Nvidia's new "brain" is here. Why are big tech companies scrambling to develop "brains" for robots?

Top-tier "shovel sellers" not only sell products but also set industry standards and build ecosystems. They attract third-party developers and service providers to develop around their "shovels," creating a powerful synergistic effect. For example, Microsoft's Windows operating system and Apple's App Store are exemplary cases of ecosystem building.

After installing NVIDIA's new robot computing platform, Jetson Thor, humanoid robots might as well say, "I've got it."

After a high-profile preview, NVIDIA finally arrives with a new brain for humanoid robots. Its AI performance has skyrocketed by 7.5 times, and the computing power has soared to 2070 TFLOPS. Such explosive data has made many people exclaim that this is the "iPhone moment" for humanoid robots.

Image source: NVIDIA

Putting the specific data aside for now, from the final result, NVIDIA's new brain, Jetson Thor, has broken through physical intelligence. Humanoid robots can directly predict the consequences of actions through generative reasoning and execute complex tasks. In simple terms, humanoid robots understand the physical world and work processes. Given a tool and materials, they know what to do next.

In addition to being able to understand, reason, and work autonomously, with the support of this brain, the overall working speed of humanoid robots has become smoother. Wang He, the CTO of Galaxy Universal, once revealed that the movement speed and fluidity of its G1 Premium robot have been significantly improved after adopting Thor.

The Embodied Learning Club once had a close experience of a robot equipped with Thor that mastered the "drifting" skill, and its movement speed seemed to be on fast forward. When carrying boxes, it could accurately identify the position of the boxes even after being interfered with by the Embodied Learning Club.

This is the "enlightenment" turning point for humanoid robots.

NVIDIA's active layout in embodied intelligence not only aligns with Jensen Huang's optimism about future physical AI and fits NVIDIA's role as a "shovel seller" but also precisely taps into another main theme: big tech companies are vying to create brains for robots.

Whether it's Tencent's general external brain launched under the slogan of "not touching hardware" or JD.com's empowerment of robots with "dialogue" capabilities, big tech companies have all quietly set their sights on the brain. The reasons behind this include both strategic layouts to make up for industrial shortcomings with their advantages and defensive strategies to avoid the risks of large-scale investment.

The championship seat in this "brainpower competition" for humanoid robots may be reserved for big tech companies.

NVIDIA's Pursuit of Both

Let's first look at a set of somewhat obscure data to see how NVIDIA pursues both the performance advantages of hardware and the systematic upgrade of software.

Jetson Thor is specially designed for generative reasoning models. Based on the latest Blackwell GPU architecture, it provides a computing performance of up to 2070 FP4 TFLOPS, which is 7.5 times higher than the previous generation, Jetson Orin. The CPU performance has increased by 3.1 times, energy efficiency by 3.5 times, and memory capacity by 2 times. Compared with 10 years ago, the AI performance has increased by as much as 7000 times.

Image source: NVIDIA

This robot computing platform can perform real-time processing of high-speed sensor data and execute visual reasoning at the edge, running multiple generative AI models simultaneously on edge devices. It solves one of the most significant challenges in the field of robotics: running multiple AI workflows, enabling robots to interact with humans and the physical world in real-time and intelligently.

In addition, Thor can give the first token response within 200 milliseconds and generate more than 25 tokens per second. This speed can almost support real-time conversations between robots and humans.

This means that humanoid robots are expected to no longer rely solely on the cloud and can adapt to and complete tasks in complex scenarios at the edge.

Jetson Thor has also made a lot of efforts in software to meet the low-latency and high-performance requirements of real-time applications and support all mainstream generative AI frameworks and AI reasoning models, including Cosmos Reason, DeepSeek, Llama, Gemini, Qwen, etc.

In simple terms, with the simultaneous upgrade of hardware and software, NVIDIA is no longer just installing a chip in a robot but giving the robot a brain capable of real-time interaction and reasoning, which is also fast and stable.

The starting price of this brain in the United States is $3499. The unit price for purchasing more than 1000 mass-produced modules is $2999. It will be available for sale to global customers, including those in China, starting today. The price in China has not been announced yet.

However, among domestic companies, Ubtech, Galaxy Universal, Unitree Robotics, Zhongqing Robotics, and Zhiyuan Robotics have all taken the lead in using Jetson Thor. In addition to Galaxy Universal mentioned above, Ubtech also said that its new generation of industrial humanoid robot, Walker S2, has deployed NVIDIA Isaac Sim and Jetson AGX Thor.

Image source: Ubtech

This brain is not entirely designed for humanoid robots. NVIDIA said that Jetson Thor is expected to speed up various robot applications, including surgical assistance robots, intelligent tractors, delivery robots, industrial robotic arms, and visual AI agents.

And this once again confirms Jensen Huang's words that in the future, all moving objects in the world will be robots.

Can Only Big Tech Companies Make Brains?

The brain has always been regarded as the top-level competition in embodied intelligence.

Without the support of brain capabilities, no matter how superior the body structure is, a robot cannot truly integrate into the physical world and interact directly with humans, production, and life. It can't even be considered a toy but only a "figurine." However, the meaning of top-level competition not only represents its high value but also the high degree of difficulty.

After all, the components of a robot's body can be handmade, but the difficulty of making a brain by hand has increased by several orders of magnitude.

This has naturally set a high threshold for entering the brain competition. Not only does one need technology but also financial support. This has also destined that at this stage, the brain track has become a "playground for giants." Previously, when the Embodied Learning Club had a conversation with the founder of a body manufacturer, the founder mentioned that the brain companies that can survive in the future are those that have raised a lot of money. Another founder also held a similar view, saying that big tech companies will ultimately have the say in the brain field because, at this stage, the book funds of body manufacturers are not enough to support the R & D of brains.

On the one hand, the high entry threshold has made the brain a niche market within a niche. On the other hand, doing a good job in the brain can not only bypass the intense competition at the body level but also promote the co - prosperity of the industry and seize the future voice. Therefore, big tech companies that already have a technological foundation have naturally become brain suppliers.

During the 2025 World Artificial Intelligence Conference, Tencent's Robotics X Laboratory and Futian Laboratory jointly released the embodied intelligence platform Tairos. This platform features "plug - and - play" to lower the entry threshold for users. It consists of two parts: model algorithms and cloud services. The model layer includes multi - modal perception models, large planning models, and large perception - action joint models. The cloud service platform includes a simulation platform, a data platform, and development tools.

Image source: Tencent

The Unitree robot equipped with Tairos can understand and execute the instruction "Come to me," see and describe the objects on the table clearly. It even has the ability of spatio - temporal memory and can remember what it was doing yesterday.

JD.com mainly provides the brain and IO input - output interaction capabilities. The IO capabilities include voice understanding in complex noisy environments and accurate understanding of children's language expressions. Robots that have accessed JoyInside include the SenseTime Yuanluobo AI chess - playing robot, the Lingtong Nian NIA - F01 humanoid robot, and the Zhongqing PM01. With the support of JoyInside, robots can have natural multi - round conversations with humans.

It is worth noting that Tencent and JD.com, which have entered the brain field, do not show the "merchant" nature. One talks about non - commercialization, and the other doesn't care about revenue. This is fundamentally different from NVIDIA.

Whether it's the strategic layout of Tencent and JD.com or NVIDIA's almost all - in attitude towards robots, at this stage, big tech companies' involvement in making brains is not only in response to the call of the times but also to seize the opportunity to strengthen their role as "shovel sellers." After all, the potential of this area is far more extensive than making a single robot.

Top-tier "shovel sellers" not only sell products but also set industry standards and build ecosystems. They attract third-party developers and service providers to develop around their "shovels," creating a powerful synergistic effect. For example, Microsoft's Windows operating system and Apple's App Store are exemplary cases of ecosystem building.

However, this also means higher technical requirements. A qualified "shovel seller" must have core technologies that competitors cannot easily replicate in the short term. This could be patents, complex algorithms, or unique manufacturing processes. They must also maintain forward - looking R & D investment. They must look further ahead than the "gold diggers" and plan for the next - generation "shovels" in advance. When the gold diggers are still using shovels, they are already researching excavators. Continuous R & D investment is the lifeline for maintaining a leading position.

From an industrial perspective, body enterprises need the empowerment of big tech companies to accelerate the process of robots becoming productive forces through smarter brains. This is a development path of mutual achievement. With the efforts of various participants in different roles, perhaps we are getting closer to general intelligence.

This article is from the WeChat official account "Embodied Learning Club," author: Lü Xinyi, editor: Di Xintong. It is published by 36Kr with authorization.