A single statement from Jensen Huang has sparked another wave of excitement

Just after the New Year, NVIDIA CEO Jensen Huang made a major move. Recently, at the CES exhibition, Huang threw out his signature viewpoint: “The ChatGPT moment for robots is arriving,” and he believes that “without real-world data, embodied intelligence can only be an illusion.”

This statement instantly ignited the industry. A clear signal began to emerge: the robotics industry will bid farewell to the inefficient phase of 'single-task programming and reliance on real-world data,' entering a general-purpose explosion period centered on physical AI.

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

With Jensen Huang’s entry, a competition for infrastructure aimed at 'continuously acquiring high-quality interactive data to drive model iteration' has begun.

Jensen Huang is also joining the race in embodied intelligence.

At this year's CES exhibition, Jensen Huang officially launched NVIDIA's new generation of foundational models for robots, NVIDIA Isaac GR00T, specifically designed for embodied intelligence. Simultaneously, he introduced the NVIDIA Cosmos platform to support the development of physical AI, offering open models, massive datasets, and toolchains to establish a core technical foundation for the widespread implementation of general-purpose robots.

As a demonstration vehicle for the new technical solution, Jensen Huang brought a special 'guest' during his presentation — Reachy Mini. Due to its resemblance to WALL-E from 'WALL-E,' this robot is also called the 'WALL-E robot.'

During the interaction with 'WALL-E,' Jensen Huang allowed 'WALL-E' to learn, observe, and mimic human actions in a simulated environment, understand the relationship between actions, results, and feedback, and transfer these capabilities into the real world.

The video footage showed that after training in a simulated virtual environment, 'WALL-E' successfully performed the action of 'falling down and getting up' on a real wooden floor while maintaining balance. This demonstration by Jensen Huang was seen as evidence that a physics simulation platform can help robots quickly learn complex physical interactions, bridging the gap between digital twins and the real world.

Under the influence of the simulated environment and world models, Jensen Huang stated that this is equivalent to moving the 'training ground' into the 'digital world.' He believes that this approach not only allows for scalable construction of training environments but also constrains and calibrates the scenarios generated by the model to ensure they have 'physical' credibility, including combinations of lighting, materials, scenes, etc.

In summary, Jensen Huang believes that the future of AI is not just about supercomputers but must be closely integrated with the physical world, and virtual simulation is key to breaking the data bottleneck.

This perspective has prompted Jensen Huang to quickly make strides in the embodied intelligence field. At the CES exhibition, Huang announced collaborations with several U.S.-based robotics companies, including Apptronik, Agility Robotics, Figure, Boston Dynamics, and Sanctuary AI. As part of the collaboration with Sanctuary AI, NVIDIA is providing a computing platform, simulation tools, and other technical support to jointly advance the development of general-purpose humanoid robots.

Therefore, after winning the 'computing power competition,' NVIDIA is attempting to recreate a 'CUDA' in the embodied intelligence space.

The key to breaking through for Chinese players

The spark ignited by Silicon Valley has also rapidly spread to the East. Unlike NVIDIA’s focus on 'high-fidelity simulation + general models,' Chinese players are leaning towards a pragmatic path of 'real-world scenario-driven + vertical closed-loop.'

In October 2025, the Ministry of Industry and Information Technology (MIIT) released the 'Embodied Intelligence Data Collection and Annotation Standards (Draft for Comments),' which provided a guiding framework for the format, quality, and security of physical interaction data. This indicates that 'data standardization' has risen to the level of a national strategy. In response, many embodied intelligence companies have already taken action.

Among them, Zhiyuan Robotics released GenieSim 3.0, the first open-source simulation platform driven by a large language model, encompassing over 200 tasks and tens of thousands of hours of simulation datasets as open-source. Despite launching the open-source simulation platform, Zhiyuan Robotics still emphasizes the core role of real machine data and considers real-world data as the foundation for model training. At the same time, it leverages simulated data as a supplement for early testing and engineering iterations.

Yinhe General focuses on developing large embodied intelligence models using synthetic data, proposing a 'three-tiered large model system' comprising hardware, skill layers, and top-level large models. Yinhe General believes that the synergy between synthetic and real-world data is crucial. On one hand, simulated data is used for large-scale foundational learning; on the other hand, real-world data is applied to validate and enhance the model's adaptability in practical scenarios, ensuring that the model can learn quickly and land precisely to form a closed loop of 'simulation pre-training → real-world data fine-tuning → model optimization.'

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

Luming Robotics, one of the 'four小龙 (young dragons)' of embodied intelligence data, chose a 'lightweight handheld gripper' approach for data collection.

Luming Robotics founder Yu Chao believes this method was chosen because while simulations can generate millions of scenarios, only real machines can perceive dust, oil stains, and material aging found in the real world. In Yu Chao's view, the industry has long been trapped in a vicious cycle of real machine data collection—namely, 'high cost, low efficiency, and poor compatibility.' For instance, traditional teleoperation can collect only 30-35 pieces of data per hour at a high cost. Additionally, data from robotic arms of different brands and models cannot interoperate, meaning each collection effort adapts to only one specific unit, resulting in significant resource waste.

Against this backdrop, Luming has independently developed the FastUMI Pro system, which enables direct reuse of data from different brands of robotic arms through a unified gripper interface, force control module, and visual calibration solution. This means that a model trained on an automotive welding line can be fine-tuned for use in 3C assembly or logistics sorting.

Yu Chao believes that its core value lies in being able to摆脱 specific robot hardware dependencies, quickly adapting to dozens of robotic arms and grippers available on the market, truly breaking data silos, and enabling cross-platform data reuse. Compared with traditional data acquisition technologies, FastUMI Pro improves efficiency fivefold, reduces costs to one-fifth, and achieves an industry-leading precision of 1-3mm.

Regarding the crucial aspect of data acquisition, an investor stated, 'The essence of embodied intelligence investment is that it must ensure a high probability of success while leaving enough room for potential returns.'

Looking at the current players in embodied intelligence, their success rate depends on choosing a sufficiently pragmatic point of entry. Simply put, this means focusing on industrial scenarios where customers have a strong willingness to pay, tasks have clear boundaries, and ROI can be quantified, such as in 3C electronics, logistics warehousing and handling, quality inspection, and defect recognition.

'This is evident from the vertical fields that players are focusing on. For Luming Robotics, achieving a 60% compression in production line cycle times with Mitsubishi also demonstrates that its technology has passed real-world commercial validation and is no longer a lab demo. This is the most fundamental guarantee of success,' Yu Chao said.

The potential upside, however, lies in the paradigm of 'body + data' or 'data × model × body.' Once an embodied intelligence company's data collection solution is widely adopted by the industry, its value will no longer depend on how many robot hardware units it sells, but rather on how many robots operate and iterate within its data ecosystem.

'In terms of direction, Luming's FastUMI Pro is heading towards becoming the 'USB interface' of embodied intelligence.' In this regard, Yu Chao stated, 'We are not just building robots; we are building the infrastructure for embodied intelligence. Luming's goal is clear: accumulate data by operating real machines in scenarios, train better models, and provide the industry with two key infrastructures—data and hardware—to promote the co-construction of universal platforms and ecosystems.'

The Eve of the Hottest Financing Track

If we were to ask which would be the hottest financing track in 2025, the answer would likely be 'embodied intelligence.' Data shows that over the past year, the domestic industry's interest has continued to rise, with 298 financing events occurring during the year, a year-on-year increase of 144%; total financing reached 32.9 billion yuan, a year-on-year increase of 291%.

Behind this trend lies both the capital market's 'strong' interest in the embodied intelligence sector and bets on the emergence of 'AI + physical interaction' as a new phenomenon. Throughout this process, industrial capital has continued to invest heavily, becoming one of the most active investment forces in this space.

Taking JD.com as an example, it invested in three embodied intelligence companies in one day—Qianxun Intelligence, Zhuni Power, and Zhongqing Robotics—and has also set its sights on RoboScience and Pasini for 2025. The purpose of these strategic investments is clear: through investment, JD.com covers multiple links from the whole machine to key components, building an embodied intelligence ecosystem while promoting technology application in logistics, warehousing, and factory inspection scenarios.

CATL, as the leading company in the power battery sector, has also turned its attention to this field. Through investment, it penetrates into the industrial chain to provide power solutions for robots, promoting their application in industries such as manufacturing and logistics.

Furthermore, Meituan entered the robotics track as early as 2020 and has since invested in more than 10 robotics and embodied intelligence companies, including some industry leaders. Through these investments, Meituan explores application scenarios like local life services, delivery, and sorting, improving service efficiency and user experience.

Facing the hottest financing track, the above-mentioned investors’ most direct observation is that 2025 clearly leans toward two extremes: 'investing early and small' versus 'investing strong and premium.'

On one hand, within 'investing early and small,' seed rounds, angel rounds, and Series A financing events together account for 74% of the total, with a large amount of capital spread out early like gold prospectors, betting on the next potential unicorn. On the other hand, within 'investing strong and premium,' Series B and later rounds account for 15%, meaning that leading companies have further strengthened their ability to secure large-scale financing, with investors generally listing 'data acquisition capability' and 'scenario implementation validation' as core due diligence indicators.

As of now, the companies referred to as the 'four小龙 (young dragons) of embodied intelligence data collection,' despite differing technologies and approaches in data collection, all aim to capitalize on underlying technological benefits by breaking through high-frequency, essential, scalable application scenarios.

According to a report from the Gaogong Robotics Industry Research Institute (GGII), by 2025, the global market size is expected to reach 6.339 billion yuan, with China accounting for over 50%. It is projected that by 2030, global humanoid robot market sales will approach 340,000 units, with the market scale potentially exceeding 64 billion yuan.

This makes high-quality interactive data critical for the large-scale implementation of humanoid robots on the eve of the humanoid robot market boom—a crucial step that the industry must navigate.

10K Views