The launch responds to rising demand for compact, power-efficient AI supercomputers capable of running foundation models directly at the edge, as general-purpose robots and autonomous machines move out of research labs and into commercial use across industries.
The Jetson T3000 delivers 865 FP4 teraflops of AI computing performance in a form factor about half the size and power consumption of the higher-end T5000, pairing an Nvidia Blackwell GPU with an eight-core Neoverse Arm CPU and 32GB of memory.
Nvidia said the module offers inference performance comparable to the T5000 on multimodal workloads while helping developers cut costs amid elevated memory prices.
The T2000, offering 400 FP4 teraflops and 16GB of memory, extends the Thor architecture to a wider range of systems, serving as an entry point for developers building visual AI agents, autonomous mobile robots and industrial manipulators.
Leading firms including 1X, Amazon Robotics, Boston Dynamics, Fanuc and Hitachi are already building systems on the Jetson AGX Thor platform, according to the company.
Nvidia said newly released Jetson agent skills allow developers to automate memory optimization, achieving in days savings that once took weeks, with early adopters such as Ubtech and Connect Tech trimming memory use by as much as 15GB.
The company also added Cosmos 3 Edge, a four-billion-parameter world foundation model, to its Thor lineup, enabling embodied systems to perceive their surroundings and predict actions in real time through on-device inference.
Copyright ⓒ Aju Press All rights reserved.
