NVIDIA just built a “brain” that can power humanoid robots and physical AI — and it only costs $3,499

by

As an analyst, I’m excited to announce that NVIDIA has unveiled a revolutionary piece of technology they’re calling the Jetson AGX Thor, often colloquially known as simply “Jetson Thor.” This device serves as the brain for robots and is versatile enough to be utilized across various industries, such as agriculture, healthcare, logistics, manufacturing, and transportation. In simpler terms, I’m looking forward to seeing this technology bring advancements in robotics to these diverse sectors.

Though you might not be familiar with the concept of physical AI or NVIDIA’s robotics software suites, chances are you’ve come across videos featuring machines that run on earlier versions of the Jetson Thor.

Among the pioneering users of Jetson Thor are Amazon Robotics, Boston Dynamics, Caterpillar, and Meta. John Deere and OpenAI are currently assessing the robust artificial intelligence capacities built into this robot’s central processing unit (CPU).

Jensen Huang, CEO of NVIDIA, stated that we’ve constructed Jetson Thor specifically for the vast number of developers focusing on robotics, who are continually shaping and influencing the real-world environment through their systems,” implies a more natural and straightforward paraphrase.

Jetson Thor offers unprecedented power and energy savings, capable of operating several AI models for generation directly on the device. This makes it the ideal top-tier supercomputer, leading the era of AI in physics and overall robotics.

Jetson Thor is fueled by a NVIDIA Blackwell-powered GPU. This architecture is utilized by NVIDIA to manufacture AI processors like the B30A AI chip. Similarly, NVIDIA’s retail RTX graphics cards are based on the same Blackwell technology.

Jetson Thor developer kits can be purchased now for $3,499.

As an analyst, I can attest that Jetson Thor’s groundbreaking performance has empowered robots to execute tasks that once demanded the processing prowess of servers.

According to Peggy Johnson, CEO of Agility Robotics, using the robust real-time processing capabilities of Jetson Thor will significantly advance Digit’s performance by improving its immediate reactivity and enabling it to master a wider range of intricate skills, making it more versatile overall.

By utilizing Jetson Thor, we bring the cutting-edge AI technology for physical applications, aiming to enhance efficiency in warehouse and factory processes for our clients.

Due to the self-contained abilities of robots driven by Jetson Thor, artificial intelligence is making a significant advancement, as stated by NVIDIA.

What is physical AI?

The latest AI processing unit from NVIDIA is built to fuel real-world artificial intelligence devices. While many might associate AI with something like ChatGPT in 2025, it’s essential to understand that there are various types of AI, and ChatGPT is just one example used by millions daily.

According to NVIDIA CEO Jensen Huang, there are four waves of AI:

  1. Perception AI: Recognizing speech, identifying images, and similar tasks.
  2. Generative AI: Creating text and images — think ChatGPT, DALL-E, and Gemini.
  3. Agentic AI: Reasoning models performing tasks independently. Agentic AI is quite new. Microsoft recently teased that Windows will be agentic.
  4. Physical AI: Integrating AI with real-world devices, such as robots.

We’re well into the second wave now and entering the first parts of the Agentic AI wave.

Moving from one wave to another significantly boosts the computational requirements. Huang highlighted that implementing Agentic AI necessitates about 100 times more data tokens and computing power compared to the original expectations.

Sebastian Scherer, affiliated as an Associate Research Professor at Carnegie Mellon University’s Robotics Institute, discussed the impact of technological advancements in computing on the field of robotics.

The extent of what we can accomplish is limited by the available computational resources. In the past, there was a significant gap between computer vision and robotics due to the slow processing speeds for real-time decision-making in computer vision tasks. However, advancements in models and computing power have now made it possible for robots to handle more complex and nuanced duties.

Given the high requirement for computational power, it’s understandable why NVIDIA would strive to create a device as potent as the Jetson Thor.

I’m far from a robot engineer, so please understand that I’m oversimplifying things quite a bit.

As an analyst, I’d rephrase that statement to: I enable devices like robots and self-driving cars with the ability to sense their surroundings and respond appropriately, thus facilitating autonomous decision-making and movement.

In simpler terms, while it’s not overly challenging to design a machine or robot to repeat a task under consistent conditions, real-world situations rarely offer such uniformity.

Consider viewing artificial intelligence with a physical presence as an advanced, autonomous cleaning appliance, similar to a highly intelligent robot vacuum cleaner. As this device navigates your home picking up dirt, it may encounter different surfaces, obstacles, or find itself trapped in a corner. To make informed decisions about its next actions, this robot requires data.

The high-powered NVIDIA Jetson Thor is capable of operating complex machinery such as surgical robots, advanced farming equipment, and industrial manipulators. Unlike devices designed to distinguish between carpet and hardwood floors, these machines demand a higher level of motor precision and spatial awareness for their operations.

Enter the scene: Jetson Thor, boasting an impressive 7.5 times the processing power compared to its predecessor, Jetson Orin. Let’s quantify this robot brain a bit more. Jetson Thor can reach up to 2070 Floating-Point 4 Teraflops in AI computation and offers a generous 128GB of memory for efficient data handling.

NVIDIA clarified that the Jetson Thor offers compute capabilities previously available only on servers, providing significantly enhanced processing power for devices such as the Atlas robot from Boston Dynamics.

Read More

2025-08-26 23:40