Will NVIDIA's new Jetson Thor T3000 and T2000 make robots accessible to everyone?

NVIDIA has expanded its offerings dedicated to robotics, edge AI, and autonomous systems with the new Jetson Thor T3000 and Jetson Thor T2000, modules designed to promote the widespread adoption of humanoid robots, intelligent industrial machines, and AI applications executed directly on the device.

The goal is to provide developers with a scalable platform capable of running foundational models, large language models (LLM), vision language models (VLM), and other multimodal workloads without relying on the cloud, with sizes and consumption compatible with systems aimed at the mass market.

The new Jetson T3000 derives directly from the high-end Thor platform and integrates an NVIDIA Blackwell GPU, an eight-core Arm Neoverse CPU, 32 GB of LPDDR5X memory, a bandwidth of 273 GB/s, and 25 GbE connectivity. The module offers 865 TFLOPS FP4 dedicated to AI calculations, packed in a form factor that requires about half the space and power compared to the T5000 model.

According to NVIDIA, the Jetson T3000 achieves inference performance very close to that of the T5000 in multimodal workloads, while consuming around 70 watts. This configuration also helps keep platform costs down, particularly at a time characterized by high memory prices.

For applications with lower requirements, the Jetson T2000 debuts, a solution aimed at visual AI agents, autonomous mobile robots, industrial manipulators, and other intelligent systems. The module provides 400 TFLOPS FP4, 16 GB of memory, and consumes around 40 watts, representing the entry point of the new Thor family.

With these products, NVIDIA expands its Jetson offering to cover a performance range between 70 TOPS and 2,000 TFLOPS, enabling developers to choose the most suitable platform based on the type of processing required.

Alongside the hardware, new Jetson Agent Skills are being introduced, software tools designed to automate tasks such as memory optimization, system configuration, and application deployment. The company explained that these features significantly reduce the time required to optimize the entire software stack from weeks to a few days.

The optimizations also allow for reduced memory consumption and the use of less expensive hardware configurations without compromising performance. NVIDIA cites several practical examples: UBTech, Agile Robots, and Connect Tech have reduced memory usage by up to 15 GB, moving from Jetson AGX Orin 64 GB to the 32 GB variant. In retail, SandStar has lowered memory requirements to 4 GB, while NoTraffic achieved a 30% reduction on the Jetson TX2 NX platform, leaving more room for new AI functionalities.

The company has also expanded the Cosmos 3 model family with Cosmos 3 Edge, a 4 billion parameter version compatible with the Thor platform. The model is designed to enable robots to analyze their surroundings, reason in real-time, and generate actions directly on-device through local inference. Developers will also be able to quickly adapt it to specific sensors and platforms, thereby narrowing the gap between simulation and real-world use.

Sharing the same hardware architecture and software stack across all Thor systems offers a uniform development path. NVIDIA emphasized compatibility with its software ecosystem dedicated to Physical AI, which includes Isaac, Nemotron, Cosmos 3, and Isaac GR00T, tools aimed at accelerating the development of next-generation robots and autonomous machines.

The emulation mode for Jetson T3000 will be available by the end of the month through JetPack 7.2.1, while the mode dedicated to Jetson T2000 will arrive with a later version. Commercial availability of both modules is expected in the first quarter of 2027, with support from numerous hardware partners already active in the Jetson ecosystem.