NVIDIA Jetson T3000/T2000: Blackwell-Powered Thor Modules for Edge Robotics Scale

<cite index="61-2,61-4">On July 15, 2026, NVIDIA introduced the Jetson T3000 and T2000, new Blackwell-powered modules based on the Thor architecture designed for mass-market robotics and edge AI applications. The T3000 delivers 865 FP4 teraflops of AI compute, combining a 1536-core Blackwell GPU, eight-core Neoverse Arm CPU, 32GB LPDDR5X memory, and 273GB/s bandwidth. The T2000 delivers 400 FP4 teraflops with 16GB memory for entry-level edge AI systems.</cite>

<cite index="61-2,61-5">NVIDIA claims the T3000 achieves similar inference performance to the larger T5000 for multimodal workloads (LLMs, vision-language models, vision-language-action models) while occupying roughly half the size and power footprint. Both modules are expected to ship in Q1 2027. The architecture enables developers to run foundation models at the edge for robotics, autonomous machines, and industrial inspection without cloud connectivity overhead.</cite>

<cite index="61-2,61-3">Leading adoption includes 1X (Neo Humanoid), Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot. NVIDIA simultaneously released Jetson agent skills that developers can use to optimize software stacks and reduce memory overhead in days instead of weeks—with reported reductions of 10–15GB enabling cost-effective deployments across smaller Orin variants.</cite>

<cite index="62-4">For architects evaluating edge AI silicon, the T3000/T2000 expand NVIDIA's portfolio from 70 TOPS (Orin Nano) to 2,070 TOPS (T5000), enabling developers to scale from microcontroller-grade to humanoid-robot-class inference on a single platform. The agent-skills software integration signals NVIDIA's investment in closing the gap between frontier-model capability and real-world deployment cost, critical for robotics-as-a-service and industrial automation.</cite>