en.Wedoany.com Reported - Momenta, a company founded on autonomous driving technology, has filed a prospectus with the Hong Kong Stock Exchange, becoming a contender for the title of "first physical AI stock" in the Hong Kong market. The company previously held a significant market share in the intelligent assisted driving sector, with its system deployed in mass-produced models of several multinational automakers. Momenta's IPO process reveals its strategic direction of expanding from an autonomous driving company into the field of physical AI.
In the realm of physical AI, the "world model" is regarded as the core foundation model, yet the technical roadmap has not yet converged. Currently, there are four mainstream approaches: the generative video route represented by OpenAI Sora, which pursues pixel-level realism; the interactive world route represented by Google DeepMind Genie, which generates real-time interactive environments based on user actions; the spatial intelligence route advocated by Fei-Fei Li (World Labs), which views the world model as a generative and interactive 3D representation; and the Joint Embedding Predictive Architecture (JEPA) route championed by Yann LeCun, which predicts the next state of the world at an abstract representation layer to save computational power. While all these routes aim to understand the physical world, their paths differ.
Fei-Fei Li once used the example of "a cup placed on a table" to explain the essence of a world model: a model that truly understands the world should be able to render it from any angle, simulate the physical process of it being knocked over, and plan how a hand would pick it up. These three capabilities share the same underlying "simulator." Meanwhile, LeCun believes that large language models are essentially statistical pattern matchers and do not truly understand the physical world. AMI Labs, founded by LeCun after leaving Meta, and Fei-Fei Li's World Labs have both received substantial capital support.
Autonomous driving is seen as the earliest "touchstone" for world models. A world model predicts possible future world states based on imagined action sequences proposed by an agent, which naturally aligns with the "action → prediction → action" cycle of autonomous vehicles.
Momenta's physical AI solution, the R7 world model, has already achieved mass production and is first deployed in the SAIC Volkswagen ID. ERA 9X. This model has accumulated over 12 billion kilometers of real-world driving mileage, from which more than 100 million segments of "golden data" have been extracted. During training, the R7 model can repeatedly simulate rare real-world hazards and alter boundary conditions for "extra practice" to improve performance in uncommon scenarios. According to data from CIC灼识咨询, from March 2025 to February 2026, among third-party urban NOA suppliers in China, mass-produced vehicles equipped with Momenta's system achieved a market share of 65% in sales volume. In terms of growth rate, the fastest delivery of 100,000 units can be completed in less than 40 days.
Momenta CEO Cao Xudong positions the company as a "builder of physical AI foundation models." The technical architecture of its R7 world model is divided into three layers: the first layer is world model pre-training, which compresses physical laws and causal relationships into the model; the second layer is world model simulation, which conducts closed-loop testing of extreme long-tail scenarios; the third layer is reinforcement learning within the model, which iteratively tries and errors through a reward and punishment mechanism, reasoning within a virtual world.
The R7 model is not merely a "real-time vehicle-side model" or a traditional "foundation large model," but is regarded as a foundation model for the physical AI era, providing a basis for AI to cognize the real physical world. At this stage, autonomous driving is the highest-value scenario that allows data scaling and commercial scaling of physical AI to form a positive feedback loop.
The prospectus shows that Momenta's revenue grew from 743 million yuan to 2.413 billion yuan from 2023 to 2025, tripling in three years with an average annual compound growth rate exceeding 80%. Among this, technology development revenue increased to 1.445 billion yuan, while licensing revenue surged from 23 million yuan to 968 million yuan, multiplying 42 times in three years. Licensing revenue is Momenta's fee model for authorizing automakers to use its physical AI system, characterized by high marginal returns. This business model is considered the most ideal revenue model for autonomous driving startups.
Momenta's technical system follows a framework of "one flywheel, two legs," namely a data-driven core mechanism and two business lines: L2 mass-produced assisted driving and L4 fully autonomous driving. These two legs share the same software algorithm architecture, sensor suite, and world model. Currently, over 900,000 L2 mass-produced vehicles provide real-world driving data and commercial revenue to support the iteration of the world model. The iterated model is then deployed on L4 Robotaxis, operating in cities such as Shanghai, Suzhou, Munich, and Abu Dhabi.
Momenta's IPO provides a new benchmark for value assessment in the physical AI field: for autonomous driving companies, it requires evaluating whether they possess a multimodal foundation model; for startups directly targeting the "ultimate physical AI brain," they need to address issues of deployment channels and data closed loops. Momenta has become the first player to prove its business logic in terms of operational data and technical system, but whether its technical system can be transferred to other physical AI terminals, such as robots, remains uncertain.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
