NVIDIA announced the T3000 and T2000 modules based on the Thor architecture, designed to power mainstream robotics and edge AI applications. These compact, power-efficient computers can run foundation models at the edge, enabling mass-market deployment of autonomous machines.
SynthePulse Insight · AI deep reading
Version 2 · 2 sources
NVIDIA introduces smaller, more affordable Jetson Thor modules T3000 and T2000, along with software optimizations, aiming to lower the barrier for edge AI deployment while maintaining performance, accelerating the adoption of physical AI applications like robotics. However, performance claims and availability timelines warrant cautious consideration.
On July 15, 2026, NVIDIA announced the Jetson Thor T3000 and T2000 modules based on the Thor architecture, designed to meet the needs of general-purpose robots and autonomous machines transitioning from labs to large-scale deployment. These modules are positioned as compact, energy-efficient AI supercomputers capable of running foundation models at the edge.
The T3000 module delivers 865 FP4 TFLOPS, featuring a Blackwell GPU, eight-core Neoverse Arm CPU, 32GB LPDDR5X memory, 273GB/s memory bandwidth, and 25GbE connectivity. The T2000 offers 400 FP4 TFLOPS and 16GB memory, targeting edge systems such as vision AI agents, autonomous mobile robots, and industrial robotic arms.
NVIDIA claims that the T3000 achieves performance close to the larger T5000 on multimodal inference tasks (including LLMs, VLMs, VLA models, and world foundation models), while being roughly half the size and power. If true, this would significantly lower the hardware barrier for edge AI.
NVIDIA also introduced Jetson Agent Skills, a suite of software tools designed to automate memory optimization, system configuration, and deployment tasks for Jetson Thor and Jetson Orin devices. The core value lies in reducing hardware requirements through software means.
According to NVIDIA, UBTech, Agile Robots, and Connect Tech reduced memory usage by up to 15GB through software optimizations, enabling migration from the Jetson AGX Orin 64GB module to the 32GB module. SandStar, GROOVE X, and NoTraffic also achieved memory reductions in retail, companion robots, and traffic systems.
These cases suggest that software optimization can reduce reliance on high-end hardware, thereby lowering costs. However, it's important to note that these figures come from specific partners and may not be universally applicable, nor have they been independently verified.
NVIDIA expanded its Cosmos 3 world foundation model family with Cosmos 3 Edge, a 4B-parameter model optimized for the Thor platform. This model targets embodied systems that need to process world information, perform real-time inference, and generate actions through on-device inference.
The introduction of Cosmos 3 Edge indicates NVIDIA's strategy to build a software ecosystem beyond hardware, offering a complete edge AI solution. However, there is no publicly available independent test data on the model's real-time performance on edge devices.
The T3000 and T2000 modules are expected to be available in Q1 2027. Developers can start developing on T3000 via simulation mode later this month using JetPack 7.2.1, with T2000 simulation support coming in future releases.
Simulation mode may not perfectly reflect actual hardware performance, so performance data obtained during early development may not fully represent final hardware capabilities.
NVIDIA states that Jetson AGX Thor has been adopted by companies such as 1X, Agile Robots, Amazon Robotics, Boston Dynamics, Fanuc, Hitachi, and Techman Robot. This indicates NVIDIA already has a foundation in the robotics edge AI market.
The new modules aim to further expand the market by attracting more small and medium-sized enterprises and emerging applications with smaller, more cost-effective solutions. However, competitors like Qualcomm and Intel are also investing in edge AI, and whether NVIDIA can maintain its lead remains to be seen.
This report is primarily based on NVIDIA's official blog and The AI Insider's coverage. Performance data, memory savings cases, and availability timelines are sourced from NVIDIA's statements and are classified as source_claim, not independently verified. Uncertainties exist regarding differences between simulation and actual hardware performance, and the generalizability of partner data.
NVIDIA aims to replicate its data center success in edge AI through the T3000/T2000 modules and software optimizations, but performance claims and ecosystem effects still require market validation. For developers, simulation mode offers early access, but final decisions should be based on actual hardware testing.
Primary report
Primary source
Also covered by 1 sources