NVIDIA palm-sized AI computer

NVIDIA’s small-box AI pitch is no longer just about giant clusters in data centers. A May 16 YouTube video circulating around the company’s hardware ecosystem spotlighted a “palm-sized” machine for running models locally, framing it as a tool for developers, robotics builders and edge deployments rather than a miniature version of a rack server. NVIDIA’s own product and developer materials show that message rests on two distinct product lines: the Jetson family for embedded and robotics work, and DGX Spark for desktop-scale AI development. The smaller form factors do not replace NVIDIA’s larger systems, but they give the company a hardware story for local inference, prototyping and on-device development. ### Which NVIDIA machine is the video actually talking about? The May 16 YouTube video identified the device as the Jetson Orin Nano Super Developer Kit, describing it as a tiny computer that can run AI models locally without an internet connection. The clip pitched the system around privacy, robotics and low-latency inference at the edge. NVIDIA introduced the Jetson Orin Nano Super Developer Kit on December 17, 2024, calling it its “most affordable generative AI supercomputer.” NVIDIA said the kit “fits in the palm of a hand,” costs $249, and delivers up to 67 INT8 TOPS, with a 1.7x gain in generative AI inference performance over its predecessor after a software upgrade. (youtube.com) ### How is Jetson different from DGX Spark? (youtube.com) NVIDIA’s Jetson Orin Nano Super is an embedded developer kit built around a Jetson Orin Nano 8GB system-on-module and a reference carrier board. NVIDIA says it is aimed at prototyping edge AI applications and can support up to four cameras, with software ties to Isaac for robotics, Metropolis for vision AI and Holoscan for sensor processing. (blogs.nvidia.com) NVIDIA’s DGX Spark is a separate product. NVIDIA says DGX Spark is powered by the GB10 Grace Blackwell Superchip, delivers up to one petaFLOP of FP4 AI performance, includes 128 GB of memory, and is designed for developers to prototype, fine-tune and deploy reasoning models from a desktop system. NVIDIA previously referred to the product as Project DIGITS. ### Why does NVIDIA keep stressing local inference? (blogs.nvidia.com) The YouTube video emphasized offline use, privacy-sensitive workloads and low-latency robotics tasks. That framing matches NVIDIA’s Jetson materials, which describe the platform as a way to run AI close to sensors, cameras and actuators rather than sending every workload to the cloud. NVIDIA’s January 5, 2026 technical blog on the Jetson T4000 used similar language for a higher-end edge module. (nvidia.com) NVIDIA said the T4000 was built for “robotics and edge AI applications,” offering up to 1,200 FP4 TFLOPs, 64 GB of memory and real-time 4K video processing in tighter power and thermal envelopes. The post also highlighted TensorRT Edge-LLM and production-ready deployment. (youtube.com) ### Who is NVIDIA trying to reach with these smaller systems? NVIDIA’s December 2024 Jetson announcement named commercial AI developers, hobbyists and students as target users. The company also said the system was suited to people building generative AI, robotics and computer vision applications. The DGX Spark page points to a second audience. NVIDIA says the desktop machine is for developers working on reasoning models and autonomous agents, and it links to playbooks, a quick-start guide, a user manual, an unboxing video and a support forum. (developer.nvidia.com) That packaging turns the hardware into a developer on-ramp as much as a finished appliance. ### What can these machines do today? Jetson Orin Nano Super supports workloads such as retrieval-augmented chatbot development, visual AI agents and AI-based robots, according to NVIDIA’s launch post. (blogs.nvidia.com) NVIDIA also says developers can use Jetson AI Lab tutorials and the broader Jetson community to work with open-source models. DGX Spark comes with a preinstalled NVIDIA AI software stack, and NVIDIA says it can handle models up to 200 billion parameters at the desktop with its 128 GB of coherent unified memory. (nvidia.com) The product page also promotes NemoClaw, part of the NVIDIA Agent Toolkit, for building and evaluating autonomous agents locally. ### What should developers watch next? NVIDIA’s current product pages point developers to immediate next steps rather than a single launch event. (blogs.nvidia.com) The DGX Spark page includes marketplace purchase options, specifications, support materials and software guides, while Jetson materials direct users to Jetson AI Lab tutorials and NVIDIA’s robotics and vision software stack. NVIDIA’s January 2026 Jetson roadmap materials also show the company extending the same local-inference message upward into newer modules such as the Jetson T4000. (nvidia.com) For developers deciding where to start, the company’s live documentation now separates the embedded Jetson path from the desktop DGX Spark path, with both available through NVIDIA’s current product and support pages. (developer.nvidia.com)