Manchi | 晚点聊 LateTalk Summary

1. Key Themes

The Autonomous Driving Market Has Already Consolidated — The Window Is Closed

Momenta and Huawei together command approximately 90% market share in China's urban NOA (Navigate on Autopilot) third-party supplier segment. Cao Xudong argues this consolidation is now irreversible, with the competitive window having closed by end of 2025/early 2026.

"The top tier — their combined market share has already reached 90%. It's just us and Huawei adding up to 90%. And the two of us have fairly strong pricing power." [00:05:27.120]

"The opportunity window that could change this landscape may have already closed, or will close by year-end." [00:04:58.800]

Autonomous Driving Is Now a "Must-Have," Not a "Nice-to-Have"

The product has crossed a threshold from toy to genuinely useful product, and is now ranked 4th in purchase decisions according to Nielsen research — behind only price, appearance, and interior space. Cao Xudong predicts by 2028 it will become a hard requirement.

"It's already moved from being a fun toy to becoming a useful product... If we maintain ten-times improvement every year, I believe by 2028, even an L2++ system will be at least at human level in terms of driving safety. At that point it crosses from 'nice to have' to 'must have.'" [00:08:53.060]

The "Flywheel" Compounds Into Adjacent Physical AI Domains

Momenta's foundational World Model (R7) has been validated to work across all autonomous driving applications — mass production, RoboOne, RoboTruck, RoboTaxi. The same foundation transfers to robotics, making each expansion progressively lower-cost.

"We verified with R7's World Model one very important thing: that a single World Model can support all autonomous driving applications, including mass production, RoboOne, RoboTruck, RoboTaxi — and they all work. And if this Foundation Model improves ten times, then every application improves ten times." [00:46:09.940]

Robotics Is the Third Curve — With a 2030 Inflection Point Target

Cao Xudong has been thinking about household robots since 2020–2021 during COVID. He projects 2030 as the commercialization inflection point based on three factors: compute costs coming down to support on-device physical brain and language brain simultaneously, hardware/actuator costs reaching ~$10,000–$20,000, and World Model training paradigms now validated.

"We're inferring that around 2030, household robots could reach an inflection point — beginning scaled commercialization. This may be an important judgment." [00:28:26.820]

"For home robots to scale, the hardware cost probably needs to get to around $10,000, or ideally $10,000 to $20,000." [00:33:54.920]

The "Good vs. Better" Industry Dynamic Favors Winner-Take-Most

Cao Xudong makes a non-obvious observation: industries where products are simply "good vs. better" (no subjective taste differentiation) tend to consolidate to 2–3 players with strong pricing power, unlike lifestyle/fashion products where many brands coexist. Autonomous driving and batteries follow this pattern.

"The 'good vs. better' industry — it just has massive linear advantages and scale effects. The final number of players won't be that many, maybe two or three domestically, three or four globally. This is determined by industry characteristics. Whereas 'different strokes for different folks' industries — like clothing — there are many brands." [00:14:13.540]

Momenta's Profitability Path Is Deliberate — Revenue-Funded R&D, Not Capital Burn

Gross margin grew from under 20% in 2023 to over 70% in 2025. The company plans to break even in 2027 and be profitable in 2028, with GPU/compute R&D investment doubling annually (~$200M this year, $400M next year, $600M in 2028). The explicit strategy: fund robotics R&D from autonomous driving gross profit rather than continuous equity raises.

"Our R&D investment in compute for training large models — this year might be $200 million, next year $400 million, 2028 perhaps $600 million. We believe this growth pace is sufficient to deliver ten-times improvement every year." [00:25:31.940]

Data-Driven Belief Must Be Genuinely Held, Not Lip-Service

Cao Xudong distinguishes between believing and truly believing in a data-driven approach. True belief means when you hit a wall, you find new architectures to make it work rather than abandoning the method.

"Truly believing means when you encounter all problems, you start from this angle to solve them. When you face difficulties and can't solve them, you don't retreat to other methods — instead you ask, how do I innovatively change my architecture and system so the data-driven approach can work." [00:49:30.780]

The "Zoomability" Framework — A Quantifiable Leadership Skill

Cao Xudong introduces the concept of "zoomability" — the ability to rapidly zoom in and out between abstraction layers. He frames it as quantifiable (e.g., how many layers up you can zoom out plus how many layers down you can zoom in), and argues this is essential for major decisions.

"This zoomability might even be quantifiable — how many layers up can you zoom out, say four layers, and how many layers down can you zoom in, say five layers. Then zoomability might equal 9. That person is very capable." [00:19:14.040]

FSD's China Entry Is Welcomed as an Accelerant, Not a Threat

Cao Xudong explicitly says he "very much welcomes" Tesla FSD entering China, believing it will accelerate industry consolidation and drive out weaker players — analogous to Tesla's EV entry into China accelerating poor EV brands' exit while strengthening good ones. He asserts Momenta's R7 will be "evenly matched" at the time of FSD's entry.

"I'm very much looking forward to it... Like Tesla's entry into China's EV market at that time, it accelerated the exit of those poor brands while good brands got stronger. So you think FSD entering China will accelerate industry consolidation — exactly, driving the whole industry to compete on safety, quality, and experience together." [00:16:38.760]

The "No Smokestacks" Architecture Principle

A key operational breakthrough from a difficult first mass production project: avoid creating too many technical "branch lines" or independent stacks. Reuse and develop the main trunk. A senior Huawei executive's framing resonated deeply.

"A very senior Huawei executive once said to me — he didn't use the word 'main line' but said: 'Xu Dong, you must never have so many smokestacks. With that many smokestacks, you'll be torn apart in five places.' I thought that was so vivid. Smokestacks are branch lines." [00:43:06.380]

2. Contrarian Perspectives

The "More Players = More Competition" Thesis Is Wrong for Autonomous Driving

Most people assumed that commoditized third-party autonomous driving suppliers would face brutal price competition and thin margins — like many hardware component industries. The opposite happened: gross margins expanded from sub-20% to over 70%. The reason is that "good vs. better" products without taste differentiation actually concentrate to oligopolies with pricing power.

"At the time I found it very strange, hard to understand. But actually quite a few industry bigwigs thought the same way and said so publicly. But it doesn't fit the industry characteristics." [00:15:39.920]

Tesla's Vertical Integration Is a Constraint on Its Supplier Ambitions, Not a Strength

While conventional wisdom holds that Tesla's FSD advantage comes from vertical integration, Cao Xudong argues this actually prevents Tesla from becoming a third-party supplier at scale — because no one will buy Tesla's technology who also competes with Tesla in selling cars. Scale effects that work for a single brand become a limiting factor when trying to serve an industry.

"A car company that makes its own cars and also wants to sell autonomous driving systems — it's actually a constraint, a limitation. Consumers won't buy that. Who would buy Tesla's system? So it can't actually become a third-party supplier." [00:07:25.360]

Explaining Your Vision to Investors Is Useless — Just Execute

Against the conventional startup advice of storytelling and vision-casting to investors, Cao Xudong argues explanations have zero value. The only thing that works is demonstrated results.

"Explanations are useless. Really, explanations have no use at all. Especially for outside investors — explanations are completely useless. You have to actually do it, and only then do people believe you. If you haven't done it, why would anyone give you such a high valuation to invest?" [00:30:56.920]

Anthropic Was the "Cheapest AI Company" at a Specific Moment in Early 2025

Contra the narrative that Anthropic was expensively valued, Cao Xudong argues that when Claude Code transitioned from L2 to L4 coding agent, Anthropic was actually underpriced on a price-to-sales basis relative to the market it was opening up — from ~$10B to potentially $10 trillion addressable.

"Anthropic may have been the cheapest AI company at that point in time — a simple PS calculation would tell you that... Anthropic made a particularly significant contribution to the whole large model industry by transforming an L2 coding agent to an L4 coding agent in a relatively short time... Once it becomes an L4 coding agent, the entire coding market size is trillions of dollars." [00:11:57.660]

The Smarter the Researcher, the More Dangerous Their Untested Theories

Cao Xudong inverts the common assumption that intelligence is a research advantage. He learned from Microsoft Research's Sun Jian that smart people's self-consistent logical frameworks are actually a liability if not relentlessly tested against raw data — because their internal coherence makes errors harder to detect.

"Smart people sometimes aren't so fond of painstakingly going through all this data and correlating their logic or theory with all the data. The truly elite person is a smart person who also puts in hard work." [00:01:04.230]

3. Companies Identified

Momenta

Description: Chinese autonomous driving company focused on mass production ADAS and full autonomy, founded 2016, recently IPO'd. Why mentioned: Primary subject of interview; holds 65%+ third-party high-end ADAS market share in China; gross margin grew from <20% (2023) to >70% (2025); operating losses narrowing from ~1B RMB to ~300M RMB; planning break-even 2027, profitability 2028.

"In the third-party high-end mass production autonomous driving market, our market share exceeds 65%." [00:02:32.330]

Huawei

Description: Chinese technology conglomerate with significant autonomous driving business (Huawei Intelligent Automotive Solution). Why mentioned: The only other company alongside Momenta named as a certain winner in China's top-tier autonomous driving supplier consolidation; combined they hold ~90% market share with strong pricing power.

"From current market share perspective, it's us and Huawei that have the highest market share. Other companies still have a very large gap." [00:04:29.440]

Anthropic

Description: US AI safety company, maker of Claude. Why mentioned: Cao Xudong identified Anthropic as potentially the "cheapest AI company" at a specific moment in early 2025 and credited them with the breakthrough of taking coding agents from L2 to L4, unlocking a multi-trillion dollar market.

"Anthropic made a particularly significant contribution to the whole large model industry — transforming an L2 coding agent to an L4 coding agent in relatively short time... Once it becomes an L4 coding agent, the entire coding market size is trillions of dollars." [00:11:57.660]

Physical Intelligence (Pi)

Description: Silicon Valley robotics AI company focused on the "brain" side of robotics. Why mentioned: Named as a representative Silicon Valley company proving out the World Model approach for robotics; validates Momenta's strategic direction.

"Companies like Pi — they also do the brain side." [00:36:22.600]

Unitree Robotics (宇树)

Description: Chinese robotics hardware company, same founding year as Momenta (2016). Why mentioned: Cited for doing pioneering work in China's robotics space, coming from the hardware/body-first direction — contrasted with Momenta's planned brain-first approach.

"Unitree has done very pioneering work in China's robotics field... They approach it more from the body angle, but I'm sure they've also moved from the small brain to the big brain." [00:35:24.160]

Tesla

Description: US electric vehicle and autonomous driving company. Why mentioned: FSD used as the competitive benchmark; discussed extensively regarding China entry timing and competitive impact; Cao Xudong claims R7 will be "evenly matched" with FSD V14 at time of China entry.

"If FSD enters at year-end, and our R7 at year-end... I think it'll be evenly matched." [00:17:07.960]

BYD

Description: Chinese EV and battery manufacturer. Why mentioned: Used as a case study of how vertical integration can be a leveraged bet that pays off in a specific window, but isn't the most fundamental reason for success — product quality was the core.

"BYD's real breakout had many reasons. This vertical integration leverage, combined with that perfect timing, greatly accelerated its development. But it's not the most fundamental reason. The most fundamental thing is that it made a product as good as the Han at that point in time." [00:22:04.580]

SensTime (商汤 / SenseTime)

Description: Chinese AI computer vision company. Why mentioned: Cao Xudong worked there in 2015 before founding Momenta; described as a formative crucible where he learned how to build products, serve customers, and operate at startup intensity.

"My overall entrepreneurial process — the most physically exhausting was at SenseTime... The company had just been founded, the Beijing office had just opened, there were very few full-time employees, and we were already starting to chase clients and deliver." [00:13:25.040]

Waymo

Description: US robotaxi company, subsidiary of Alphabet. Why mentioned: Cited as the catalyst for the recent wave of RoboTaxi entrants in China; its success and large fundraise inspired new players entering the space.

"I think it's related to Waymo having achieved very good results in RoboTaxi. On the other hand, Waymo also raised very large amounts of funding at good valuations." [00:44:12.240]

Pony.ai (小马智行)

Description: Chinese autonomous driving company, primarily RoboTaxi. Why mentioned: Noted as an already-listed company with RoboTaxi as its main business; gross margin cited at approximately 15% in 2025, significantly lower than Momenta's 70%+, illustrating the fleet ownership model's margin drag.

"Among already-listed companies with RoboTaxi as their main business, Pony and WeRide's 2025 gross margins are both much lower than ours — one around 15%, one around 30%." [00:40:14.560]

WeRide (文远知行)

Description: Chinese autonomous driving company. Why mentioned: Same context as Pony.ai; gross margin cited at approximately 30% in 2025.

"One is around 15%, one around 30%." [00:40:14.560]

Generalist (Physical Intelligence or similar unnamed Silicon Valley company)

Description: A Silicon Valley robotics company doing pioneering work on body-agnostic pre-training for robotics. Why mentioned: Validated the World Model approach by using large-scale embodiment-agnostic data for pre-training, then fine-tuning with embodiment-specific data — improving task success rates from ~50% to ~90%.

"A representative company like Generalist — it uses large-scale data, and moreover body-agnostic data, for pre-training, and then uses body-specific data for fine-tuning. This pushed success rates from maybe 50% up to 90%. This has brought significant proof points to the robotics field." [00:29:55.940]

Microsoft Research Asia (微软亚洲研究院)

Description: Microsoft's flagship research lab in Asia. Why mentioned: Formative institution for Cao Xudong; where he worked under Sun Jian and developed his research philosophy around data and experimentation over pure theory.

"After interning at the place I mentioned, then Sony Research, then Yahoo Research, constantly accumulating experience... only when I applied to Microsoft Research did I get the opportunity to intern there." [00:01:04.230]

Didi / Gaode (滴滴 / 高德)

Description: Chinese ride-hailing and mapping platforms. Why mentioned: Named as companies Momenta explicitly chooses NOT to compete with in RoboTaxi operations in China; Momenta's strategy is to be the algorithm/software provider (ASP) and let platform companies handle fleet operations.

"We're not positioning ourselves as a platform competing with Didi or Gaode. We're more focused on being an ASP, earning ASP revenue sharing." [00:40:44.480]

Zhipu AI (智谱)

Description: Chinese large language model company, publicly listed. Why mentioned: Used as an example of how the new generation of AI companies (large models) commands dramatically higher valuations than the previous wave (computer vision), because their value creation potential is structurally larger.

"Zhipu AI went public and in just over half a year has already broken through 1 trillion RMB in market cap. Why is their commercial impact and capital market impact so different?" [00:10:58.340]

4. People Identified

Cao Xudong (曹旭东)

Description: Founder and CEO of Momenta; started company at age 30 in 2016; background in statistical physics, self-taught AI, worked at Microsoft Research Asia under Sun Jian, then SenseTime before founding Momenta. Why mentioned: Primary interview subject; recognized as one of China's first-generation AI researcher-turned-entrepreneur founders; praised for strategic foresight, conviction, and the ability to hold a long-term thesis across 10 years.

"Over the past six years I've interviewed Cao Xudong four times... Momenta, after ten years, holds over 65% market share in the third-party high-end mass production autonomous driving segment." [00:02:02.910]

Sun Jian (孙剑)

Description: Legendary computer vision researcher; led the vision group at Microsoft Research Asia; co-creator of ResNet; later joined Momenta. Why mentioned: Described as profoundly influential on Cao Xudong's research philosophy — specifically instilling the value of rigorous experimentation and paying close attention to results that contradict expectations.

"Sun Jian's help to me was really enormous. His working style was extremely grounded, extremely focused on experiments, and especially on experimental phenomena that were very inconsistent with your expectations — those might be where the most significant discoveries lie." [00:01:04.230]

Wang Shengxing (王生星)

Description: Founder of Unitree Robotics. Why mentioned: Mentioned as someone Cao Xudong follows online and respects for pioneering work in Chinese robotics, though they haven't met personally.

"Unitree has done very pioneering work in China's robotics field. Have you talked with Wang Shengxing? No, but I often see him online." [00:35:24.160]

5. Operating Insights

Build R&D Milestones That Yield Positive Feedback Within Weeks, Not Years

Cao Xudong argues that any R&D system designed such that validation cycles take a year or more is fundamentally flawed. The architecture of the research program itself must be designed for short feedback loops — weeks to at most three months — otherwise the team loses direction and morale.

"Your R&D system design, the design of R&D problems, and the staging of R&D milestones must ensure that in short cycles — as little as one week to one month — you get positive feedback. At most three months. You can't design an R&D path where the feedback cycle is one year or three years. That R&D can't be done." [00:31:55.200]

The "GPT Bounty" Internal Incentive for AI Tool Adoption

Momenta created a formal internal rewards program in early 2023 where employees who used LLMs (initially GPT) in an innovative way to solve a real problem AND successfully rolled it out to at least one internal user received a monetary prize. This drove rapid internal adoption of AI tooling across engineering workflows.

"We had an internal bounty called the GPT Bounty — if you used GPT with an innovative method to solve an actual problem, and this method or tool wasn't just used well by yourself but was also deployed to at least one internal user, you could claim this prize." [00:47:05.340]

Vehicle Verification Platform (VVP) Powered by LLMs — From Rules to Agents

Their internal tool for vehicle sensor signal verification evolved from rule-based systems to LLM-powered agents. When a new vehicle variant is onboarded, engineers now query an agent rather than search for a human expert — the agent crawls data, applies historical pattern analysis, and surfaces root causes.

"We have a tool called VVP — Vehicle Verification Platform... Now we use large models to verify. So when a problem occurs, many times people don't go find the engineer — they just ask an agent directly. The agent crawls the data, and based on past experience, analyzes and tells you the root cause." [00:46:05.960]

The "Main Line" Razor: Avoid Smokestacks, Build Organizational Thrust

When facing a problem with multiple solution methods, prioritize methods that (a) reuse the existing main architecture, (b) build organizational convergence rather than divergence, and (c) are cumulative over time. Solutions that create parallel "branch lines" or "smokestacks" fragment organizational energy and should be strongly deprioritized.

"Main Line Razor — if you have many methods to solve the same problem, use the most concise one... Main Line Convergence — when solving problems, choose architectures, systems, and methods that allow the organization to form aligned force, rather than dispersed force." [00:44:36.960]

Low-Cost, Short-Cycle Hypothesis Testing Before Committing to Major Decisions

Momenta's operating discipline for major strategic decisions involves beginning to explore and discuss them years in advance, decomposing them into testable checkpoints at lower levels of abstraction, and only "deciding" on timing/entry point/pace rather than a binary yes/no — because the real decision has already been made through years of staged validation.

"When we truly commit to doing something, it's not a 'do it or not' decision. It's more: what timing, what entry point, what pace, what path. Because we may have been discussing and researching this for two to three years, even three to five years by that point." [00:51:51.120]

6. Overlooked Insights

The Physical AI Compute Investment Trajectory Signals Momenta Is Effectively a Foundation Model Company

Cao Xudong reveals a very specific compute spending roadmap: $200M this year, $400M next year, $600M in 2028 — all directed at training compute, not headcount. This trajectory, combined with the validation that one World Model serves ALL autonomous driving and will transfer to robotics, means Momenta is structurally becoming a physical world Foundation Model company, not merely an autonomous driving supplier. This is almost entirely unremarked upon in the conversation, but the implication is enormous: the correct comparable is not Mobileye or Bosch but something closer to a domain-specific version of OpenAI, whose moat compounds with every dollar of compute and every kilometer of driving data simultaneously.

"More importantly, the investment in data centers for training large models — this year maybe $200 million, next year $400 million, 2028 maybe $600 million... We believe this growth pace is enough to let us achieve ten-times improvement every year." [00:25:31.940]

The "Embodiment-Agnostic Pre-Training" Paradigm Shift Makes Momenta's Existing Data Archive a Surprise Robotics Asset

Cao Xudong briefly mentions that the Generalist robot approach validated using body-agnostic data for pre-training before fine-tuning with embodiment-specific data — improving success rates from 50% to 90%. The non-obvious implication: Momenta's massive autonomous driving dataset (accumulated across 100+ vehicle models at scale) is itself body-agnostic physical world data that can be used as robotics pre-training data. This gives Momenta a robotics data head-start that pure robotics startups — who must start data collection from scratch — cannot replicate, regardless of how much funding they raise. This connection is never explicitly stated in the interview.

"A representative company like Generalist uses large-scale data, and moreover body-agnostic data, for pre-training, then uses body-specific data for fine-tuning — pushing success rates from 50% up to 90%... Silicon Valley robotics companies are all moving in this direction and showing very good validation." [00:29:55.940]