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 1 · 1 source
The T3000 and T2000 modules deliver near-flagship inference performance in a smaller footprint and lower power, paired with software optimization tools to reduce robot deployment costs and accelerate physical AI from lab to mass market.
On July 15, 2026, NVIDIA released the Jetson T3000 and T2000 modules based on the Thor architecture, designed to meet the needs of general-purpose robots and autonomous machines moving from lab to large-scale deployment. The T3000 module integrates a Blackwell GPU, an 8-core Neoverse Arm CPU, 32 GB LPDDR5X memory, and 273 GB/s bandwidth, delivering 865 FP4 TFLOPS at roughly half the size and power of the T5000. Despite the smaller footprint, the T3000 achieves inference performance close to the T5000 on tasks such as large language models and vision-language models, helping reduce costs in a high-memory-price environment. The IGX T3000 version adds functional safety support and can run the Halos full-stack safety system.
The T2000 module offers 400 FP4 TFLOPS and 16 GB memory, positioned as a broader entry point for edge AI systems, covering visual AI agents, autonomous mobile robots, and more. With these additions, the NVIDIA Jetson platform now spans performance from 70 TOPS to 2000 TFLOPS.
NVIDIA introduced Jetson Agent Skills, which use AI agents to automate tasks such as memory optimization and system configuration, reducing development cycles from weeks to days. This tool supports the entire Jetson product line, including Thor and Orin series. Multiple companies have achieved significant memory savings: humanoid robot companies UBTech and Agile Robots reduced memory usage by up to 15 GB, enabling migration from 64 GB modules to 32 GB modules; smart retail company SandStar saved 4 GB, allowing use of 8 GB modules; smart traffic company NoTraffic reduced memory usage by 30% on Jetson TX2 NX, freeing up space for new AI features.
These optimizations lower system costs and deployment barriers, enabling developers to run more complex workloads on lower memory configurations. NVIDIA emphasizes that Agent Skills make Jetson an "agent-ready" physical AI platform.
NVIDIA also released the Cosmos 3 Edge model, a lightweight 4B-parameter version of the Cosmos 3 series, specifically optimized for the Thor platform. This model supports embodied systems in performing real-time visual reasoning, action prediction, and policy generation on-device. Developers can use the Cosmos framework to post-train the model for specific sensors and morphologies in about a day, narrowing the sim-to-real gap. The launch of Cosmos 3 Edge aims to bring cutting-edge world model capabilities to edge devices, supporting real-time decision-making.
The new modules share the same chip and software stack with the existing Thor architecture, providing a seamless development path. Developers can now simulate T3000 and T2000 performance using simulation mode on the Jetson AGX Thor developer kit (available through channel partners). T3000 simulation mode will be available with JetPack 7.2.1 at the end of July 2026, with T2000 simulation mode to follow. Actual modules are expected to launch in Q1 2027.
NVIDIA listed multiple partners, including hardware manufacturers ADLINK, Advantech, AAEON, and software partners Antmicro, Neurealm, providing simulation and migration support.
This article is primarily based on an official NVIDIA press release; information has been confirmed by the company but lacks independent third-party verification. Performance data (e.g., FP4 TFLOPS) and memory savings cases come from statements by NVIDIA and its partners and may involve selective disclosure.
NVIDIA aims to lower edge AI deployment costs while maintaining high performance through the T3000/T2000 modules and software optimizations, but actual effectiveness and ecosystem adoption remain to be verified after product launch in 2027.
Primary report
Primary source