Momenta IPO后再访曹旭东：就是想做没有尽头的AI

Summary

This LateTalk interview revisits Cao Xudong around Momenta’s IPO window and frames the company as an AI organization moving from mass-production autonomous driving toward broader Physical AI. Cao argues that the high-end assisted-driving supplier market is concentrating, that Momenta’s advantage comes from an Autonomous Driving Data Flywheel, and that the same physical-world model stack can eventually support Robo One, Robotruck, Robotaxi, and home robots. The episode also turns into a founder-method interview: Cao traces his path through Microsoft Research Asia, SenseTime, early Momenta organizational pain, mainline architecture discipline, and Low-Cost Short-Cycle Validation.

Key Claims

• Cao expects third-party advanced-driving suppliers to consolidate to two or three firms in China and three or four globally; in his account, Momenta and Huawei together already hold roughly 90% share among third-party city NOA suppliers and have meaningful pricing power.
• He says advanced driving has moved from a fun toy toward a usable product, and predicts that by 2028 even L2++ city-driving systems may at least reach human-driver levels in driving ability and safety.
• IPO is presented less as a financing event than as a trust-building and brand event: Momenta wants consumers and investors to associate the company with safety, reassurance, and repeated product overdelivery.
• The source gives a financial and compute-investment path from 2026 strategic losses to 2027 breakeven and 2028 profitability, with training-resource spending described as roughly $200 million, $400 million, and $600 million across those years.
• Cao describes Autonomous Driving Data Flywheel as the original strategic trunk: data-driven architecture plus mass-produced assisted driving generates data, revenue, and customer pressure, while Robot applications form the second leg toward full autonomy.
• World Models are treated as the bridge from cars to robots: Cao says the R7 World Model showed that physical-world data can support one foundation model across mass-production driving, Robo One, Robotruck, Robotaxi, and later robots.
• Momenta plans to extend from autonomous driving into robotics, but Cao explicitly does not frame 2026 as the year of a formal robot launch; he instead points to a possible 2030 home-robot inflection as edge compute, body reliability, precision, cost, and supply chains improve.
• For the next five years, Cao ranks Scalable Robo opportunities as Robo One first, Robotruck second, and Robotaxi third; in China, he says Robotaxi should be done through ASG partnerships with ecosystem players such as Didi, AutoNavi / Gaode, and T3 Chuxing, not by Momenta operating its own fleet.
• The organizational story centers on Low-Cost Short-Cycle Validation: Momenta had to move from a loose research-institute style toward a customer-centered startup, use product value to discipline technical innovation, and design short feedback cycles rather than waiting years for validation.
• Cao says the first mass-production delivery around 2021-2022 exposed weak tooling and too many vehicle, actuator, sensor, chip, and domain-controller problems; later mainline reuse, VVP, internal agents, and GPT-style tooling compressed some new-vehicle delivery work from hundreds of people over more than a year to roughly ten people over three months.

Key Quotes

“Better AI, Better Life” - Momenta’s long-horizon mission frame.

“好玩的玩具” to “好用的产品” - Cao’s description of advanced driving’s product transition.

“一流工作很多时候不是想出来的，而是做出来的” - Cao’s lesson from Sun Jian’s experimental culture.

“相信和真的相信不一样” - Cao’s distinction between verbal agreement with data-driven methods and actually rebuilding systems around them.

Connections

• Momenta and Cao Xudong - the central company and founder-method case.
• Huawei and Tesla - comparison points for city NOA competition and FSD entering China.
• Microsoft Research Asia, Sun Jian, and SenseTime - Cao’s training path from research and experiments into AI product delivery.
• Unitree Robotics - robotics body-side contrast for Momenta’s brain-first route.
• Didi, AutoNavi / Gaode, and T3 Chuxing - Robotaxi ecosystem partners named for the Chinese ASG cooperation model.
• Autonomous Driving Data Flywheel, Physical World Data Flywheel, World Models, and Physical AI - technical and strategic concepts extended by the source.
• AI Organization Design, Fast Feedback Loops, and Low-Cost Short-Cycle Validation - organizational method and execution concepts.
• Robotaxi Economics and Autonomous Vehicle Safety Benchmark - adjacent mobility economics and safety-evaluation context.

Contradictions

• No direct contradiction found. The episode creates a useful tension with prior robotics sources that say robots lack a Tesla-like fleet data loop: Cao’s claim is narrower and source-scoped, arguing that Momenta’s autonomous-driving data and world-model stack may transfer into adjacent robot businesses, not that general household robots already have a mature fleet flywheel.