Nvidia launched the Jetson Orin Nano 2 on Monday, a compact robotics computer that doubles the inference performance of its predecessor while consuming 40 per cent less power. The module delivers 78 trillion operations per second of AI compute in a 15-watt envelope — enough to run frontier language and vision models on drones, home robots, and industrial sensors that previously needed a full workstation.
The price stays at $249 for the developer kit, putting datacentre-class reasoning within reach of hobbyists and startups.
What changed
The Orin Nano 2 keeps the same physical form factor as the existing Jetson Orin Nano Super, meaning it drops into existing designs without hardware changes. The gains come from improved Tensor Cores and higher memory bandwidth — architectural improvements rather than a bigger chip. Eight gigabytes of memory and an eight-core Arm CPU round out the spec sheet.
According to The Robot Report’s briefing with Nvidia, the company frames this as the moment edge hardware catches up with model compression. A year ago, frontier models needed 600-billion to 1-trillion-parameter chips in a datacentre. Now, models that match that accuracy can run on an entry-level board.
“Today’s small and medium frontier models have reached the accuracy of last year’s largest frontier models, unlocking real-time intelligence for edge devices,” said Deepu Talla, Nvidia’s vice president of robotics and edge AI.
Who is building on it
Nvidia says more than 3 million developers and 10,000 companies are already shipping or developing products built on the Jetson platform. Early partners for the Orin Nano 2 include:
- Wing (Alphabet’s drone delivery subsidiary), evaluating it for real-time perception and reasoning in delivery drones
- Matic Robots, using it to give home cleaning robots conversational AI, gesture detection, and semantic understanding of home environments
- Cognex, for industrial vision systems
The key claim is that companion and home robots — long limited by weak on-device intelligence — can now run models like Gemma 4, Qwen 3, and Nvidia’s own Cosmos and Nemotron at the edge, in real time. Talla called it the moment “we can actually put frontier AI models such as LLMs and VLMs on top of the other autonomous capabilities.”
This matters because the bottleneck for useful home robots has never been the motors. It has been the ability to understand a messy room, follow a spoken instruction, and adapt when a child leaves a toy on the floor. If that bottleneck is cracking open at $249, the implications go well beyond drones.
The three-computer strategy
The Orin Nano 2 is the deployment end of Nvidia’s “three-computer” approach to physical AI: Omniverse with Cosmos for simulation, DGX for training, and Jetson for runtime deployment. The company refreshed its higher-end Jetson lineup last month with the Thor T3000 and T2000 for mainstream robotics. The Orin Nano 2 fills the entry-level tier.
The board runs Nvidia’s full software stack — Isaac robotics framework, TensorRT for inference optimisation, and the new Jetson agent skills for memory-efficient model deployment. It shares the same GPU architecture as Nvidia’s datacentre chips, meaning models trained in the cloud can be exported to the edge without recompilation.
What this means
The edge AI hardware cycle has compressed dramatically. The Orin Nano Super launched in late 2025 at roughly half the performance. A year later, the Nano 2 doubles it at lower power. If that cadence holds — and Nvidia’s silicon roadmap suggests it will — the gap between what a datacentre can run and what a pocket-sized robot can run will close within two product cycles.
For New Zealand, the calculus is straightforward. Local robotics startups — and there are a handful building agricultural and inspection robots — can now prototype with frontier AI at component cost that doesn’t require venture funding. A $249 developer kit and a Raspberry Pi budget are no longer different categories of project.
The Orin Nano 2 module and developer kit will be available in the first half of 2027.
❓ FAQ
What is the Jetson Orin Nano 2? Nvidia’s entry-level robotics computer, designed to run AI models on small robots, drones, and edge devices. It sits at the bottom of Nvidia’s Jetson lineup, below the Thor T2000/T3000 and the high-end T4000/T5000.
How much does it cost? The developer kit is $249. Module-only pricing for production runs has not been disclosed.
Can it run large language models? Yes. Nvidia specifically calls out Gemma 4, Qwen 3, Cosmos, and Nemotron as models optimised for the platform. The 8GB memory limits model size, but quantised versions of 4B-parameter models are within reach.
What is edge AI? Running AI models directly on a device (robot, drone, sensor) rather than sending data to a cloud server. Edge AI means lower latency, no internet dependency, and better privacy — but requires compact, power-efficient hardware.
