Jensen Huang’s One Remark Sparks the Next Big Wave

Just after the new year, NVIDIA CEO Jensen Huang made a decisive move. At CES, he delivered a signature point of view: “Robotics’ ChatGPT moment is coming.” He also stressed that without real-world data, embodied intelligence is little more than an illusion.

That message immediately set the industry buzzing. A clear signal is emerging: robotics is poised to move beyond an inefficient era of single-task programming and heavy reliance on limited real-world datasets, and into a breakout phase centered on Physical AI and more general-purpose capabilities.

In fact, capital in China had already sensed the shift. In the domestic embodied-intelligence landscape, if 2024 was dominated by the keyword “hardware (the robot body)”, then by the second half of 2025 the focus had evolved into “hardware + data”—or a broader paradigm upgrade of “data × models × hardware.”

With Huang now fully in the mix, a new competition is accelerating: an infrastructure race to continuously acquire high-quality interaction data and use it to drive faster, tighter model iteration.

Jensen Huang enters the embodied intelligence race

At CES, Huang officially unveiled NVIDIA Isaac GR00T, a new-generation foundation model built specifically for robotics. At the same time, NVIDIA introduced the NVIDIA Cosmos platform to support Physical AI development—bringing together open models, large-scale datasets, and a full toolchain intended to form the core technical foundation for generalizable robotics in the real world.

To demonstrate the approach, Huang introduced a special “guest” on stage: Reachy Mini. Because of its resemblance to the character WALL·E, it’s often nicknamed the “WALL·E robot.”

In the on-stage interaction, Huang showed how the robot could learn in a simulated environment—observing and imitating human actions, then building an understanding of the link between action, outcome, and feedback—before transferring those skills to real-world settings.

In the demo footage, the robot trained in simulation and then successfully performed a “fall down → get back up” sequence on a real wooden floor while maintaining balance. Observers interpreted the demonstration as evidence that high-fidelity physical simulation can help robots learn complex interactions faster, narrowing the gap between digital twins and the real world.

With simulation plus world models, Huang described this as effectively moving the “training ground” into the “digital world.” The advantage, he argued, is not only the ability to build training environments at scale, but also to constrain and calibrate generated scenarios so they remain physically credible—including realistic combinations of lighting, materials, and environments.

The broader message was consistent: the future of AI isn’t only about supercomputers—it must be deeply connected to the physical world, and simulation is a key lever for breaking today’s data bottlenecks.

That belief is also shaping NVIDIA’s partnerships. At CES, Huang said NVIDIA is working with several U.S. robotics companies, including Apptronik, Agility Robotics, Figure, Boston Dynamics, and Sanctuary AI. In the collaboration with Sanctuary AI, NVIDIA provides computing platforms and simulation tools to help advance the development of general-purpose humanoid robots.

In other words, after winning the “compute race,” NVIDIA is now trying to build a robotics-era equivalent of CUDA—a foundational platform others will depend on.

The deciding factor for China’s breakthrough

The spark ignited in Silicon Valley is spreading quickly in the East. But compared with NVIDIA’s bet on high-fidelity simulation + general-purpose models, many Chinese players are leaning toward a more pragmatic path: real-scenario driven execution + vertical closed loops.

In October 2025, China’s Ministry of Industry and Information Technology released a draft for public consultation titled “Embodied Intelligence Data Collection and Annotation Specification”, providing a first framework to guide the format, quality, and security of physical interaction data. That move signaled that data standardization is rising to a national strategic level—and many embodied-intelligence companies have begun acting accordingly.

One example is Zhiyuan Robotics, which released an open-source simulation platform driven by large language models—GenieSim 3.0—covering more than 200 tasks and releasing a simulation dataset totaling tens of thousands of hours. Even while launching simulation as open infrastructure, Zhiyuan continues to emphasize that real-robot data is foundational for training, with simulation data serving as a complement for early testing and engineering iteration.

Galaxy General has taken a synthetic-data-driven approach to embodied foundation model R&D, proposing a “three-layer” system spanning hardware, skills, and a top-level foundation model. In its view, the collaboration between synthetic and real data is essential: simulation data supports large-scale learning of core capabilities, while real data validates and strengthens adaptability in practical scenarios—forming a loop of simulation pre-training → real-world fine-tuning → model optimization.

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

And as one of the so-called “four rising players” in embodied-intelligence data collection, Luming Robotics chose a “lightweight handheld gripper” approach for data capture.

Luming founder Yu Chao explains the logic simply: simulation can generate millions of scenarios, but real-world dust, oil residue, and material aging can only be sensed by real machines. In his view, the industry has long been trapped in a painful loop for real-world data collection—high cost, low efficiency, and low transferability. With traditional teleoperation, for example, one hour may collect only 30–35 data samples at a high cost. Worse, datasets often can’t be shared across robot brands or models, meaning a single collection effort may only fit one specific robot body—wasting resources.

Against that backdrop, Luming developed FastUMI Pro, aiming to enable data reuse across different robotic arms by standardizing gripper interfaces, force-control modules, and visual calibration methods. The implication is powerful: a model trained on an automotive welding line could, with modest fine-tuning, be adapted for 3C assembly or logistics sorting.

Yu argues the core value is reducing dependence on specific robot hardware—enabling fast adaptation across dozens of arm and gripper configurations, breaking data silos and allowing cross-platform reuse. Compared with traditional approaches, he claims FastUMI Pro can raise efficiency by five times, reduce cost to one-fifth, and achieve high precision in the 1–3 mm range.

On the investment side, one investor summed up the mindset: the essence of embodied-intelligence investing is to pursue both a higher probability of success and enough upside imagination.

Today, “probability of success” often comes from choosing pragmatic entry points—industrial scenarios with strong willingness to pay, clear task boundaries, and measurable ROI. Typical examples include 3C electronics, logistics warehousing and handling, and quality inspection and defect recognition.

Yu points to Luming’s reported results with a Mitsubishi production line—compressing cycle time by 60%—as proof of real commercial validation rather than a lab demo, forming a baseline for confidence.

The “upside imagination,” meanwhile, comes from the new paradigm of hardware + data (or data × models × hardware). If a company’s data-collection approach becomes widely adopted, its value may no longer depend primarily on how many robots it sells, but on how many robots run, learn, and iterate within its data ecosystem.

Yu describes FastUMI Pro as moving toward an “embodied intelligence USB interface.” The goal, he says, is not merely building robots, but building the infrastructure of embodied intelligence—accumulating real-world operational data, training better models, and providing the industry with foundational capabilities in both data and hardware, ultimately pushing toward a shared ecosystem.

On the eve of the hottest fundraising track

If you ask what the hottest fundraising sector of 2025 was, “embodied intelligence” would be a leading candidate. Data cited in the article indicates the domestic market heated up sharply over the past year: financing events reached 298, up 144% year over year; total funding規模 reached 32.9 billion RMB, up 291% year over year.

Behind this surge is both strong capital interest and a broader bet on the rise of AI + physical interaction. Industrial capital, in particular, stayed highly active.

JD.com, for example, reportedly invested in three embodied-intelligence companies in a single day—Qianxun Intelligence, Zhujidongli, and Zhongqing Robotics—and later in 2025 also targeted RoboScience and Pasini. The strategic goal is clear: cover multiple links from complete machines to key components, build an embodied-intelligence ecosystem, and accelerate deployment across logistics, warehousing, and factory inspection.

CATL, a leader in power batteries, has also turned attention to the space—seeking to penetrate parts of the supply chain by providing power solutions that help robots move into industrial and logistics applications.

Meituan began investing in robotics as early as 2020 and has since backed more than ten robotics and embodied-intelligence companies, including top-tier names—building scenario exploration in local services, delivery, and sorting to improve operational efficiency and user experience.

One investor’s most direct takeaway: in 2025, investment clearly polarized toward two ends—“invest early and small” and “invest strong and best.”

On the early end, seed, angel, and Series A deals accounted for 74% of all events, with capital casting wide nets for the next potential unicorn. On the other end, Series B and later rounds accounted for 15%, reflecting stronger fundraising power among leading companies. Investors increasingly treat data acquisition capability and validated scenario deployment as core diligence criteria.

So far, the companies described as the “four rising players” in embodied-intelligence data collection may be taking different routes—but they share the same ambition: break through a high-frequency, strong-demand, scalable scenario to capture the underlying technological dividends.

A report cited from the High-Tech Robotics Industry Research Institute (GGII) estimates that in 2025 the global market size will reach 6.339 billion RMB, with China accounting for more than 50%. By 2030, global humanoid robot sales are projected to approach 340,000 units, and market size may exceed 64 billion RMB.

That’s why—on the eve of a potential humanoid-robot boom—data, especially high-quality interaction data, is becoming the critical bottleneck for scaling real-world deployment. It is the essential link the industry must solve.