Overview
- NVIDIA introduced two new Jetson AGX Thor modules, the T3000 for humanoids and robotics and the T2000 for broader edge AI, as part of a push to move embodied AI from labs toward mainstream deployment.
- The T3000 is specified to deliver 865 FP4 teraflops with 32GB LPDDR5X memory and 25 GbE connectivity in a package NVIDIA says is roughly half the size and power of the T5000, while the T2000 is listed at 400 FP4 teraflops with 16GB of memory.
- NVIDIA released Jetson agent skills that automate memory optimization and system configuration, and the company cited partner results showing memory cuts of up to 15GB that let teams run workloads on lower‑memory, lower‑cost modules.
- Developers can emulate the T3000 this month using JetPack 7.2.1 and expect physical T3000 and T2000 modules in Q1 2027, with the lightweight 4‑billion‑parameter Cosmos 3 Edge model available to run on‑device; independent reports claiming different specs and earlier availability are disputed versus NVIDIA’s official announcements.
- The announcement builds on the Jetson family and Blackwell/Thor architecture and includes a broad partner ecosystem — from robot makers to board vendors — aiming to shorten sim‑to‑real development and cut deployment cost and power for real‑world robots.