem, it is not about retreating to other methods, but about how to innovatively change your architecture and change your system.
遇到困难，解决不了这个问题的时候，不是说我又退回到其他的方式方法上去，而是说我怎么创新性的去改我的架构和改我的体系
Cao Xudong1:48:54
Figures
| Combined market share of first-tier third-party suppliers | 90% (Momenta and Huawei) | 5:17 |
| Cumulative R&D investment in autonomous driving | Over 1 billion USD, less than 2 billion USD | 17:58 |
| Target cost of home robot body | 10,000 USD or 10,000-20,000 USD | 33:56 |
| Expected retail price of home robot | About 100,000 RMB | 54:54 |
| Zhipu's market cap breakthrough after IPO | Over 1 trillion RMB in just over half a year | 1:10:13 |
| Human driving safety benchmark | One major accident per 100 million kilometers | 1:28:03 |
Glossary
- world model
- A generative model that predicts the evolution of the physical world over time; Momenta uses it to unify the foundation for autonomous driving and robotics.
- RoboVan
- An autonomous vehicle used for intra-city logistics delivery; Cao Xudong predicts it will be the best business in the next 5 years.
- RoboTruck
- An autonomous truck used for bulk commodity logistics, less difficult than long-haul logistics.
- post-train
- Fine-tuning and aligning a pretrained model with domain data to give it specific behaviors or skills.
- data scaling
- The law that model capability improves as training data volume grows; it is the theoretical basis of the data flywheel.
How to listen
Autonomous driving and robotics founders, automaker intelligentization leads, autonomous driving investors, and engineers following AI company strategy.
The last less than 10 minutes have poor audio quality and can be skipped; the rest is worth listening to in full.