NVIDIA's New Entry-Level Robotics Chip Doubles Edge AI Power
NVIDIA's Jetson Orin Nano 2 delivers 78 TOPS, twice its predecessor's inference speed, at 40% less power, shipping in developer kits in H1 2027.
NVIDIA said on August 25 that its Jetson Orin Nano 2 robotics computer, the successor to the Jetson Orin Nano Super, delivers 78 trillion operations per second of AI compute, twice the inference performance of its predecessor, while cutting power draw 40% at matched performance in the module’s 15-watt mode. The module and developer kit are expected to ship in the first half of 2027, aimed squarely at the entry-level tier of the robotics-compute market: the cheapest, lowest-power boards developers put inside cost-sensitive robots, delivery drones and inspection systems rather than the flagship Thor-class modules NVIDIA sells for humanoids and autonomous vehicles.
The pitch, in NVIDIA’s own words from vice president of robotics and edge AI Deepu Talla, is about what now fits in a small power envelope: “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. 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 at the edge.”
The spec sheet, and what it’s built to run
Orin Nano 2 keeps the same compact form factor as the Orin Nano Super and the same 8GB of memory and 8-core Arm CPU, but NVIDIA says improved Tensor Cores and higher memory bandwidth get it to 78 TOPS of AI compute, twice its predecessor’s inference throughput. NVIDIA frames the power-efficiency gain the same way: in 15-watt mode, the new module draws 40% less power to deliver the performance its predecessor needed more power to reach. None of those figures come from an independent test lab; they are NVIDIA’s own published comparisons against its own prior-generation part.
The module is built to run large language models and vision-language models specifically optimized for memory-constrained edge inference, and NVIDIA names four example families: its own Cosmos world-foundation-model platform and Nemotron models, plus third-party models Gemma 4 and Qwen 3. Worth flagging precisely: NVIDIA’s release text names “Gemma 4,” not the more widely deployed Gemma 3 generation, so that’s what this story reports rather than assuming a typo. The idea behind naming all four is that a robot running on Jetson Orin Nano 2 can do more than fixed-function perception, it can hold a real-time conversation, reason about a scene in language, or run a physical-AI world model locally rather than round-tripping to the cloud.
Who’s actually building on it
NVIDIA names Cognex, Doosan Bobcat and Matic Robots as among the first companies adopting and exploring Jetson Orin Nano 2. Matic, which makes a home-cleaning robot, gave the most concrete use case: CEO Navneet Dalal said the extra headroom lets the company run “state-of-the-art AI models at the edge in a compact home robotics platform built for real-time perception, interaction and navigation,” specifically citing conversational AI, gesture detection and precision mapping as capabilities the new compute unlocks.
Wing, Alphabet’s drone-delivery subsidiary, offers a more cautious read. It currently runs the prior-generation Jetson Orin Nano Super in its delivery-drone fleet today and says it plans to evaluate, not yet deploy, the new module. Wing’s head of perception, Dinuka Abeywardena, framed the interest in terms of real-time understanding: “Drone delivery depends on AI that can enable fast, reliable understanding of the real world. Wing is exploring Jetson Orin Nano 2 to give us a path to more responsive, energy-efficient drones that can help make deliveries quicker and more dependable for customers.” That’s an evaluation commitment, not a production deployment, an important distinction NVIDIA’s release does not blur but a casual read of the announcement could.
Beyond the three named early adopters, NVIDIA lists more than a dozen hardware partners already building carrier boards and reference systems around the module, including AAEON, Advantech, Aptiv, Connect Tech and Seeed Studio, and says more than 3 million developers are building on its Jetson robotics stack overall. That developer-count figure is also NVIDIA’s own reporting, with no methodology disclosed for how it’s measured.
Where this sits in a crowded edge-compute field
Jetson Orin Nano 2 is not launching into an empty market. AMD opened its own Robotics Partner Network and Ryzen AI Embedded X100 processor family in July, a rival embedded chip line topping out around 50 TOPS inside a 55-watt envelope, already showing up in real integrations such as Avnet and Weston Robot’s quadruped inspection platform. NVIDIA’s own broader robotics push extends well past this one module, from the Cosmos physical-AI stack Japanese manufacturers FANUC, Yaskawa, Kawasaki Heavy Industries and Fujitsu are integrating to the safety-certified Jetson Thor compute inside Agility Robotics’ Digit v5 humanoid, a reminder that Orin Nano 2 sits at the low-cost, low-power end of a much larger NVIDIA compute ladder rather than as a standalone product.
The specific claim worth tracking is the TOPS-per-watt gain, since that’s the number that actually determines whether a robot maker can run a genuinely capable language or vision-language model on a battery-powered device rather than a tethered one. NVIDIA’s own numbers say entry-level edge hardware just got meaningfully closer to running frontier-class small models in real time; nobody outside NVIDIA has independently confirmed that yet.
Sources
- NVIDIA Announces Jetson Orin Nano 2 Robotics Computer to Redefine Entry-Level Edge AI — NVIDIA Newsroom, Aug 25, 2026
Frequently asked questions
What actually changed between the Jetson Orin Nano Super and the new Orin Nano 2? +
NVIDIA says Orin Nano 2 delivers 78 trillion operations per second (TOPS) of AI compute, twice the inference performance of the Orin Nano Super, in the same compact form factor, through improved Tensor Cores and higher memory bandwidth. Memory stays at 8GB with an 8-core Arm CPU, and in 15-watt mode the new module draws 40% less power than its predecessor to hit equivalent performance. All of these are NVIDIA's own published specifications, not results from an independent benchmark lab.
What AI models is Jetson Orin Nano 2 actually built to run? +
NVIDIA's own announcement names NVIDIA Cosmos, NVIDIA Nemotron, Gemma 4 and Qwen 3 as example large language and vision-language models optimized for memory-efficient edge inference on the module. Note that NVIDIA's release specifically says "Gemma 4," not the more widely known Gemma 3 line, which this story reports as written in the source rather than assumed.
Which companies are already using or evaluating Jetson Orin Nano 2? +
NVIDIA names Cognex, Doosan Bobcat and Matic Robots as among the first to adopt and explore the new module, and Wing, Alphabet's drone-delivery subsidiary, which currently runs the older Jetson Orin Nano Super in its delivery drones and says it plans to evaluate the new chip. NVIDIA also lists more than a dozen hardware partners, including AAEON, Advantech, Aptiv and Seeed Studio, building carrier boards and reference systems around the module.
When can a robotics developer actually buy this? +
NVIDIA says the Jetson Orin Nano 2 module and developer kit are expected to be available in the first half of 2027. There is no announced price.
More news
