NVIDIA Jetson AGX Thor

2025-08-25 · Publisher : NVIDIA · Domain : Robotics

NVIDIA ships Jetson AGX Thor, an onboard computer with a Blackwell GPU and 128 GB of memory delivering up to 2,070 FP4 TFLOPS within 130 W; the developer kit starts at $3,499.

### Background On August 25, 2025, NVIDIA announced general availability of the Jetson AGX Thor developer kit and Jetson T5000 production modules. These onboard computers are designed to run generative AI models, including VLAs, directly on robots. ### Specifications - **Architecture**: NVIDIA Blackwell GPU and 128 GB of memory. - **Performance**: up to 2,070 FP4 TFLOPS within a 130 W envelope (manufacturer figures). - **Gains**: 7.5x higher AI compute and 3.5x better energy efficiency than Jetson Orin, per NVIDIA. - **Price**: developer kit from $3,499. - **Ecosystem**: Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta are among cited early adopters; 1X, John Deere, OpenAI and Physical Intelligence are evaluating it. ### Why it matters Robot foundation models must run onboard, with low latency and no network dependency. Jetson Thor gives humanoids and mobile robots the power to run several generative models at once. ### Key takeaway Thor completes NVIDIA's "three computers" strategy: train in the data center, simulate with Omniverse and Cosmos, run onboard with Jetson.

Tags : Puce, Calcul embarqué, NVIDIA, Blackwell