NVIDIA expands Jetson Thor lineup with T3000 and T2000 AI

NVIDIA has introduced the Jetson T3000 and T2000 modules for robotics and edge AI systems, extending its Thor-based line for autonomous machines.

The modules target developers building robots, visual AI agents, industrial systems and other machines that run AI models locally rather than in the cloud. The launch expands the Jetson range to cover performance from 70 TOPS to 2,000 teraflops.

The Jetson T3000 is positioned as the more powerful of the two. NVIDIA said it delivers 865 FP4 teraflops of AI compute in a compact format while using about half the size and power of the T5000.

It combines a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth and 25 GbE connectivity, according to NVIDIA. An IGX T3000 variant adds functional safety support and runs the Halos for Robotics safety system for robots operating alongside humans.

NVIDIA said the T3000 is intended for multimodal inference workloads, including large language models, vision language models, vision language action models and world foundation models. It added that the module can deliver inference performance similar to the T5000 in those workloads while using a smaller footprint.

The Jetson T2000 is designed as a lower entry point in the Thor family, offering 400 FP4 teraflops of compute and 16GB of memory for developers working on autonomous mobile robots, industrial manipulators and visual AI agents, NVIDIA said.

Developer tools

Alongside the hardware, NVIDIA released Jetson agent skills, software tools designed to automate memory optimisation, system configuration and deployment tasks across the Jetson portfolio. The tools support both Jetson Thor and Jetson Orin devices.

NVIDIA said the software can cut optimisation time from weeks to days and help customers use lower-memory configurations without affecting performance. That matters as memory costs remain elevated and edge systems are often deployed under tight hardware constraints.

NVIDIA cited several companies as early users of the optimisation tools. UBTech, Agile Robots and Connect Tech reduced memory use by as much as 15GB, allowing a shift from Jetson AGX Orin 64GB to the 32GB module.

In retail, SandStar cut memory usage by up to 4GB, enabling deployment on the Jetson Orin NX 8GB module instead of the 16GB version, according to NVIDIA. In transport systems, NoTraffic reduced memory usage by 30% on Jetson TX2 NX, creating room for additional AI functions on the same hardware.

Companion robot maker GROOVE X is also using Jetson's mix of AI accelerators to redistribute workloads and reduce memory requirements, NVIDIA said. The company presented these examples as evidence that software changes can lower system costs as much as hardware changes.

Foundation model

NVIDIA also expanded its Cosmos 3 family with Cosmos 3 Edge, a 4-billion-parameter world foundation model compatible with Thor platforms. The model is intended for embodied AI systems that need to analyse surroundings, reason in real time and generate actions on-device.

Developers can adapt Cosmos 3 Edge for specific robots and sensors in about a day using NVIDIA's open Cosmos framework, the company said. After that, the model can be deployed on Jetson Thor for vision analysis and robot policy running on the device itself.

The launch reflects NVIDIA's broader effort to link hardware sales to a software stack for robotics, simulation and model deployment. That stack includes Isaac for robotics simulation and perception, as well as NVIDIA models such as Nemotron, Cosmos 3 and Isaac GR00T.

Industry uptake

Jetson AGX Thor is already being used by robotics groups including 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi and Techman Robot, according to NVIDIA. The list signals an attempt to establish Thor as a common computing base for both humanoid robots and industrial automation systems.

A broader group of hardware partners, including ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK and YUAN, is already offering Thor-based systems, NVIDIA said. Software partners including Antmicro, Neurealm, REBOTNIX and RidgeRun are set to provide emulation and migration support for customers moving to the new modules.

Developers can begin work with the existing Jetson AGX Thor developer kit and emulate the performance of the T3000 and T2000, since the modules share the same chip architecture and software stack across the Thor family. NVIDIA said T3000 emulation mode will be available later this month with JetPack 7.2.1, while support for T2000 emulation mode will arrive in a later release.

The Jetson T3000 and T2000 modules are scheduled to become available in the first quarter of 2027.