NVIDIA is extending its Jetson Thor line with two smaller modules that keep the Blackwell GPU architecture while cutting size, power and memory to match the economics of volume robotics and edge AI. The T3000 and T2000 sit below the existing T5000 and give developers a clear path from high-end humanoids down to industrial manipulators and visual AI agents without leaving the same software stack.
The Jetson T3000 delivers 865 FP4 teraflops inside a module roughly half the size and power of the T5000. It pairs the Blackwell GPU with an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory and 273GB/s bandwidth plus 25 GbE connectivity. IGX T3000 adds functional safety hardware so the same performance can run NVIDIA Halos for Robotics when machines work next to people.
Despite the smaller footprint, the T3000 maintains inference throughput on multimodal workloads that include large language models, vision-language-action models and world foundation models. The reduction in memory capacity also lowers bill-of-materials cost at a time when high-bandwidth memory prices remain elevated.
The Jetson T2000 opens the Thor architecture to a wider set of edge systems. It supplies 400 FP4 teraflops and 16GB of memory, serving as the entry point for autonomous mobile robots, industrial manipulators and visual AI agents that do not need the full T3000 specification.
With these additions, NVIDIA now offers a single edge AI platform that scales from 70 TOPS to 2,000 teraflops. Developers can therefore match hardware to workload without switching ecosystems or rewriting application code.
New Jetson agent skills automate memory optimization, system configuration and deployment across the entire Jetson portfolio, including both Thor and Orin modules. Tasks that once required weeks of manual tuning can now be completed in days, allowing teams to move down one memory SKU while preserving performance.
UBTech and Agile Robots have each cut memory usage by up to 15GB and shifted from the 64GB Orin module to the 32GB configuration. SandStar achieved a 4GB reduction to run on the Orin NX 8GB module, while GROOVE X and NoTraffic reported similar gains that freed headroom for additional AI features without raising hardware cost.
The commercial signal is clear: NVIDIA is optimizing for the capital-efficient ramp of humanoid and mobile manipulator fleets rather than chasing peak benchmark numbers on a single flagship board. By releasing lower-power Thor variants alongside automated software tools, the company is removing the last practical barriers that kept foundation-model robotics in pilot mode.
Source: NVIDIA Blog.