On August 25, NVIDIA released the Jetson Orin Nano 2 robotic computer, aimed at the entrance edge AI, robotics, distribution and inspection drones and visual AI systems. The official specifications include 78 trillion operations per second, 8GB RAM and 8nucleary Arm CPU; compared with Jason Orin Nano Super, the reasoning performance was increased to a factor of two and the outer size remained unchanged. In the 15-watt model, it is reduced by 40 per cent by the time it operates with the same performance.
Particular emphasis needs to be placed on the fact that the product is now “declared” and not universally available. NVIDIA indicates that the modules and development packages are expected to be listed in the first half of 2027. Cognex, Doosan Bobcat and Matic were among the first companies to adopt or explore, while Wing planned to evaluate the product; the “exploration” “plan assessment” was in different states and could not be uniformly written as deployed. No price, specific date of sale or all regional supply arrangements are currently available.
78TOPS is not an overall robotic capability. Memory and software determine whether or not the model really works.
TOPS measures the amount of operation that can be performed per second at a given accuracy and is suitable to describe the peak of the chip, but robots need to process cameras, sensor integration, positioning, planning, language interaction and security control simultaneously. 78TOPS does not mean that all models can run at the same speed and cannot be directly equated with mission accuracy. Model structures, quantification, memory occupancy and data removal often create bottlenecks earlier than theoretical calculations.
The 8GB memory determines that the device is more suitable for compressed small and medium models rather than directly carrying the data centre level supermodel. NVIDIA lists the optimised open models Cosmos, Nemotron, Gemma 4 and Qwen 3, but each model still has to choose the appropriate parameter size and precision. The developers also have to set aside space for the simultaneous operation of the camera stream, system processes and multiple models. The so-called "front-level generation AI capability" should be understood to allow for the operation of a new generation of model loads on the edges, not to imply that any forward model can be fully loaded into the equipment.
Declining power consumption is particularly important for mobile robots. When battery capacity is limited, the calculation of each watt saved may be translated into a longer, smaller heater or more sensor budget. However, “a 40 per cent drop in the 15-watt model” is limited to a 40 per cent reduction in all load-down power consumption in comparison with the performance of the previous generation. Electricity, cameras, communications and power consumption continue to consume energy, and the actual renewal requires full machine testing.
Maintaining the same shape helps partners to follow the shell and panel design, but does not mean that old modules can be unconditionally upgraded. Interfaces, power, heat-dispersion, solids and JetPack versions are verified. NVIDIA lists a number of board, hardware systems and reference programme partners, reflecting the fact that for products to actually enter robotics, complete engineering ecology around modules is required.
The value of borderline reasoning is real time and local control, but security certification cannot be omitted.
The completion of sensory and reasoning on the equipment can reduce delays, bandwidth and privacy pressures associated with the continued uploading of the cloud. Household cleaner robots need to understand people, gestures and objects in dynamic space; drones need to identify barriers and drop points when connections are unstable; and industrial visual systems need to judge at the speed of the production line. These scenarios require a millisecond response, and edge calculations are more reliable than remote calls.
Matic indicated that new modules were planned to support dialogue, handprint testing, fine maps, semantic understanding and self-cleaning. Wing is exploring more real-time, energy-efficient drone awareness and reasoning. Both indicate a potential workload but are not a guarantee of final safety or commercial effectiveness. Unlike the cost of failure for domestic use and air robots, model outputs must be subjected to certainty control layers, sensor cross-certification and safety shut-off mechanisms, and the generation of models cannot be a direct substitute for all control logic.
NVIDIA claims that more than 3 million developers already use its robotic software warehouse. The large-scale developer base can reduce the cost of software migration for new modules, but the ecological scale is not equal to the volume produced for each project. From demonstration to product, robots also cross data collection, long-term reliability, certification, after-sale and cost thresholds. The development of packages is appropriate for the prototype, and the final equipment usually requires customized pallet, sensor and power design.
For the development team, the model memory, end-to-end delay and fault boundary should be confirmed during the waiting period for listing, rather than only the TOPS reservation structure. Quantification and performance imagery could be completed on the existing Jetson platform to assess whether Nano 2 ' s additional calculus addressed real bottlenecks. If bottlenecks come from cameras, storage or control algorithms, the change of chips does not necessarily result in a double system upgrade.
The positioning of Jason Orin Nano 2 is clear: a compact, low-capacity, entry-level platform for a stronger generation and visual language model. Its technical parameters and the first cases of cooperation were of concern, but the products were not expected to be listed until the first half of 2027. The real tests will come from volume prices, the development of packages, real model benchmarks, overall power consumption and the reliability of robots on long-term sites, rather than a single peak figure.
Source: NVIDIA, NVIDIA Innocent Orin Nano 2 Robotics Company to Redefine Entry-Level Edge AI, 25 August 2026, https://nvidianews.nvidia.com/news/nvidia-announces-jetson-orin-nano-2-robotics-computer-to-redefine-entry-lever-edge-ai