NVIDIA adds smaller Jetson Thor modules for robotics and edge AI systems

NVIDIA has introduced two new Jetson Thor modules aimed at moving robotics and edge AI systems from high-end prototypes into more compact, lower-power products. The company said the Jetson T3000 and T2000 are based on its Thor architecture and are designed for humanoid robots, autonomous mobile robots, industrial manipulators, visual AI agents and other machines that need to run modern AI models close to sensors rather than in a remote data center.

The higher-end T3000 is the more consequential part of the announcement. NVIDIA says the module delivers 865 FP4 teraflops of AI compute in a form factor roughly half the size and power of the existing T5000. The configuration combines a Blackwell GPU, an eight-core Arm Neoverse CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth and 25 GbE connectivity. A related IGX T3000 version adds integrated functional safety support for robots that may operate around people.

Why it matters

Edge robotics is increasingly constrained by memory, power and cost. NVIDIA is pitching the new modules as a way to run larger multimodal workloads, including language, vision-language, vision-language-action and world foundation models, without requiring the largest Jetson configuration for every deployment. The lower-cost T2000 broadens the same Thor architecture to systems that need 400 FP4 teraflops and 16GB of memory.

The launch also includes a software angle. NVIDIA released Jetson agent skills intended to automate memory optimization, system configuration and deployment tasks across the Jetson portfolio. The company cited customer examples where software optimization reduced memory use enough to move products to lower-memory modules, including up to 15GB in some humanoid and industrial robotics workloads and a 30% reduction for NoTraffic on Jetson TX2 NX.

• The Jetson platform now spans from 70 TOPS to 2,000 teraflops, according to NVIDIA.
• Cosmos 3 Edge, a 4-billion-parameter model for embodied systems, is being brought to Thor platforms for on-device inference.
• T3000 emulation mode is due later in July with JetPack 7.2.1, while T2000 emulation will follow later.
• Jetson T3000 and T2000 modules are scheduled for general availability in Q1 2027.

The announcement does not mean mainstream robots are suddenly inexpensive or simple to deploy, but it does show NVIDIA filling in the middle of its robotics compute stack. For developers and hardware partners, the practical question is whether the smaller Thor modules can preserve enough AI performance while reducing bill-of-materials pressure, energy use and integration complexity.

Sources

• blogs.nvidia.com

Cover photo by Lisha Dunlap on Pexels, used under the Pexels License.