Two or three assisted-driving suppliers will be left, and the window shuts by year-end
Momenta's Cao Xudong (曹旭东) argues that assisted driving is like batteries — there is only good and better — so the endgame leaves two or three third-party suppliers in China and three or four globally; the first tier already holds 90% market share, a data flywheel plus mass production is the ticket in, 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 competitive window closes at the end of this year
Cao Xudong still holds to the position that the assisted-driving contest ends in 2026: 2-3 third-party suppliers left in China, 3-4 globally. The mechanism offered is that this industry has extremely strong first-mover advantage and scale advantage at once — the leaders roll revenue and gross margin back into R&D, while the second tier can neither win customers nor afford core R&D. The first tier today (Momenta and Huawei) already accounts for 90% of the market combined, with clear pricing power; on this reading the window of opportunity may close by the end of this year, and concentration will keep tightening.
— Cao XudongAssisted driving stops being nice to have by 2028
Two years ago, Cao Xudong says, assisted driving was still a ‘fun toy’; now it is a ‘usable product’ — but only just over the usable threshold. At 10x improvement a year, by 2028 the driving ability of an L2++ passenger car can reach human level, which is what turns it from nice to have into must have. One piece of evidence is organizational change on the customer side: the head of assisted driving became the head of intelligent vehicles, then the head of R&D or even the chairman — it is now a program owned by the number-one executive. Another is a Nielsen survey in which assisted driving already ranks fourth in the car-buying decision, behind only price, styling and interior space.
— Cao XudongAssisted driving is a battery business: only good and better
Cao Xudong uses batteries as the analogy for assisted driving: there is only ‘good and better’, not different strokes for different folks, and differentiation accounts for just 10%-20%. Industries of this kind have extremely strong first-mover advantage and scale effects, so the final players come down to two or three in China and three or four globally. Against the common prediction — many third-party suppliers, a price war, thin gross margins — the counter is that the number of carmakers doing assisted driving will not shrink because of electrification, but the supplier side will concentrate heavily; Momenta's gross margin went from under 20% in 2023 to over 70% in 2025.
— Cao XudongOne world model becomes the shared base for cars and robots
Cao Xudong says there will be no robot work this year, with investment planned for 2027. Two judgments support that. First, reasoning backwards from China as the starting point, home robots reach an inflection point in 2030 and begin commercializing at scale. Second, R7 validated the world model — pre-train on massive physical-world data, then post-train on proprietary scenarios; the R&D paradigm for autonomous driving and robotics is highly similar and the infra is reusable. The observation from Silicon Valley: robotics companies were talking about LA a year ago and are all talking about world model this year — one of them, for instance, pre-trains on data collected without a robot body, possibly lifting success rates from 50% to 90%.
— Cao XudongRoboVan beats RoboTruck and RoboTaxi over the next five years
Asked about the new players crowding into last-mile delivery — JD.com (京东), Meituan (美团), Hello (哈罗) — Cao Xudong ranks the businesses for the next five years: RoboVan first, RoboTruck second, RoboTaxi third. RoboVan handles intra-city delivery, covering 4.2-metre Jinbei (金杯) vans and small three-wheelers; RoboTruck handles bulk logistics (hauling coal and ore) on suburban and rough roads, and stays away from the hardest case, trunk-line freight. Trunk-line logistics gets called ‘the last piece of the autonomous-driving puzzle’ — hard on technology, business and policy alike, and therefore the thing to do last.
— Cao XudongA self-run fleet can never collect enough data; only mass production can
Cao Xudong explains the conviction, held since 2016, that autonomous driving had to be data-driven: trained as a physicist, the habit is to test a hypothesis with dimensional analysis. Human driving safety is about one serious accident per 100 million kilometres; driverless cars need to be at least 10 times safer to be adopted at scale, and statistical confidence then requires roughly 100 validations — so the amount of data needed could never be accumulated by a self-operated fleet, and can only be collected by mass-production vehicles. That derivation fixed the two-legged strategy of ‘data flywheel + mass production + full driverless’ on day one, and it has not changed in ten years.
— Cao XudongWhoever ships customer value gets promoted — that ended the over-promising
Cao Xudong looks back on the company's transformation from the end of 2018 into 2019: before it, the company was like a loose research institute and over-promised to investors, badly enough that outsiders could see it. In 2019 the fix was guidance on principles plus organizational change — whoever delivered a product measured by customer value moved up — and it took half a year to turn around; the same thing now takes a month. What settled out of that transformation is the principle of ‘reuse the main line, develop the main line’: when a problem appears, solve it with the main-line architecture first and never open a branch outside the main line, or ‘chimneys’ will drag you down. The first customer project is what made it clear that the R&D and delivery system has to be data-driven and highly automated.
— Cao XudongBig decisions are never made at the moment of deciding
Cao Xudong says major decisions do not go down the ‘decide only now’ route; instead they are brewed and argued through inside the organization one or two years — even three to five years — ahead of time, so that when the moment to act arrives the question is no longer whether, but when and from which entry point. The repeated emphasis is on going back to the front line for first-hand information rather than listening to reports; for a CEO the most important thing is not efficiency but whether the perception and the judgment are right. The closing summary of the methodology is ‘the fox knows many things, the hedgehog knows one big thing’: underlying principles matter more than multiple paths.
— Cao XudongIn their own words · checked verbatim
You just run it through first-mover advantage and scale advantage, and you get exactly the conclusion we were just discussing.
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 run into difficulty, when you can't solve the problem, it's not that I retreat back to some other method — it's how do I innovatively change my architecture and change my 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 spend on autonomous driving | over US$1 billion, under US$2 billion | 17:58 |
| Hardware cost target for a home robot | US$10,000, or US$10,000-20,000 | 33:56 |
| Expected retail price of a home robot | about RMB 100,000 | 54:54 |
| Zhipu's market cap after listing | past RMB 1 trillion in a little over six months | 1:10:13 |
| Human driving safety benchmark | one serious accident per 100 million kilometres | 1:28:03 |
Glossary
- world model
- A generative model that predicts how the physical world evolves over time; Momenta uses it as the unified base for assisted driving and robots.
- RoboVan
- An autonomous vehicle for intra-city logistics delivery; Cao Xudong predicts it is the best business of the next five years.
- RoboTruck
- An autonomous truck for bulk-commodity logistics; less difficult than trunk-line freight.
- post-train
- Fine-tuning and aligning a pre-trained model on domain data so it acquires specific behaviors or skills.
- data scaling
- The regularity that model capability improves as the volume of training data grows; the theoretical basis of the data flywheel.
How to listen
Assisted-driving and robotics founders, OEM executives running intelligent-vehicle programs, autonomous-driving investors, and engineers who follow AI company strategy.
The audio quality is poor in the final stretch of under 10 minutes and can be skipped; listen to everything else.