What happened
NVIDIA announced Jetson Orin Nano 2, a new compact robotics computer designed to run AI models locally in robots, drones and vision systems. The company says it provides 78 trillion operations per second, 8GB of memory and an eight-core Arm CPU, with twice the inference performance of Jetson Orin Nano Super in the same form factor. NVIDIA says the device uses 40% less power at equivalent performance in 15-watt mode.
NVIDIA said on August 25, 2026, that it had announced Jetson Orin Nano 2, a new robotics computer for entry-level edge AI. The company positioned the product for robots, delivery and inspection drones, and vision AI systems. Its stated purpose is to let developers run AI capabilities inside physical devices rather than relying solely on larger computing systems elsewhere. The source describes this as part of a broader shift toward smaller and more efficient AI models that can interpret language and images, understand context and act in real time.
The announced hardware includes 78 trillion operations per second of AI compute, 8GB of memory and an eight-core Arm CPU. NVIDIA says Jetson Orin Nano 2 maintains the same compact form factor as its predecessor while delivering twice the inference performance of Jetson Orin Nano Super through improved Tensor Cores and higher memory bandwidth. In a 15-watt operating mode, NVIDIA says it consumes 40% less power while delivering the same performance as the predecessor. These are vendor-reported specifications and comparisons; the source does not provide test conditions, benchmark results or independent verification.
The computer is tied to NVIDIA’s software stack for Jetson, including what the company calls Jetson agent skills and other development tools. NVIDIA says the platform supports the latest large language models and vision-language models optimized for memory-efficient edge inference. The announcement names NVIDIA Cosmos, NVIDIA Nemotron, Gemma 4 and Qwen 3 as examples of open models that developers can use to build applications. It does not specify which model versions, quantization settings, context sizes or operating limits apply to each example.
NVIDIA said more than 3 million developers are building on its robotics stack, a figure that comes from the company. It identified Cognex, Doosan Bobcat and Matic as among the first companies to adopt or explore Jetson Orin Nano 2. Matic says it is adopting the computer for home-cleaning robots with conversational AI, gesture detection, precision mapping, semantic understanding and autonomous cleaning. Wing currently uses Jetson Orin Nano Super in its delivery-drone fleet and says it plans to evaluate Jetson Orin Nano 2. NVIDIA also listed numerous partners developing carrier boards, hardware systems, customized software and reference solutions. The module and developer kit are expected to be available in the first half of 2027.
Read the primary source: nvidianews.nvidia.com ↗
Why it matters
The announcement targets developers building physical systems that need local processing for perception, language and decision-making. NVIDIA says Jetson Orin Nano 2 can run memory-efficient large language and vision-language models, including selected open models, on compact edge devices. That could make more capable local AI practical for robotics and drones, although the announcement provides no independent testing, pricing or current availability.
The product’s significance rests on where NVIDIA says the computation will happen: at the edge, inside or near a physical machine. For a robot, drone or vision system, a compact computer that can process models locally could support the real-time perception and reasoning these devices require. NVIDIA’s announcement specifically connects the product to understanding environments, interpreting language and images, and acting in response. Those are company descriptions of the intended use, not evidence that every listed application has been demonstrated on Jetson Orin Nano 2.
The claimed performance and power improvements could matter most for developers constrained by physical size or energy budgets. NVIDIA says the new computer fits the same form factor as Jetson Orin Nano Super, doubles inference performance and reduces power use at equivalent performance in a 15-watt mode. If those claims generalize to practical workloads, developers could have more room to run larger or more capable models in compact systems. The source does not state battery life, operating temperature, sustained performance or the effect of connected sensors and other components on total system power.
The software support broadens the announcement beyond a hardware specification. NVIDIA is presenting Jetson Orin Nano 2 as part of an ecosystem that includes its own software, agent skills, model families and third-party hardware. That may reduce development work for organizations that already use Jetson tools or partner components. However, the source does not establish how easy migration will be, what licensing terms apply to the named models, or whether all listed models can perform equally well on the device.
The potential public impact is therefore prospective rather than immediate. The product could help developers build more responsive home robots, inspection systems, delivery drones and other devices, but the announcement does not document a completed Jetson Orin Nano 2 deployment at scale. Matic describes intended capabilities for its robots, and Wing describes a future evaluation. NVIDIA’s claim that more than 3 million developers use its robotics stack also measures ecosystem reach, not the number of systems operating with this new computer.
What to watch next
The module and developer kit are not expected to be available until the first half of 2027. Key questions include whether NVIDIA’s performance and power claims hold across representative workloads, how much the system and supporting hardware will cost, and which models can run at useful speeds. The announced partner activity also varies: some companies are adopting or exploring the platform, while Wing says it plans to evaluate it.
Availability is the first practical milestone. NVIDIA says both the Jetson Orin Nano 2 module and developer kit are expected in the first half of 2027, so the product is an announcement rather than a currently available purchase. The source gives no price, order process, production volume or more precise shipping date. Those details will determine whether the platform is accessible to independent developers and smaller robotics companies as well as established partners.
Independent testing should examine the headline comparisons under clearly described conditions. Important details missing from the source include the workloads used to establish twice the inference performance, the models and precision formats tested, whether performance is sustained over time, and how power consumption changes across different operating modes. Comparisons should also measure complete systems, because sensors, storage, cooling, wireless connections and motors can affect the power and responsiveness of a finished robot or drone.
The partner announcements require careful interpretation. NVIDIA says Cognex, Doosan Bobcat and Matic are among the first to adopt and explore the product, while Wing says it plans to evaluate it. Those statements do not establish commercial deployment, production volume, customer availability or improved safety. Follow-up reporting should distinguish prototypes, evaluations, reference designs and shipped products, and should seek evidence about real-world reliability in changing environments.
The announcement says little about safeguards for systems that can act physically. NVIDIA and its partners discuss perception, navigation, autonomous cleaning and faster, safer delivery, but the source provides no safety testing, failure-rate data, human-override procedures or account of how models behave when sensor data are incomplete or wrong. NVIDIA also lists manufacturing, software defects, market acceptance, integration performance, third-party supply and regulatory changes among risks that could affect results. Those stated uncertainties remain material until the product is available and independently evaluated.
