Fresh off the release of the Jetson AGX Orin 3000 and 2000 last month, NVIDIA today introduced another robot computer targeting entry-level Edge AI—the Jetson Orin Nano 2.
Behind these successive hardware updates lies a unified industry trend: optimized and compressed frontier large models are no longer exclusive to cloud servers; rather, there is an urgent need to deploy them into real-world physical devices.
Image source: NVIDIA
Deepu Talla, Vice President of Robotics and Edge AI at NVIDIA, pointed out that the deployment of Physical AI has never been achieved by a single chip alone. The complete pipeline consists of three parts: training, simulation, and edge deployment. After model training is completed in the data center, extensive simulation tests are conducted in a simulated environment to avoid the risks and costs of real-world debugging, ultimately running the models on the local hardware of robots and drones.
Therefore, around this complete workflow, NVIDIA has open-sourced a series of foundation models, including Nemotron, Cosmos, and GR00T. For various development teams, the greatest value of open-source models lies in the ability to overlay their own industry data and business experience on top of the foundational base to create proprietary dedicated AI. This allows them to avoid complete reliance on closed-source solutions and gain more technological initiative.
Turning to the product itself, as a newly released entry-level device, the Jetson Orin Nano 2 retains the hardware form factor of its predecessor while achieving a significant iteration in internal computing power. It is equipped with an Ampere architecture GPU, an 8-core ARM CPU, and 8GB of LPDDR5x memory, with a memory bandwidth reaching 120GB/s and an INT8 computing power of 78 TOPS. The overall power consumption can be dynamically adjusted between 15W and 40W. Relying on improved Tensor Cores and higher memory bandwidth, its inference performance is directly doubled compared to the previous generation; if only the same performance as the previous generation is required, the overall power consumption can be reduced by 40%.
In terms of the software ecosystem, the Jetson Orin Nano 2 maintains form factor and software stack compatibility with the previous generation, while fully inheriting NVIDIA's unified computing architecture from the data center to the edge. This means that all frameworks and models running on data center GPUs can be seamlessly deployed across the Jetson product line. Developers can directly run the latest Small Language Models (SLMs) and Vision-Language Models (VLMs), such as Gemma 4 and Qwen 3.5, on the Orin Nano 2, even including the open models from the NVIDIA Cosmos and Nemotron series. The inference speed exceeds 20 tokens/s, which is sufficient for most robotic applications.
Regarding this, Deepu Talla stated that for existing Jetson users, the threshold for migration is not high. The hardware and software form factors, along with the entire software stack, remain compatible, meaning existing projects can be ported with just a few days of debugging.
Furthermore, he mentioned that previously, many teams working on high-end robots had to choose the relatively expensive Orin NX (16GB version) due to strict performance requirements. Now, with the introduction of the Orin Nano 2, as long as the memory (8GB) is sufficient, they can completely migrate their workloads to this more affordable board through software trimming. The performance remains unchanged, but the hardware cost per device can be significantly reduced.
A demo was showcased at the launch event: two Reachy Mini robots running simultaneously on a single Orin Nano 2, independently completing speech recognition, Vision-Language Model inference, and speech synthesis in an offline state. A computing board with restrained power consumption and cost can enable two robots to "see," "hear," and "respond" simultaneously, which will have a substantial impact on the cost structure and product definition methods across the entire robotics industry.
Deepu Talla emphasized that NVIDIA does not build robots itself; its role is to provide core technologies and bring products to market through hundreds of partners—those making carrier boards, hardware systems, customized software, and reference designs. Currently, over 10,000 companies are developing commercial and industrial robots based on the Jetson platform, and more than 3 million developers are active on NVIDIA's robotics technology stack.
Regarding availability, the Jetson Orin Nano 2 module and its corresponding development kits are expected to be officially available in the first half of 2027. Currently, many industry enterprises have begun to evaluate the practical deployment possibilities of this hardware. Among them, Wing (a drone delivery company under Alphabet), Matic (home cleaning robots), Cognex (machine vision), and Doosan Bobcat (construction equipment) are already using or planning to use the Orin Nano 2.