A single statement from Jensen Huang has sparked another craze.

Right after the new year, Jensen Huang, the CEO of NVIDIA, made a significant move. Recently, at the CES, Jensen Huang put forward his iconic view: "The ChatGPT moment for robots is coming," and he believes that "without real - world data, embodied intelligence can only be an illusion."

This statement instantly set the industry on fire. A clear signal began to emerge: the robotics industry will bid farewell to the inefficient stage of "single - task programming and real - data dependence" and enter a general - purpose explosive period centered around physical AI.

Previously, domestic capital had already sensed this change. Looking at the development of domestic embodied intelligence, if the keyword in 2024 was "body", then by the second half of 2025, the track had transformed into a paradigm upgrade of "body + data" or "data × model × body".

Coupled with Jensen Huang's entry, a battle for infrastructure to "continuously acquire high - quality interactive data to drive model iteration" has begun.

Jensen Huang Joins the Fray in Embodied Intelligence

At this CES, Jensen Huang officially launched the new - generation embodied intelligence foundation model NVIDIA Isaac GR00T specifically designed for robots, and simultaneously introduced the NVIDIA Cosmos platform to support physical AI development. By opening up the model, massive datasets, and toolchains, it builds a core technical foundation for the general - purpose implementation of robots.

As a demonstration carrier for the new technical solution, at the speech site, Jensen Huang brought a special "guest" - Reachy Mini. Since this robot looks very similar to Wall - E in "WALL - E", it is also known as the "Wall - E robot".

During the interaction with "Wall - E", Jensen Huang let "Wall - E" learn, observe, and imitate human actions in a simulated environment, then understand the relationship between action - result - feedback, and transfer these abilities to the real world.

The video footage shows that after training in the simulated virtual environment, "Wall - E" successfully completed the "falling - getting up" action on a real wooden floor and maintained balance. Jensen Huang's demonstration is regarded by the outside world as evidence that the physical simulation platform can help robots quickly learn complex physical interactions to narrow the gap between the digital twin and the real world.

Under the effect of the simulation environment and the world model, Jensen Huang said that this is equivalent to moving the "training ground" into the "digital world". He believes that in this way, not only can the training ground be built on a large scale, but also the situations generated by the model can be constrained and calibrated to have "physical" credibility, and this physical credibility includes combinations such as lighting, materials, and scenes.

In short, Jensen Huang believes that the future of AI is not only about supercomputers but also closely related to the physical world, and virtual simulation is the key to breaking the data bottleneck.

This view has enabled Jensen Huang to quickly gain momentum in the field of embodied intelligence. At the CES, Jensen Huang announced that NVIDIA has cooperated with several American robot companies such as Apptronik, Agility Robotics, Figure, Boston Dynamics, and Sanctuary AI. Among them, in the cooperation with Sanctuary AI, NVIDIA provides technical support such as computing platforms and simulation tools to jointly promote the research and development of general - purpose humanoid robots.

Therefore, after winning the "computing power battle", NVIDIA is trying to create another "CUDA" in the field of embodied intelligence.

The Key to Success for Chinese Players

The spark ignited in Silicon Valley has also quickly spread in the East. Different from NVIDIA's bet on "high - fidelity simulation + general model", Chinese players prefer the practical path of "real - scenario - driven + vertical closed - loop".

In October 2025, the Ministry of Industry and Information Technology released the "Specifications for Embodied Intelligence Data Collection and Annotation (Draft for Comment)", which for the first time put forward a guiding framework for the format, quality, and security of physical interaction data. This means that "data standardization" has risen to the national strategic level. In response, many embodied intelligence companies have taken action.

Among them, Zheyuan Robotics released the first open - source simulation platform driven by a large - language model, GenieSim3.0, which includes more than 200 tasks and tens of thousands of hours of open - source simulation datasets. Although it has launched an open - source simulation platform, Zheyuan Robotics still emphasizes the core position of real - machine data and believes that real - world data is the basis for model training. At the same time, it uses simulation data as a supplement for early testing and engineering iteration.

Galaxy Universal drives the research and development of embodied large models with synthetic data and proposes a "three - level large - model system" including the hardware layer, skill layer, and top - level large model. In the view of Galaxy Universal, the synergy between synthetic data and real data is very important. On the one hand, simulation data is used for large - scale basic ability learning; on the other hand, real data is used to verify and improve the model's adaptability in real - world scenarios, ensuring that the model can learn quickly and be accurately implemented, thus forming a closed - loop of "simulation pre - training → real - data fine - tuning → model optimization".

Tashizhihang focuses on human video data, expanding semantic coverage through large - scale human behavior videos.

As one of the "Four Leading Companies in Embodied Intelligence Data", Luming Robotics chooses the method of "lightweight handheld grippers" for data collection.

Yu Chao, the founder of Luming Robotics, believes that the reason for adopting this method is that simulation can generate millions of scenarios, but only real machines can sense the dust, oil stains, and material aging in the real world. In Yu Chao's view, in the past, the industry was trapped in a vicious circle of real - machine data collection, that is, "high cost, low efficiency, and low adaptability". Taking traditional teleoperation as an example, only 30 - 35 pieces of data can be collected per hour, with high costs. At the same time, the data of mechanical arms from different brands and models cannot be interchanged, and one - time collection can only be adapted to a single body, resulting in a large waste of resources.

Against this background, Luming self - developed the FastUMI Pro system. By unifying the gripper interface, force - control module, and visual calibration scheme, the data of mechanical arms from different brands can be directly reused. This means that a model trained on an automobile welding line can be slightly adjusted and used for 3C assembly or logistics sorting.

Yu Chao believes that its core value lies in the ability to get rid of the dependence on specific robot hardware, quickly adapt to dozens of mechanical arms and grippers on the market, truly break the data silos, and achieve cross - platform data reuse. Compared with traditional data collection technologies, the efficiency of FastUMI Pro is increased by 5 times, the cost is reduced to one - fifth, and the accuracy reaches the high - level industry standard of 1 - 3mm.

Facing the crucial link of data collection, an investor said, "The essence of investing in embodied intelligence lies in ensuring both a high probability of winning and leaving enough room for imagination in terms of odds."

Looking at the current players in embodied intelligence, the probability of winning for all players lies in choosing a practical entry point. Simply put, it means focusing on fields where industrial - scenario customers have a strong willingness to pay, clear task boundaries, and quantifiable ROI, such as 3C electronics, logistics warehousing and handling, quality inspection, and defect recognition.

"This can be seen from the fact that players are focusing on vertical fields. For Luming Robotics, the result of reducing the production line cycle time by 60% with Mitsubishi also shows that its technology has been verified in real - world business and is no longer a laboratory demo. This is the most basic guarantee of the probability of winning," Yu Chao said.

The room for imagination in terms of odds is that from the paradigm of "body + data" or "data × model × body", once the data collection solution of an embodied intelligence company is widely adopted in the industry, its value will no longer depend on the number of robot hardware carriers sold, but on the number of robots running and iterating on its data ecosystem.

"In terms of direction, Luming's FastUMI Pro is moving towards the 'USB interface' of embodied intelligence," Yu Chao said. "We are not simply making robots but building the infrastructure for embodied intelligence. Luming's goal is clear: to accumulate data through real - machine operations in scenarios, train better models, and provide the industry with two major infrastructures of data and hardware to promote the industry to jointly build a general - purpose body and ecosystem."

On the Eve of the Hottest Financing Track

If we ask which is the hottest financing track in 2025, the answer may well be "embodied intelligence". Data shows that in the past year, the temperature of the domestic industry has been rising continuously. The number of financing events within the year reached 298, a year - on - year increase of 144%; the total financing scale reached 32.9 billion yuan, a year - on - year increase of 291%.

Behind this is not only due to the "strong" interest of capital in the embodied intelligence track but also a bet on the rise of the emerging concept of "AI + physical interaction". In the process, industrial capital has continuously placed bets and has become one of the most active investment forces in this track.

Take JD.com as an example. It invested in three embodied intelligence companies in one day, namely Qianxun Intelligence, Zhujidongli, and Zhongqing Robotics. At the same time, in 2025, it "targeted" RoboScience and Pasini. The purpose of these strategic investments is also clear. Through investment, JD.com covers multiple links from the whole machine to key components. While building an embodied intelligence ecosystem, it promotes the application of technology in scenarios such as logistics, warehousing, and factory inspections.

CATL, as a leading enterprise in the field of power batteries, has also turned its attention to this field. Through investment, it penetrates into the industrial chain and provides power solutions for robots, promoting the application of robots in industrial, logistics, and other fields.

Moreover, Meituan entered the robotics track in 2020 and has invested in more than 10 robotics and embodied intelligence companies so far, including many leading enterprises. Through investment, Meituan explores application scenarios such as local - life services, delivery, and sorting, improving service efficiency and user experience.

Facing the hottest financing track, the most direct feeling of the above - mentioned investor is that in 2025, the investment clearly gravitates towards two ends, namely "investing in early - stage and small - scale projects" and "investing in strong and excellent projects".

On the one hand, in the "investing in early - stage and small - scale projects", the total number of financing events in the seed round, angel round, and Series A accounted for 74% of the total. A large amount of capital, like gold - diggers, is casting a wide net in the early stage, betting on the next potential unicorn. On the other hand, in the "investing in strong and excellent projects", the proportion of Series B and later rounds reached 15%, which means that the ability of leading enterprises to obtain large - scale financing has been further enhanced. Investors generally list "data acquisition ability" and "scenario implementation verification" as the core indicators for due diligence.

So far, although the four companies known as the "Four Leading Companies in Embodied Intelligence Data Collection" have different technologies and paths in data collection, they all try to connect with a high - frequency, essential, and scalable implementation scenario to reap the benefits of underlying technologies.

A report from the GGII shows that in 2025, the global market size is expected to reach 633.9 million yuan, with China accounting for more than 50%. It is estimated that by 2030, the global sales volume of humanoid robots will be close to 340,000 units, and the market size is expected to exceed 64 billion yuan.

This makes high - quality interactive data the key to the large - scale implementation of humanoid robots on the eve of the explosion of the humanoid robot market, and it is a link that the industry must break through.

This article is from the WeChat official account "Dongshisitiaojie Capital" (ID: DsstCapital), written by Chen Mei and published by 36Kr with authorization.