NVIDIA Launches Entry-Level Robotics Chip with Doubled Inference Performance and 40% Lower Power Consumption
Nashnova编辑部
Nvidia unveiled Jetson Orin Nano 2, an entry-level robotics chip that doubles inference performance and cuts power 40% at equal workloads — targeting not the data center but the 'on-board brain' market for robots and drones.
What exactly got upgraded?
Jetson Orin Nano 2 packs an 8-core CPU, 8 GB of memory, and 78 TOPS (trillion operations per second) of AI compute.
Versus the prior generation, inference performance is 2× higher; at 15 watts it matches the old chip's output while drawing 40% less power.
This means → the headline selling point isn't "faster" — it's "just as fast on far less power." For a battery-powered device, that matters more than peak throughput.
Why does "performance per watt" matter most at the edge?
Robots and drones are small and battery-powered — they can't house the high-wattage processors that sit in data centers.
They must perceive, understand, and decide continuously while moving — compute can't pause, and neither can the battery.
In plain terms = data centers compete on who computes the most; robots compete on who computes enough on the least power.
A 40% power saving lets device makers extend battery life, improve cooling, or free space for extra sensors.
Who is already using Nvidia's robotics chips?
Wing, Google's drone-delivery unit, already runs the prior-gen Jetson Orin Nano Super on its delivery drones and plans to evaluate the new chip.
Home-robot maker Matic plans to use Jetson Orin Nano 2 for natural-language interaction, gesture recognition, and spatial mapping.
Industrial players Doosan Bobcat and Cognex have also begun adopting or exploring the platform; hardware partners including Aetina, ADLINK, and Advantech are developing carrier boards and reference designs.
What business is Nvidia really building here?
Nvidia robotics head Tara said mid-size frontier models now match last year's large-model accuracy — creating the conditions for real-time intelligence on edge devices.
The new platform is compatible with Nvidia's robotics software stack and can run memory-optimized LLMs and VLMs including Cosmos, Nemotron, Gemma 4, and Qwen 3.
This reflects Nvidia's data-center playbook applied to a new domain: sell not just chips but a chip + software + model stack that locks in the entire dev chain. Over 3 million developers already build on its robotics stack.
When can you buy it — and does it move the earnings needle?
The Jetson Orin Nano 2 module and dev kit are expected to ship in H1 2027.
This means → at least a year before real revenue arrives; the near-term earnings impact is minimal.
In plain terms = this is a long-term bet that every robot will need an on-board AI chip. No big numbers yet — but it's the clearest test of whether Nvidia's growth story extends beyond the data center.
Content is for reference only, not financial advice.