NVIDIA Corporation (NASDAQ: NVDA) has introduced Jetson Orin Nano 2, a new entry-level robotics computer delivering 78 trillion operations per second of AI compute as the semiconductor company extends its physical AI strategy from high-end humanoids and industrial robots toward smaller autonomous machines. NVIDIA says the new system provides twice the inference performance of Jetson Orin Nano Super while maintaining the same compact form factor, with 40% lower power consumption when delivering equivalent performance.
The platform combines 8GB of memory with an eight-core Arm CPU and is intended for robots, delivery and inspection drones, vision AI systems and other machines that need to perform inference locally instead of sending every sensor input to a distant data centre. More than three million developers have already built on NVIDIA’s robotics software stack, giving the company a sizeable installed development ecosystem into which the new hardware can be introduced.
Commercial availability is not immediate. NVIDIA expects the Jetson Orin Nano 2 module and developer kit to become available during the first half of 2027, making the August announcement a product roadmap milestone rather than a near-term hardware revenue event.
How much faster is Jetson Orin Nano 2 than NVIDIA’s current entry-level platform?
The existing Jetson Orin Nano Super delivers up to 67 TOPS of AI performance, 8GB of memory and 102GB per second of memory bandwidth, with configurable power between 7 watts and 25 watts. NVIDIA currently prices the developer kit at US$249.
Jetson Orin Nano 2 raises headline AI compute to 78 TOPS, only about 16% above the predecessor’s 67-TOPS peak figure. That makes NVIDIA’s claim of twice the inference performance particularly important because the performance improvement cannot be understood simply by comparing headline TOPS.
The company attributes the inference gain to improved Tensor Cores, greater memory bandwidth and architectural changes that allow AI models to execute more efficiently.
That distinction matters in practical robotics. TOPS is a theoretical compute metric; developers ultimately care about how quickly a particular vision-language model, perception pipeline or robotic policy can generate an output under real memory and power constraints.
A doubling of actual inference performance could consequently matter much more than a 16% increase in the nominal TOPS figure if applications scale as NVIDIA’s benchmark claims suggest.
Why could the 40% power reduction matter more than raw AI performance?
Robots and drones operate under a constraint that data-centre servers largely avoid: every watt consumed by compute competes with motors, sensors, communications equipment and overall battery life.
NVIDIA says Jetson Orin Nano 2 can deliver the same level of performance as its predecessor while consuming 40% less power.
If an application previously required 25 watts of module power and could genuinely achieve equivalent performance using 40% less, the theoretical reduction would be about 10 watts. Actual power consumption will depend on workload and configuration, but the example demonstrates why efficiency can be commercially important in small autonomous systems.
A delivery drone has to carry its battery into the air. A mobile robot needs enough stored energy to complete a shift. A remotely deployed inspection system may operate where charging opportunities are limited.
Reducing compute power therefore has the potential to improve battery life, allow smaller power systems or free energy for additional sensors and actuation.
For edge AI, performance per watt can be more economically important than maximum standalone compute.
Which companies are already evaluating or adopting Jetson Orin Nano 2?
NVIDIA identified Cognex, Doosan Bobcat and Matic among the first companies adopting or exploring the new platform. Alphabet subsidiary Wing also plans to evaluate Jetson Orin Nano 2 after using Jetson Orin Nano Super and NVIDIA software in its drone-delivery fleet.
Matic intends to use the processor in home cleaning robots for workloads including conversational AI, gesture detection, mapping, semantic understanding and autonomous navigation.
Wing’s interest highlights another edge-AI requirement. Delivery drones need to understand visual environments quickly while operating under tight weight, energy and connectivity constraints. Performing more reasoning onboard can allow the aircraft to respond without depending entirely on cloud connectivity.
NVIDIA also named a broad hardware ecosystem including AAEON, ADLINK, Advantech, Aetina, Antmicro, Aptiv, Connect Tech, Seeed Studio and other partners developing carrier boards, complete systems, software and reference solutions around the new module.
That ecosystem is important because NVIDIA generally does not build the entire robot. Its commercial position depends on supplying the compute and software layer around which robot manufacturers, component companies and system integrators design their own products.
Why is NVIDIA emphasizing smaller AI models at the edge?
The economics of edge inference have changed as AI models have become more capable at smaller parameter counts.
A few years ago, advanced language and multimodal reasoning often required large cloud accelerators with substantial memory. Smaller models can now perform increasingly sophisticated vision, language and planning tasks using dramatically less compute.
NVIDIA says Jetson Orin Nano 2 can run optimized large language and vision-language models including NVIDIA Cosmos and Nemotron as well as open models such as Gemma 4 and Qwen 3.
Moving those models onto a robot creates several potential advantages. Latency can fall because sensor information does not need to travel to a data centre and back. Connectivity becomes less critical. Sensitive visual or operational data can remain on the device. Cloud inference charges can also be reduced for high-frequency workloads.
The trade-off is that an embedded computer cannot match the memory or raw compute of NVIDIA’s data-centre GPUs. Developers therefore need smaller models, quantization, optimization and specialized software to fit useful intelligence into the available power and memory envelope.
That is where NVIDIA’s software stack becomes strategically significant.
How does the new Jetson product strengthen NVIDIA’s wider physical AI strategy?
NVIDIA’s competitive position in robotics is not based solely on selling an embedded module.
Developers can use JetPack, Isaac ROS and other NVIDIA tools to create applications, simulate and train robotics systems and then deploy inference on Jetson hardware. The objective is to create continuity between data-centre model development, simulation and the computer installed inside the physical machine.
That resembles the strategy NVIDIA successfully used in data centres: hardware becomes more difficult to replace when developers also depend on the software ecosystem around it.
The scale is already substantial. NVIDIA says more than three million developers are using its robotics stack.
Jetson Orin Nano 2 targets the lower end of that ecosystem rather than competing with the company’s most powerful robotic processors. Jetson AGX Orin can deliver up to 275 TOPS, while newer Jetson platforms target far more computationally demanding autonomous systems.
That segmentation gives NVIDIA a path from relatively inexpensive development hardware through industrial and high-performance robotics compute without requiring developers to move to a completely different software environment as products become more demanding.
Does NVIDIA disclose a price for Jetson Orin Nano 2?
Not yet in the August 25 announcement.
That omission matters because the existing Jetson Orin Nano Super developer kit sells for US$249 and has been positioned as NVIDIA’s affordable entry point into generative edge AI.
The predecessor originally cost US$499 before NVIDIA cut the developer-kit price to US$249 while increasing performance through a software update. That move dramatically improved compute-per-dollar for developers.
Jetson Orin Nano 2 therefore faces an unusually strong internal benchmark. Twice the inference performance and better energy efficiency are attractive, but adoption among smaller robotics companies, students and makers will also depend on whether NVIDIA preserves the low-cost economics that helped the Nano line build its ecosystem.
NVIDIA has disclosed only that the new module and developer kit are expected during the first half of 2027. Pricing, specific module configurations and broader commercial details remain to come.
Why is Jetson Orin Nano 2 strategically relevant despite NVIDIA’s much larger data-centre business?
The direct revenue contribution from an entry-level embedded computer is unlikely to resemble NVIDIA’s multibillion-dollar data-centre GPU business in the near term. Its strategic value lies in establishing NVIDIA architecture inside devices that could eventually be produced at much larger unit volumes.
Robotics is still fragmented. Industrial robots, autonomous mobile machines, drones, home robots, agricultural equipment, inspection systems and smart cameras all have different performance and cost requirements.
An inexpensive common compute platform allows developers to prototype on NVIDIA software before deciding whether to deploy a smaller Jetson module or scale upward to more powerful hardware.
That makes Jetson Orin Nano 2 partly an ecosystem-acquisition product. A developer who builds perception, navigation and generative AI applications around NVIDIA’s stack has a reason to stay within the platform as the product progresses toward commercialization.
The first half of 2027 will determine whether the hardware itself is compelling enough. The larger strategic test will take longer: whether NVIDIA’s three-million-plus developer base converts into a broad installed fleet of robots and autonomous machines using Jetson as their standard compute layer.
For now, the most important technical shift is not simply 78 TOPS. NVIDIA says it can double inference performance while using 40% less power at equal performance, exactly the combination needed to push more capable AI away from centralized servers and into physical machines with limited batteries, cooling and space.
Discover more from Business-News-Today.com
Subscribe to get the latest posts sent to your email.
