Nvidia has announced general availability of the Jetson AGX Thor, its next-generation on-robot computer built on the Blackwell architecture. The developer kit sells for $3,499; the production Jetson T5000 module is $2,999 per module for orders of more than 1,000 chips. Nvidia says Thor delivers 2,070 FP4 teraflops and carries 128GB of memory within a 130-watt envelope — up to 7.5 times the AI compute, and 3.5 times the energy efficiency, of the previous Jetson Orin generation.

The pitch is that a robot can now run multiple generative AI models in real time at the edge. Nvidia lists a long roster of early users — Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic, and Meta among them — with 1X, John Deere, OpenAI, and Physical Intelligence named as evaluating it. The platform is compatible with Nvidia’s Isaac software stack, Metropolis for vision AI and Holoscan for real-time sensor processing.

CEO Jensen Huang calls Thor “the ultimate supercomputer to drive the age of physical AI and general robotics.” His robotics chief, Deepu Talla, stresses that Nvidia’s role is to arm the industry rather than build the robots itself. The combined automotive and robotics unit booked $567 million in quarterly sales in May, up 72% year over year. Robotics remains about 1% of Nvidia’s total revenue.