NVIDIA shipped JetPack 7.2 at COMPUTEX 2026, adding NemoClaw agentic AI skills to Jetson edge devices and cutting physical AI deployment timelines from weeks to days. The update brings CUDA 13 to Jetson Orin, Multi-Instance GPU support to Jetson Thor, and a 20% performance boost on Jetson AGX Orin 32GB reaching 241 TOPS.
What Happened
JetPack 7.2 is a full software stack update for the Jetson Orin and Thor device families. NVIDIA paired the release with NeMo framework integration for edge hardware, delivering four major additions:
- NemoClaw agent skills: prebuilt agentic workflow modules for robotics, industrial inspection, and automation pipelines
- CUDA 13 on Jetson Orin: the latest GPU compute stack now available on embedded hardware
- Multi-Instance GPU on Jetson Thor: multiple isolated AI workloads on a single module
- Yocto Linux support: customizable OS builds for production edge deployments
Memory optimization in JetPack 7.2 reduces VRAM requirements by 29 to 40% depending on workload. Companies already shipping on the platform include Solomon (humanoid robots), Advantech (factory automation), SandStar (AI retail), NoTraffic (intelligent traffic systems), and Zipline (autonomous delivery drones).
Why It Matters
Agentic AI at the edge means systems make decisions without cloud round-trips. For creators and developers building AI-powered physical products, from interactive installations to on-device video processing rigs, this lowers the barrier to production inference without a data center dependency.
The NemoClaw skill framework packages common agent patterns as drop-in modules rather than requiring code-from-scratch implementations. The Jetson documentation hub now includes NemoClaw agent skill references alongside CUDA 13 migration notes, making it the single stop for upgrading existing deployments.
Key Details
- Jetson AGX Orin 32GB: 241 TOPS, up 20% from prior JetPack
- NemoClaw agent skills available for both Orin and Thor families
- MIG on Thor enables isolated multi-tenant workloads on one device
- 29 to 40% memory reduction enables larger models on constrained hardware
- Yocto support targets stripped-down OS builds for production edge deployments
- Announced at COMPUTEX 2026 in Taipei
What to Do Next
Download JetPack 7.2 from the NVIDIA developer portal. The migration guide covers upgrading from JetPack 6.x and setting up NemoClaw agent skills for your target device family. If you are building AI workflows that need to run near a camera, sensor, or robotic arm rather than in the cloud, JetPack 7.2 is the current production target for the Jetson platform.
