Nvidia launches Jetson Thor to power humanoid brains

For years, robotics engineers faced a brutal tradeoff: either offload heavy vision models to the cloud and pray the connection didn't drop mid-step, or settle for dumbed-down algorithms on device. NVIDIA wants to kill that compromise for good.

The company officially launched Jetson Thor, its next-gen edge computing platform purpose-built for physical AI and general robotics.

A server cluster packed into 130 watts

Built on the Blackwell architecture, Jetson Thor delivers up to 2,070 FP4 TFLOPS of AI compute, a massive 7.5x leap over the previous Jetson Orin generation. It also brings a 3.1x boost in CPU performance via a 14-core Arm Neoverse-V3AE chip and up to 128 GB of LPDDR5X memory, all running within a configurable 40 to 130-watt thermal budget.

Raw math is useless if sensor data gets bottled up. Thor tackles this with support for 4x 25 GbE networking and NVIDIA's Holoscan Sensor Bridge, letting cameras, radar, and lidar dump high-speed streams straight into GPU memory with ultralow latency.

On the software side, it runs popular reasoning and transformer models natively at the edge, including DeepSeek, Llama, Gemini, Qwen, Cosmos Reason, and domain-specific stacks like Isaac GR00T N1.5. No cloud tether required.

The Jetson AGX Thor developer kit is available now starting at $3,499, while the production-ready Jetson T5000 module starts at $2,999 in 1,000-unit quantities. NVIDIA also opened preorders for its DRIVE AGX Thor platform for autonomous vehicles, slated to ship in September.

Atlas and Digit step up to the plate

The biggest names in humanoids are already locking it in. Boston Dynamics announced it is integrating Jetson Thor into its Atlas robot to execute complex reasoning onboard. Agility Robotics also confirmed that its upcoming sixth-generation Digit warehouse robot will use Thor as its primary compute brain.

Research labs at Stanford, Carnegie Mellon, and the University of Zurich are tapping the hardware for perception and navigation models. CMU AirLab head Sebastian Scherer noted that computer vision has historically been too sluggish for real-time robot decision-making, but compute has finally caught up to handle nuanced, unstructured environments like search and rescue.

Why it matters

Physical AI cannot wait on cloud ping times. If an autonomous warehouse robot or surgical assistant loses connectivity for two seconds, it either freezes or breaks something. Jetson Thor brings multimodal reasoning directly to the edge, making true split-second autonomy viable outside pristine lab conditions.

For robotics teams, a $3,499 dev kit is not pocket change, but building custom compute carriers with consumer desktop GPUs wastes months of engineering time while bleeding battery life. Thor offers a battle-tested, high-bandwidth platform that plugs directly into the Isaac software stack and ROS workflows right out of the box.

Strategically, NVIDIA is locking down the physical AI runtime before rivals can even organize their silicon. By delivering the base model foundations with GR00T, the simulation sandbox, and now the onboard Blackwell silicon, NVIDIA is ensuring that almost every advanced bipedal robot walks on its stack. While 130 watts is still power-hungry for smaller drones, for full-scale humanoids, this looks like the new industry baseline.