Nvidia unveiled the Jetson Orin Nano 2 on August 25, an entry-level edge AI computer aimed at robots, delivery and inspection drones, and vision systems. On paper, the module delivers 78 TOPS of AI compute, 8GB of memory and an 8-core Arm CPU, doubling inference performance over the previous-generation Jetson Orin Nano Super while keeping the same form factor.
Nvidia attributes the performance gain to an improved Tensor Core paired with higher memory bandwidth — there's no mention of a process-node shrink. A different figure says more than the peak-compute number: at matched performance, power draw in the 15W operating mode is down 40 percent from the prior generation.
The bottleneck at the edge was never peak performance
Do the rough math and 78 TOPS at the 15W tier works out to roughly 5 TOPS per watt. That figure wouldn't turn heads in a data center, but edge products live under different constraints — an inspection drone's flight time is set directly by power draw, and a home robot has no room for active cooling. For people actually building products, that 40 percent power cut at equal performance matters more than the headline "twice as fast."
Memory is the other point worth watching, and the other limit. 8GB sets a clear ceiling on what models can fit. Nvidia says the module can run Cosmos, Nemotron, Gemma 4 and Qwen 3 at the edge — two of those are Nvidia's own, the other two are open-source families from Google and Alibaba, and what they share is that each has a small-footprint version. The list itself draws a product boundary: what runs at the edge is distilled, quantized small models, not flagship parameter counts.
Nvidia's VP of Robotics and Edge AI, Deepu Talla, framed the launch around reach.
"The Jetson Orin Nano 2 computer puts that breakthrough within reach of millions of developers, delivering the performance and energy efficiency needed for real-time reasoning."
The company says more than 3 million developers already use its robotics stack, and that number is the real moat behind the Jetson line — not the spec sheet of a single module, but the CUDA-to-Isaac toolchain that already works end to end.
Launch is one thing, availability is next year's problem
One detail easy to miss under the spec sheet: the module and developer kit are expected to ship in the first half of 2027. Counting from the August 2026 announcement, that's roughly a ten-month gap.
That's not unusual for the hardware industry — design cycles in robotics and industrial vision already run long, and customers need to lock in a platform and lay out development schedules well ahead of time. The early-access and evaluation list includes Cognex in machine vision, Doosan Bobcat in construction equipment, Matic Robots in home cleaning robots, and Wing, Alphabet's delivery-drone unit — all teams whose hardware cycles are measured in years. Nvidia counts more than 20 ecosystem partners, including carrier-board and system makers AAEON, ADLINK, Advantech, Connect Tech and Seeed Studio, and Aptiv has also announced support.
Wing's perception lead, Dinuka Abeywardena, described the company as "evaluating" the module, while Matic co-founder and CEO Navneet Dalal said it lets the company run state-of-the-art AI models on a compact home-robot platform. The caution in that language tracks the timeline — for now, only intent can be made public.
Seen another way, announcing ten months ahead of availability is also a way of staking a claim. There's no shortage of competition at this layer of edge AI — Qualcomm, Rockchip, Horizon Robotics and a cluster of domestic NPU platforms are chasing the same customers, and several of them ship on shorter timelines. By putting numbers on the table now, Nvidia gives teams still weighing their options a reason to wait.
For robotics and smart-hardware teams in China, the practical value of this announcement is a benchmark: 78 TOPS, 8GB and 15W is the bar that entry-level edge platforms will need to match by the first half of next year. Whether the module will actually be available to buy, and at what price, Nvidia didn't say this time, and it gave no separate word on supply for the China market. The spec sheet is ready to read; procurement plans still have to wait.
Sources: Nvidia press release, The Robot Report, CocoLoop; verified TOPS compute, memory capacity, CPU core count, the 15W-mode power reduction figure, and the availability-window language.