NVIDIA announces Jetson Orin Nano 2, touting 78 trillion operations per second for inference

NVIDIA announced the Jetson Orin Nano 2, a robot computer for edge AI. The company says it delivers AI computing performance of 78 trillion operations per second and twice the inference performance of the previous-generation Jetson Orin Nano Super. The Jetson Orin Nano 2 module and developer kit are scheduled to be available in the first half of 2027.

For designers of edge devices, NVIDIA’s claim that power consumption is 40% lower at equivalent performance offers a point of comparison when integrating inference into power-constrained robots and drones. However, actual suitability depends on the models and workloads being run, as well as the device’s power requirements, so the announced performance figures alone do not establish the benefits for specific applications.

The 78 trillion operations per second and twice the inference performance describe the product’s capabilities. But the business focus of the announcement is on putting that computing capacity into compact devices and using it for real-time perception, interaction and movement. The design combines 8GB of memory with a 40% reduction in power consumption at equivalent performance in 15W mode, addressing the power and installation constraints of edge devices. However, these figures come from NVIDIA, and the announcement does not specify the comparison conditions for performance with individual models or real-world tasks. The product builds on the existing Jetson stack, continuing the expansion of the developer platform. But the companies cited differ in how they describe their use cases. Matic is moving ahead with adoption in cleaning robots, while Wing plans to evaluate the product for a next-generation device. The announcement groups Cognex and Doosan Bobcat as companies adopting or considering it. The companies therefore should not be understood to have reached the same implementation stage at the time of the announcement. Wing also said its existing aircraft use the Jetson Orin Nano Super, so it is necessary to examine individually which functions and power requirements the product update will address. The division of roles with the model layer is another consideration. Runway releases “Praxis-1,” connecting video pretraining to robot control covered the release of a model for robots, while the Jetson Orin Nano 2 is a computing platform for running inference on physical devices. Releasing a model alone does not enable it to operate in the real world; evaluation must also cover execution under memory constraints, sensor-input processing, latency and power consumption. NVIDIA’s announcement, however, did not provide measurements of a robot running a specific model. Going forward, points to watch include pricing, a specific availability date, measured performance by application, power requirements and when the companies cited will incorporate the product into their devices. The announcement does not say which models the companies will actually run, or how they will assess operating conditions and safety after deployment. The figure of more than 3 million developers indicates the scale of the software platform, but does not directly show the number of production devices using it or continued use.

The device has 8GB of memory and an eight-core Arm CPU, while retaining the same compact form factor as the previous generation. NVIDIA says power consumption is 40% lower than the previous generation when delivering equivalent performance in 15W mode. It is intended for running small language models and vision-language models at the edge, with NVIDIA Cosmos, NVIDIA Nemotron, Gemma 4 and Qwen 3 cited as examples. NVIDIA says more than 3 million developers use its Jetson robotics software stack. Matic is moving ahead with adoption in home-cleaning robots. Wing uses the Jetson Orin Nano Super in its existing aircraft and plans to evaluate the Jetson Orin Nano 2. Cognex and Doosan Bobcat were also cited as companies adopting or considering it.