NVIDIA Launches Jetson Thor T3000/T2000 Modules for Mass-Market Robotics and Edge AI
Tags AI · Infrastructure · Enterprise
NVIDIA introduced Jetson and IGX T3000 (865 FP4 TFLOPS, 32GB LPDDR5X, 273GB/s bandwidth, 25GbE) and T2000 (400 FP4 TFLOPS, 16GB) modules based on the Thor/Blackwell architecture, targeting humanoid robotics, autonomous mobile robots, and industrial manipulators. T3000 matches T5000 inference performance at half the size/power. Cosmos 3 Edge (4B parameters) enables on-device world modeling. Developer kits available now with emulation modes; modules ship Q1 2027. Adoption by 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, Techman Robot.
Technical significance
Thor brings datacenter-class Blackwell GPU architecture to a compact, functionally safe (IGX) form factor, closing the gap between prototype and deployable robotics. The 865 TFLOPS in T3000 enables running multimodal VLMs, VLAs, and world models on-device — critical for latency-sensitive autonomy. Agent skills automating memory optimization (15GB savings reported by UBTech/Agile) reduce BOM costs by enabling lower-memory SKUs. Q1 2027 availability means design wins now determine 2027-2028 robot fleets.