Tech • AI • Robotics
China Has Already Won the Robot Race (and No One Is Talking About It)
9.4/10TL;DR
A new surge in physical AI is driving billions into robotics as companies shift from single-model humanoid dreams toward layered systems that can plan, predict, act and verify in the real world.
KEY POINTS
Investment boom accelerates
Robotics startups have raised more than $18 billion since January, already surpassing all of 2025 while the year is only halfway through. Agility Robotics is preparing a public listing at about $2.5 billion, Figure is valued at $39 billion, and a German robotics company recently secured nearly $1.5 billion in a single round. The capital surge reflects a wider belief that AI’s next major market will be physical rather than purely digital.
Why robotics is returning now
Earlier waves in robotics repeatedly fell short. Collaborative robot arms promised factory disruption around 2015, warehouse automation surged in 2020, and humanoids became a headline theme after Tesla Optimus appeared in 2021, yet many systems remained rigid and brittle outside tightly controlled settings. What has changed is the simultaneous progress of hardware and AI: lighter actuators, better embedded electronics and more controllable bodies now meet multimodal models that can connect vision, language and action.
From one brain to a chain of command
The industry is increasingly abandoning the idea that a single large model can simply be placed inside a robot. Instead, developers are building a cognitive stack that splits tasks across specialized layers: a planner to understand goals, a predictor to estimate what will happen next, an action model to move the body, and a verification layer to check whether the task actually succeeded. The emerging robot looks less like one brain and more like a staff of coordinated systems running at different speeds.
Why large language models are not enough
A core obstacle is the lack of a true “internet for movement.” Text models succeeded because they were trained on vast stores of written material, but physical skills are mostly tacit and rarely documented in precise, reusable form. Roboticist Ken Goldberg has noted that text and image data used for major AI models amount to roughly 100,000 years of human reading and viewing, while one of the largest robot-training datasets once contained only about one year of teleoperated human demonstrations.
The rise of VLA models
To bridge that gap, companies are extending vision-language systems into VLA models, short for vision-language-action. These systems take in a scene and a command such as picking up a cup, then generate the motor actions needed for a specific robot. Figure’s Helix architecture uses three levels, including a slow scene-understanding layer, a fast motion layer running at 200 Hz, and a whole-body coordination layer at 1000 Hz, while Google and Nvidia are developing similar action stacks through Gemini Robotics and Groot.
World models become the next battleground
Beyond choosing actions, robots must anticipate consequences: whether an object will slip, a container will overflow, or a trajectory will destabilize the body. That has pushed world models to the center of the field. Fei-Fei Li’s World Labs has raised $1 billion to build systems that can perceive, generate, reason and interact in three-dimensional environments, while Nvidia Cosmos aims to generate realistic futures that can supply synthetic training data and support robot simulation.
Two rival approaches emerge
One camp, led prominently by Nvidia, tries to generate possible future states of the world, effectively giving robots an imagination engine for training and planning. The other, associated with Yann LeCun and Meta’s JEPA approach, argues that predicting every future pixel wastes compute and that robots should instead learn abstract representations of physical structure. Meta showed this path can work by training a model largely on video and adapting it with about 62 hours of robotic data before using it to plan grasping and movement in new environments.
Google pushes high-level robotic reasoning
Google has separated motor execution from high-level planning through Gemini Robotics ER2, which acts as a supervisory brain rather than a direct controller of joints. It can break a task into steps, monitor progress over time, detect the moment a critical event occurs, and trigger retries only where needed instead of restarting an entire mission. Google has also demonstrated this orchestration across multiple robots, including Spot, Apollo, and Franka, suggesting that future robotic intelligence may coordinate fleets rather than inhabit one machine.
Nvidia’s advantage may be the tollbooth
Even if no single architecture wins, Nvidia is positioned across nearly every layer of the stack. It offers Cosmos for world modeling, Groot for robotic action, Isaac for training and simulation, Jetson for onboard execution, and the GPU infrastructure that powers most of the pipeline. That reach means many competitors may still depend on Nvidia somewhere between data generation, model training and deployment.
China’s edge may be manufacturing scale
While Google and Nvidia compete to control the cognitive stack, China may hold the strongest position in scaling the physical machines themselves. Robots require motors, reducers, batteries, sensors and supply chains, and those are areas where China already has depth. Unitree shipped about 5,500 humanoids in 2025, roughly a third of global volume, and plans a Shanghai listing to raise around 4.2 billion yuan, underscoring how mass production could become as decisive as AI architecture.
CONCLUSION
The new robotics race is no longer just about building smarter humanoids. It is about controlling the full stack of physical AI, from reasoning and simulation to components, manufacturing and deployment at scale.
Explain this