Nvidia Puts Agentic AI on Every Robot and Factory Floor with JetPack 7.2

In short, an AI agent that previously required a server rack can now, with a single command, run directly on a Jetson Orin module that retails for as little as $249. This isn't just about making robots smarter; it's about creating a unified, secure, and portable software stack that locks developers into Nvidia's ecosystem from the cloud all the way down to the factory floor.

NemoClaw: The Secure Bridge to Physical AI

The central piece of this puzzle is Nvidia NemoClaw. It's not a new model, but an open-source reference stack that solves a critical problem for enterprise deployment: security. NemoClaw bundles the popular OpenClaw multi-agent orchestration framework with OpenShell's secure runtime and Nvidia's Nemotron models. This provides privacy and security guardrails out of the box for always-on assistants operating in sensitive environments like factories, hospitals, and warehouses.

The critical update is that JetPack 7.2 comes pre-configured for one-command NemoClaw deployment. Previously, running OpenClaw-based workflows on an edge device required manual environment setup and dependency wrangling. Now, it works out of the box. This dramatically lowers the barrier for developers to build complex agentic workflows—an agent that senses, reasons, and acts on physical data—without ever needing a cloud round-trip.

To complement this, Nvidia released a major collection of open-source physical AI agent skills and tools spanning its Omniverse, Cosmos, Alpamayo, and Metropolis platforms. Complex robotics training, autonomous vehicle simulation, and industrial digital twin workflows can now be turned into tasks that an AI agent can directly execute.

Inside JetPack 7.2: The Hardware-Software Features

JetPack 7.2 is a foundational update that brings the edge platform into architectural alignment with Nvidia's server-grade offerings.

• CUDA 13 on Jetson Orin: This is a unification milestone. Jetson Orin now runs on the same Arm Server Base System Architecture (SBSA) CUDA 13 toolkit as Nvidia's data center GPUs. A single CUDA toolkit can now target everything from a GH200 superchip to a tiny module inside a delivery robot.
• MIG (Multi-Instance GPU) on Jetson Thor: This is a crucial feature for mixed-criticality applications like humanoid robots. MIG allows the GPU on a Jetson Thor SoC to be partitioned into two fully isolated instances, each with its own dedicated memory, cache, and compute. A safety-critical task like Simultaneous Localization and Mapping (SLAM) can run in a completely deterministic environment, safe from interference by less time-sensitive AI tasks.
• Yocto Project Support: For OEMs and system builders, official Yocto Project integration means they can build deeply customized, production-hardened embedded Linux distributions for Jetson hardware that are optimized to the byte for a specific industrial application.
• Super Mode for AGX Orin 32GB: A significant performance uplift and a strategic cost play. The new 32GB module replaces the earlier 16GB production variant, delivering a substantial generational AI performance gain while delivering greater cost efficiency for high-volume edge deployments.

The Strategic High Ground in Physical AI

This is a chess move aimed squarely at Qualcomm's Snapdragon dominance in edge and Intel's x86 hold on industrial controllers.

• Competitive Moat: By combining a unified CUDA server-to-edge toolkit, out-of-the-box secure agentic deployment (NemoClaw), and hardware isolation features (MIG) into a single integrated platform, Nvidia creates a developer experience that rivals struggle to match. A Qualcomm developer has to stitch together multiple fragmented stacks to achieve the same deterministic, secure agentic deployment.
• Ecosystem Lock-in: "One-command deployment" is the stickiest two words in developer relations. An enterprise that builds its industrial agentic AI infrastructure on NemoClaw + CUDA 13 + Jetson faces massive software switching costs. The investment in Nvidia's board support package, agent skills library, and simulation tools becomes a foundational layer of their operations.
• Advantech's AI Factory Brain: The partnership with Advantech, a titan of industrial computing, is the proof-of-concept writ large. They announced an "AI Factory Brain" architecture using NemoClaw, Nvidia's Factory Operations Blueprint, RTX PRO, and Jetson Thor. This targets end-to-end operational intelligence, bringing decision-making agents from the IT office directly into real-time factory operations. Crucially, Advantech's leadership noted they continue to work with Qualcomm and Intel, signaling that the physical AI platform wars are now a multi-front conflict.

The "Agentic Computing Pattern"

Jensen Huang’s vision is disarmingly simple and profoundly ecosystem-driven. In his keynote and following press conferences, he described a single, repeating pattern that will become the new computing paradigm: an agent that is a model, wrapped in a harness, using tools that have skills, running in a runtime.

This pattern, he insisted, is "exactly the same" whether that runtime lives in a massive cloud data center or inside a robot arm's compute module. Companies will use different models and harnesses, but the architectural pattern is universal. With JetPack 7.2 and NemoClaw, Nvidia has ensured its Jetson platform is the first place this universal pattern boots up in the physical world.

The Broader GTC Taipei Picture

The JetPack 7.2 announcement was not in isolation. Nvidia orchestrated a series of major ecosystem reveals at the event, all pointing toward agentic and physical AI dominance:

• Jetson Thor, Vera Rubin NVL72, and Alpamayo each won a COMPUTEX 2026 Best Choice Award, signaling industry and media recognition that three of Nvidia's core platforms are best-in-class building blocks for the AI era.
• Nemotron 3 Ultra, a new 550-billion-parameter mixture-of-experts model, was announced, offering 5x faster inference at a claimed ~30% lower cost than leading open models. This provides the powerful "model" side of the agentic pattern formula.
• To fill the "runtime" side of the pattern for personal computing, Nvidia unveiled the RTX Spark PC chip. This chip, built in collaboration with MediaTek and Microsoft, is designed to put local AI agent capabilities directly into Windows laptops and desktops, competing with AMD, Intel, and Apple.

While some rumored announcements, such as a "DGX Station for Windows" or a Cadence chip design agent, were not confirmed by the primary sources from the event, the overall picture is unmistakable. JetPack 7.2 and NemoClaw are the linchpins in Nvidia's strategy to ensure that when AI agents go from being a digital curiosity to a physical infrastructure layer, they run on Nvidia hardware, with Nvidia software, and inside Nvidia's ecosystem.