NVIDIA Launches New Jetson Thor Modules

NVIDIA Blog · rss · 2026-07-16

NVIDIA has announced the next-generation **Jetson Thor** modules for robotics and edge AI: the T3000 and T2000. ### Key Information - **T3000 / IGX T3000**: Based on the Thor architecture, it claims to deliver **865 FP4 TFLOPs** of compute. Paired with a Blackwell GPU, 8-core Arm CPU, 32GB LPDDR5X, and 25GbE, it targets humanoid robots, industrial robotics, and other real-time multimodal workloads. - **T2000**: Positioned for a broader range of edge AI scenarios, delivering **400 FP4 TFLOPs** and 16GB of memory, aimed at visual AI agents, mobile robots, robotic arms, etc. - NVIDIA highlights that the new modules achieve inference performance close to the T5000 in a smaller footprint and with lower power consumption, helping customers deploy on lower-memory SKUs. ### Software and Ecosystem - Introduced **Jetson agent skills** for automating memory optimization, system configuration, and deployment, which reportedly compresses what used to take weeks of manual optimization into just days. - The post lists multiple robotics companies and ecosystem partners already developing on or providing migration paths for Jetson Thor. - **Cosmos 3 Edge** has also been expanded to the Thor ecosystem, serving as a 4-billion-parameter world model/robotics foundation model capable of running on-device. ### Timeline - Developers can get started now using dev kits and simulation modes. - **T3000 / T2000 are expected to launch in Q1 2027**.

Original post →

More from Embodied

Embodied channel →