Nvidia Targets Entry-Level Physical AI: Jetson Orin Nano 2 Doubles Inference Performance and Reduces Power Consumption by 40%

Nvidia is extending the capabilities of artificial intelligence (AI) even further from data centers to real-world devices such as robots.

On Tuesday the 25th Eastern Time, Nvidia announced the release of its new generation entry-level robotics computing platform Jetson Orin Nano 2, targeting edge AI applications in robots, drones, and other cost-, power- and real-time-sensitive scenarios.

Nvidia stated that while maintaining the same compact size, Jetson Orin Nano 2’s inference performance is doubled compared to the previous Jetson Orin Nano Super; in 15W power consumption mode, it achieves the same performance as its predecessor with 40% less power.

From a hardware perspective, Jetson Orin Nano 2 is equipped with an 8-core processor and 8GB of memory, with AI computing power reaching 780 TOPS. Nvidia claims the new platform boosts AI inference capabilities by upgrading tensor cores and increasing memory bandwidth, while remaining compatible with Nvidia’s software ecosystem for robots and edge AI. It can run memory efficiency-optimized large language models and vision-language models, including Cosmos, Nemotron, Gemma 4, and Qwen 3.

The significance of this new product release is not just about launching a more powerful small AI computer. As AI models continue to shrink and inference efficiency keeps improving, tasks like visual comprehension, language interaction, and real-time decision-making that were previously dependent mainly on the cloud are gradually shifting to terminal devices such as robots and drones. Nvidia hopes Jetson Orin Nano 2 will further lower the threshold for physical AI deployment in entry-level hardware.

Small AI Models Accelerate Implementation as Edge Smart Robotics Demand Heats Up

The launch of Jetson Orin Nano 2 comes as the AI industry extends from cloud to edge devices.

Unlike data center AI, robots and drones face stricter constraints on computing platforms: these devices are limited in size and usually can’t mount high-power processors, and must continuously handle visual perception, environmental understanding, and decision-making on the move. Thus, how much AI computing power is available per unit of power consumption matters more than simply pursuing peak performance.

Nvidia notes that in 15W mode, Jetson Orin Nano 2 can deliver the same performance as its predecessor with 40% lower power consumption. For battery-powered robots and drones, this means manufacturers can reduce energy usage for the same performance requirements or use the saved power space to improve endurance, heat dissipation, or other sensor configurations.

Meanwhile, rapid improvement in small AI models is also expanding the boundaries of edge AI applications.

Tara, Nvidia’s Head of Robotics and Edge AI, said that some mid- and small-scale frontier models now match last year’s large frontier models in accuracy, creating conditions for real-time intelligence on edge devices.

Previously, edge AI mainly handled tasks like object detection and image classification; as model capabilities improve, terminal devices are beginning to possess more complex language understanding, visual reasoning, and multimodal interaction abilities.

This transition enables robots to evolve from "perceiving the environment" to "understanding the environment and taking action."

From Visual Recognition to Multimodal Understanding, Real-Time Decision-Making Becomes Key

Nvidia is positioning the Jetson platform as a critical computing foundation for robots acquiring "physical intelligence."

Currently, the information that robots need to process goes beyond camera images—it now includes natural language commands, voice, spatial data, and input from various sensors. For robots, locally fusing such information quickly and responding with low latency is an essential prerequisite for autonomous operation.

Jetson Orin Nano 2 thus targets multiple typical scenarios, including home robots, visual AI systems, delivery drones, and inspection drones.

For example, Google-owned drone delivery company Wing has previously adopted Jetson Orin Nano Super and Nvidia’s software stack in its delivery drones and plans to evaluate Jetson Orin Nano 2. Nvidia says improved performance and energy efficiency will further enhance Wing’s real-time AI perception and inference capabilities for drones.

Home robot company Matic plans to use Jetson Orin Nano 2 to improve robots’ natural language interaction, gesture recognition, spatial mapping, and semantic understanding, allowing robots to autonomously complete cleaning and other tasks.

In addition, Nvidia showcased the application of Jetson Orin Nano 2 on the Reachy Mini robot. According to Nvidia, this platform can simultaneously handle language models, speech models, and real-time visual AI tasks, demonstrating the entry-level edge platform’s ability to run multiple AI models at once.

For robotics manufacturers, the importance of such capabilities lies in enabling robots not just to execute pre-programmed fixed actions, but to understand the environment based on visual and voice commands and then dynamically adapt their behavior.

This also explains why Nvidia has emphasized "physical AI" in recent years: artificial intelligence is stepping out from software services on screens to the real world and will ultimately power robots, vehicles, drones and other physical devices.

Developers and Partners Expand as Nvidia Continues to Strengthen the Robotics Ecosystem

Beyond chip performance, software ecosystem has always been a key competitive advantage of the Nvidia Jetson platform.

Nvidia states there are now over three million developers building on its robotics stack. With the launch of Jetson Orin Nano 2, Nvidia is further expanding the software, hardware, and developer ecosystem around robotic AI. Cognex, and Matic

On the software side, Jetson Orin Nano 2 continues to support Nvidia’s robot development platform and can run edge-device memory-efficient large language and vision-language models, including Nvidia’s Cosmos, Nemotron, as well as Gemma 4 and Qwen 3 models.

This means developers can deploy language understanding, visual comprehension, and reasoning abilities directly onto robots under relatively limited hardware resources, without fully depending on cloud servers.

The hardware ecosystem is also expanding in sync.

Nvidia revealed that partners such as AAEON, ADLINK Technology, Advantech, Aetina, Antmicro, Aptiv, Connect Tech, and Seeed Studio are developing carrier boards, hardware systems, AI software, and reference designs around Jetson Orin Nano 2, helping robotics manufacturers shorten their product development and time to market.

Nvidia also noted that the first wave of companies adopting or exploring Jetson Orin Nano 2 includes Cognex, Doosan Bobcat, Matic, and Wing, which also plans to evaluate the platform.

Looking at the longer term, Nvidia seeks to build not just a single robotics chip business but a comprehensive ecosystem covering AI models, software tools, computing platforms, and robotic end devices.

This model closely mirrors Nvidia’s competitive edge in the data center AI sector: synergy between hardware performance and software ecosystem, amplified by a massive developer community to further enhance platform appeal.

New Product Expected in the First Half of Next Year; Commercial Contribution Still Pending

However, it’s important to note that Jetson Orin Nano 2 is still at the launch stage and will not hit the market right away.

Nvidia says both Jetson Orin Nano 2 modules and developer kits are expected to become available in the first half of 2027. Therefore, the direct contribution of the new product to Nvidia’s recent financial results remains to be seen.

From a product strategy viewpoint, Nvidia is steadily refining its robotics and physical AI product lineup: on one hand, meeting the sophisticated needs of complex applications with higher-performance robotics computing platforms; on the other, pushing AI inference further down into more cost- and power-sensitive devices using entry-level platforms like Jetson Orin Nano 2.

If small language models and vision-language models continue to rapidly improve while inference costs keep dropping, demand for edge AI computing in robots, drones, and similar devices could expand further in the future.

For Nvidia, this means the market it’s competing for is gradually shifting from traditional data center AI computing to the "brains" of robots, drones, and other real-world devices. Jetson Orin Nano 2 represents an important step in Nvidia’s reach into the broader edge device market.