NVIDIA introduced the Jetson T3000 and T2000 hardware modules to bring general-purpose robotics and foundation models from research labs into mass-market edge deployment. The company said the new devices, built on its Thor architecture, provide compact, power-efficient processing for autonomous machines. Robotics developers including 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot are currently building on the platform.
The Jetson T3000 module delivers 865 FP4 teraflops of AI compute in a physical package roughly half the size and power of the T5000. It features an NVIDIA Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory with 273GB/s bandwidth, and 25 GbE connectivity. NVIDIA also introduced the IGX T3000, which pairs identical compute performance with integrated functional safety features to run the NVIDIA Halos for Robotics full-stack safety system. According to NVIDIA, the T3000 matches the inference performance of the T5000 on multimodal workloads such as large language models and world foundation models.
The entry-level Jetson T2000 module provides 400 FP4 teraflops of compute and 16GB of memory. NVIDIA designed the T2000 for visual AI agents, autonomous mobile robots, and industrial manipulators. With the additions, NVIDIA's edge platform spans performance from 70 TOPS to 2,000 teraflops.
Software and Memory Savings
Software automation tools released alongside the hardware aim to trim deployment costs. NVIDIA released Jetson agent skills across its Thor and Orin product lines to automate memory optimization and system configuration. The company stated that the software reduces memory requirements in days rather than weeks, allowing applications to move to lower-memory hardware SKUs.
Several companies reported memory savings through the new software optimizations. Humanoid robotics companies UBTech and Agile Robots, along with hardware provider Connect Tech, reduced memory usage by up to 15GB, enabling a transition from the 64GB Jetson AGX Orin to a 32GB module. Smart retail firm SandStar cut memory usage by up to 4GB to deploy on the Jetson Orin NX 8GB module, while smart traffic provider NoTraffic reduced memory usage by 30% on the Jetson TX2 NX.
Model Integration and Launch Timeline
NVIDIA expanded its open foundation model lineup with Cosmos 3 Edge, a lightweight 4-billion-parameter model designed for Thor platforms. The model allows embodied systems to process vision, reason in real time, and generate actions through on-device inference. Developers can post-train Cosmos 3 Edge for specific hardware and sensor setups in about a day within the open Cosmos framework.
Engineers can begin software development immediately using the Jetson AGX Thor developer kit. NVIDIA said T3000 emulation mode will launch later this month alongside JetPack 7.2.1, with T2000 emulation mode following in a future release. Physical Jetson T3000 and T2000 modules are scheduled to ship in the first quarter of 2027. Ecosystem partners supporting the new hardware include ADLINK, Advantech, AAEON, Seeed Studio, Antmicro, and RidgeRun.
