Nvidia has introduced the Jetson Orin Nano 2, a compact edge computer intended to run language, vision and robotics models directly inside drones, robots and inspection systems. The company is positioning it as an entry-level route into physical AI, where power consumption and response time matter as much as raw model capability.
The board delivers a claimed 78 trillion operations per second, includes 8GB of memory and uses an eight-core Arm processor. Nvidia says smaller and medium-sized models can now reach accuracy levels that previously required much larger systems, making local inference more useful on machines with limited energy and space.
More inference inside a 15-watt envelope
Jetson Orin Nano 2 is rated at twice the inference performance of the existing Orin Nano Super. Nvidia attributes the improvement to newer Tensor Cores and greater memory bandwidth while retaining the same compact form factor. In a 15-watt mode, the company says the new board uses 40 percent less power while matching the older system’s performance.
Those figures are vendor claims rather than results from an independent benchmark in the source. Performance will depend on the model, software optimization and workload, while an efficiency comparison at matched performance does not establish the same advantage at every operating point.
The design is meant to keep perception and reasoning close to sensors. A delivery drone or mobile robot can process images and language without waiting for every decision to travel to a data center. Local execution can reduce latency and network dependence, but the source does not quantify real-world response times or battery-life improvements.
A model ecosystem moves toward the edge
The board runs Nvidia’s open Jetson software stack and its agent skills. Nvidia lists its Cosmos and Nemotron models alongside Gemma 4 and Qwen 3 as examples of language and vision-language systems that developers can deploy with memory-efficient inference.
Deepu Talla, Nvidia’s vice president for robotics and edge AI, argues that intelligence once limited to data centers can now operate in real time on entry-level Jetson hardware. That statement describes the company’s product thesis; it does not mean every frontier-scale model can fit within 8GB or preserve its original performance after optimization.
Nvidia says more than three million developers already use its robotics stack. A wide group of manufacturers and software companies is preparing carrier boards, customized software and reference designs around the new module. That ecosystem can shorten integration work, although the report does not provide availability, pricing or deployment volumes for those partner products.
Early adopters separate trials from deployment
Alphabet subsidiary Wing currently uses the Orin Nano Super in its drone fleet and plans to evaluate Orin Nano 2 for more responsive, energy-efficient perception. Wing has not announced a migration schedule or committed the new board to production delivery flights.
Matic Robots is taking a firmer step by adopting Orin Nano 2 for home-cleaning robots. Nvidia says the hardware will support conversational interaction, gesture detection, precision mapping and semantic understanding alongside autonomous cleaning. Those are intended capabilities, not independent measures of how reliably the finished robots will perform in varied homes.
Cognex and Doosan Bobcat are also among the companies Nvidia identifies as adopting or exploring the platform. The wording matters: evaluation can demonstrate technical interest without guaranteeing commercial scale.
Jetson Orin Nano 2 brings a clearer performance and power target to edge AI, but its impact will be determined by complete machines. Developers must fit models, sensors, software and thermal limits into systems that work continuously outside a laboratory. Nvidia has supplied a more capable computing component; the partner deployments will show whether that component makes physical AI meaningfully cheaper, faster or more dependable.
