Nvidia has introduced the Jetson T3000 and Jetson T2000 modules, designed to expand the capabilities of mainstream robotics and edge AI systems.
The launch was announced in a company blog by Chen Su, head of edge AI product marketing at Nvidia. The new modules, based on the Nvidia Thor architecture, are intended for deployment in applications that require compact, high-performance, and efficient AI processing at the edge.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The Jetson T3000 module provides 865 FP4 teraflops of AI compute and integrates an Nvidia Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s memory bandwidth, and 25 GbE connectivity.
Chen Su wrote that the T3000 is roughly half the size and power of the Jetson T5000 while delivering similar inference performance across workloads, including large language models and vision-language models.
The IGX T3000 matches this compute performance but adds integrated functional safety and runs the Nvidia Halos for Robotics full-stack safety system for robotic applications involving interaction with humans.
The Jetson T2000 module extends the Thor architecture to a broader set of use cases outside high-end robotics. It features 400 FP4 teraflops of compute and 16GB of memory.
Chen Su wrote that the T2000 is suitable for visual AI agents, autonomous mobile robots, industrial manipulators, and other edge AI deployments requiring lower memory and power.
With these additions, Nvidia’s Jetson edge AI platform now covers a compute performance range from 70 TOPS to 2,000 teraflops. This scalability is said to enable developers to address a wide variety of edge AI workloads in robotics, automation, and autonomous machines.
To support these capabilities, Nvidia has introduced Jetson agent skills intended to automate software optimisation, memory management, and configuration tasks.
According to Chen Su, these skills enable developers to reduce memory usage and system costs within days rather than weeks. This, in turn, allows more advanced workloads to run on lower-memory modules across the Jetson range, including Jetson Thor and Jetson Orin.
Chen Su wrote that companies such as FANUC, Hitachi, Techman Robot, 1X, Boston Dynamics, Amazon Robotics, and Agile Robots have adopted the Jetson AGX Thor platform for their robotics applications.
Hardware partners already providing Thor-based solutions include Seeed Studio, TZTEK, AAEON, NEXCOM Robotic Solutions, Advantech, ADLINK, Aetina, Auvidea, Twowin, AVerMedia, Connect Tech, Realtimes, JWIPC, ForeCR, and YUAN.
Software migration and emulation services will be provided by partners such as REBOTNIX, RidgeRun, Antmicro, and Neurealm.
Nvidia has also expanded its foundation model portfolio with Cosmos 3 Edge, a 4bn parameter model designed for embodied AI systems and compatible with Thor platforms.
Chen Su wrote that Cosmos 3 Edge enables real-time perception, reasoning, and action generation through on-device inference. Developers can post-train Cosmos 3 Edge on specific sensor and embodiment setups within about a day and then deploy these capabilities on Jetson Thor modules.
The new modules share chip architecture and the software stack with the rest of the Nvidia Thor family, which is intended to streamline development and migration.
Developers can access the Jetson AGX Thor developer kit now and will be able to use T3000 emulation mode with JetPack 7.2.1 later this month, with T2000 emulation support to be released subsequently.
Nvidia expects the Jetson T3000 and T2000 modules to become available in the first quarter of 2027.
Last month, Nvidia signed multi-year agreements with major South Korean firms including SK hynix, Naver, SK Telecom, Doosan Group, and LG Group. The deals focus on memory technology, expanding AI infrastructure, and supporting new AI “factories” for global AI computing needs.
