IN Brief:
- Aetina is preparing carrier boards, systems, integration, and software support around NVIDIA Jetson Orin Nano 2.
- The module provides up to 78 TOPS with 8GB memory and an eight-core Arm CPU.
- NVIDIA specifies twice the predecessor’s inference performance and 40% lower power at equivalent performance in 15W mode.
Aetina is preparing carrier boards, edge systems, peripheral integration, and software support for NVIDIA’s new Jetson Orin Nano 2 module, extending its existing Jetson-based embedded AI portfolio towards the next generation of compact robotics and machine-vision hardware. The module provides up to 78 TOPS of AI performance with 8GB of memory and an eight-core Arm CPU.
NVIDIA announced Jetson Orin Nano 2 on 25 August, with modules and developer kits expected to become available during the first half of 2027. Aetina’s announcement therefore concerns engineering support and forthcoming system platforms rather than a finished Orin Nano 2 product already in volume production.
The new module is specified to deliver up to twice the inference performance of Jetson Orin Nano Super while retaining the same compact form factor. NVIDIA attributes the increase to improved Tensor Cores and higher memory bandwidth, while maintaining a maximum 40W power envelope.
At a 15W operating point, NVIDIA says the module can deliver equivalent performance to its predecessor while consuming 40% less power. That performance-per-watt improvement is particularly relevant to embedded systems because the processing module competes with sensors, storage, communications, actuators, and power conversion within a fixed electrical and thermal budget.
A robotics computer that draws another 10W or 20W does not simply require a larger power supply. The additional heat has to be removed inside an enclosure that may be sealed against dust or moisture, while battery-powered machines have to carry enough stored energy to maintain operating time. Fan size, heatsinks, regulators, enclosure volume, and ambient-temperature limits can all change as compute power rises.
Memory remains another practical constraint. Generative and multimodal models require storage for parameters, activations, context, image data, and intermediate results, making optimisation important when the complete platform provides 8GB. Quantisation, model selection, context length, and scheduling therefore determine how much of the headline AI capability can be used simultaneously in a deployed system.
NVIDIA is positioning Orin Nano 2 for language and vision-language models in addition to conventional computer-vision inference. Its software environment supports models including Cosmos, Nemotron, Gemma 4, and Qwen 3, allowing a compact edge system to combine image interpretation with language-based instructions and higher-level reasoning.
For an industrial robot or vision system, the useful consequence is the possibility of running more of that processing locally. A camera system can analyse images without continuously sending them to a cloud service, while a mobile machine can continue perception and decision workloads when network latency or connectivity is variable. Local processing can also reduce the amount of raw sensor data that has to leave the equipment.
Aetina’s work begins beyond the compute-module specification. A production system needs carrier-board interfaces, camera connectivity, networking, storage, power conversion, thermal design, environmental protection, and a software image that consistently exposes the required peripherals. The company says its support will include carrier boards, edge AI systems, peripheral integration, BSP customisation, and technical assistance.
Those integration tasks can consume a substantial part of an embedded programme. A development kit may prove that a model runs, but converting it into equipment expected to operate continuously in a factory, vehicle, inspection cell, or robot requires validation of connectors, regulators, boot behaviour, peripheral drivers, thermals, and long-term software stability.
The first-half 2027 availability date also leaves a significant engineering interval before volume deployment. Developers can begin software and system architecture work around the announced specification, but final designs will still need qualification against production modules, final software releases, actual thermal behaviour, and commercial supply once hardware becomes available.
Aetina already develops Jetson-based DeviceEdge systems and customised carrier-board solutions, so Orin Nano 2 extends an existing hardware route rather than introducing an unfamiliar platform. Compatibility with the wider Jetson ecosystem should help software migration, but the increased compute capability also encourages workloads that put more pressure on memory, thermal design, and power delivery.
The module therefore shifts the entry point for embedded AI without removing the traditional constraints of embedded engineering. Seventy-eight TOPS may determine which models can run, but power integrity, cooling, interfaces, storage, BSP stability, and environmental qualification will determine whether those models can keep running once the hardware leaves the laboratory.
