NVIDIA Jetson Orin Nano 2 Announced: Doubling Inference for Robots and Delivery Drones at 15W

Forum topic · 小凯 · 2026-08-25

Summary

On August 25, 2026, NVIDIA announced the Jetson Orin Nano 2, a compact entry-level edge AI module for robotics and physical AI applications. The module features an 8-core Arm CPU, 8GB of memory, and 78 TOPS of AI compute, delivering 2x the inference performance of the previous Jetson Orin Nano Super while consuming 40% less power for equivalent throughput (about 9W in a 15W power mode). NVIDIA positions it as the entry point of its physical AI ecosystem, as small and mid-size frontier models now match last year's largest models, enabling generative AI to run directly on-device. Early adopters include Wing (Alphabet's drone delivery company), Matic Robotics (consumer home cleaning robots, running conversational AI, gesture detection, and semantic home mapping fully on-device), Cognex (industrial vision), and Doosan Bobcat. Over 20 partners including AAEON, ADLINK, Advantech, Aetina, and Seeed Studio are building carrier boards and reference systems, with 3 million+ developers on NVIDIA's robotics stack. The platform supports edge-optimized models such as Cosmos, Nemotron, Gemma 4, and Qwen 3, with shipping expected in the first half of 2027.

NVIDIA announced the Jetson Orin Nano 2 on August 25, 2026 — a new-generation entry-level robot computing module that doubles inference performance over its predecessor while achieving equal throughput at 40% lower power. The launch signals that frontier generative AI is rapidly moving into consumer robots, drones, and home devices — the start of a "Cambrian explosion" for entry-level physical AI.

Key specs and performance

8-core Arm CPU, 8GB memory, 78 TOPS AI compute in the same compact form factor
2x inference performance vs. Jetson Orin Nano Super, thanks to improved Tensor Cores and higher memory bandwidth
In 15W mode, equal inference throughput at 40% lower power — meaning manufacturers can cut power budgets from 15W to roughly 9W
Key metric for battery-powered devices is TOPS-per-watt, which dictates battery life and thermal design

> Note: TOPS (Tera Operations Per Second) measures AI inference throughput; for edge devices, TOPS/Watt often matters more than absolute TOPS.

Early adopters

Wing (Alphabet's drone delivery subsidiary) already runs Jetson Orin Nano Super in its fleet and plans to evaluate Orin Nano 2 for real-time perception. Per Wing's perception lead Dinuka Abeywardena: drone delivery depends on AI that quickly and reliably understands the real world.
Matic Robotics (consumer home cleaning robots) announced adoption, planning to use the extra compute for conversational AI, gesture detection, semantic high-precision home mapping, and autonomous cleaning in dynamic environments — all on-device, no cloud required.
Other ecosystem adopters include Cognex (industrial vision) and Doosan Bobcat (heavy equipment).

The physical AI stack

NVIDIA VP of robotics and edge AI Deepu Talla summarized the industry inflection point: "small and mid-size frontier models today have unlocked real-time intelligence for edge devices." Three converging trends drive this:

1. Model miniaturization — distillation, quantization (INT8/INT4), and pruning; open models like Cosmos, Nemotron, Gemma 4, and Qwen 3 offer memory-efficient edge variants. 2. Hardware efficiency gains — improved Tensor Cores and memory bandwidth let the same models run faster at lower power, moving vision, language, and real-time decision workloads from cloud to device. 3. Unified software stack — JetPack SDK, Isaac simulation/Isaac ROS, CUDA, and TensorRT cover the full pipeline from cloud training to edge deployment.

Ecosystem

NVIDIA claims 3 million+ developers on its robotics stack. More than 20 partners — including AAEON, ADLINK, Advantech, Aetina, and Seeed Studio — are building carrier boards, hardware systems, and reference solutions. Orin Nano 2 sits at the lowest price and power tier of NVIDIA's pyramid: Jetson Thor / IGX Thor at the top (thousands of TOPS for autonomous driving and industrial robotics), Jetson Orin NX / AGX Orin in the middle (hundreds of TOPS), and Orin Nano 2 (78 TOPS) as the highest-volume entry point.

Why "entry-level" matters

For robots, drones, and home devices — constrained by size, battery, thermals, and cost — per-watt compute and deployment density matter more than peak performance. Orin Nano 2 can handle real-time visual perception, language understanding, and path planning within a 9–15W envelope without cloud connectivity, shifting from cloud-dependent inference (high latency, privacy risk, offline failure) to autonomous edge inference with millisecond response. As Matic CEO Navneet Dalal put it, home robots need to understand people, map spaces precisely, and clean autonomously in dynamic environments — workloads that fit this module's sweet spot.

Availability

The Jetson Orin Nano 2 is expected to ship in the first half of 2027, giving partners time to build products around the chip.

Key numbers: 78 TOPS / 8GB / 8-core Arm CPU (NVIDIA, 2026-08-25); 2x inference uplift and 40% power reduction vs. Orin Nano Super; 3M+ developers; H1 2027 availability; support for Cosmos, Nemotron, Gemma 4, Qwen 3 edge-optimized models.

*Sources: NVIDIA Newsroom / Wallstreetcn / Inside AI / AI Future Front / Unite.AI / Wing statements / Matic Robotics CEO interview*