NVIDIA Unveils Jetson Thor Modules and Cosmos 3 Edge Model, Pushing AI Robotics to the Edge
Summary
NVIDIA unveils Jetson Thor T3000 and T2000 modules powered by Blackwell architecture, delivering up to 865 FP4 teraflops for edge AI robotics, alongside the Cosmos 3 Edge world model enabling real-time on-device inference, with new agent skills slashing memory usage by up to 15GB — modules set to ship in Q1 2027.
Key Points
- NVIDIA is launching two new Jetson Thor modules — the T3000 and T2000 — powered by the Blackwell architecture, targeting mass-market robotics and edge AI with compact, power-efficient designs delivering up to 865 FP4 teraflops of AI compute.
- New Jetson agent skills are enabling developers to dramatically reduce memory usage across the entire Jetson portfolio, with companies like UBTech, Agile Robots, and SandStar cutting memory consumption by up to 15GB, lowering hardware costs and speeding up deployment.
- NVIDIA is also releasing Cosmos 3 Edge, a lightweight 4-billion-parameter world foundation model optimized for NVIDIA Thor platforms, allowing robots to perform real-time vision analysis and on-device inference, with T3000 and T2000 modules set to become available in Q1 2027.