There were some major announcements made at GTC 2026 in Taipei on June 1, 2026. A big part of that is NVIDIA DGX Spark Window PCs. But more importantly, there is an important Jetson announcement. JetPack 7.2 is being released!
JetPack 7.2 Comes to Jetson Orin
Jetson Thor introduced the JetPack 7 software stack. JetPack 7.2 brings that stack to Orin: Ubuntu 24.04, Linux kernel 6.8, and CUDA 13. That should mean that everyone is can work on the same software platform on the current Jetson offerings.
Jetson AGX Orin 32GB Super Mode
Just like the Orin Nano offering a Super mode, NVIDIA introduces the Jetson AGX Orin 32GB Super, moving from 200 TOPS to 241 TOPS. The GPU max frequency moves from 930 MHz to 1.3 GHz, and the maximum power envelope moves from 40W up to 60W.
NVIDIA shows the 32GB Super configuration gaining roughly 1.1x to 1.3x across several model workloads compared with the standard AGX Orin 32GB baseline.
Better Memory Management
Memory is one of the practical limits in edge AI.
The memory on the module is not always the memory available to the application. Some of it can be reserved for hardware blocks, firmware, kernel behavior, display, cameras, or services that the application does not use.
JetPack 7.2 introduces memory optimization agent skills for Jetson. NVIDIA’s examples cover the bootloader, kernel, and user space: DRAM carveouts, reserved-memory regions, unused camera subsystems, display modes, and unused user-space functionality. These agent skills should be able to help gather up some of that statically allocated memory so we can better utilize it.
Yocto Gets Official Support
Yocto is an open-source project that lets developers build custom Linux distributions for embedded systems.
For Jetson developers, the practical distinction is simple. JetPack on Ubuntu is the fast development path. Yocto is the product-image path.
With JetPack 7.2, Yocto Project is officially supported on Jetson. That includes validated recipes and images, OE4T contributions, CI/CD ownership, documentation, forums, and support.
Community Yocto work around Jetson has existed for a long time. Official support changes the risk calculation for product teams.
Production systems often want smaller images, fewer services, reproducible builds, tighter package control, and a software stack that is easier to audit and maintain.
Yocto is not the path every developer needs. If you are learning Jetson, building demos, or testing a new idea, the Ubuntu-based JetPack image is still the natural starting point.
If you are building a product around Jetson, official Yocto support is worth investigating.
A Thor-Only Note: MIG
JetPack 7.2 brings MIG support to Jetson Thor.
MIG, or Multi-Instance GPU, partitions the GPU into isolated GPU instances. This allows isolation of different GPU programs, so if there is an issue with one app running on a GPU, it doesn’t crash everything running there. This is a Thor only feature, as it is hardware based.
Jetson Skills
Jetson Skills are packaged instructions that an AI agent can follow for Jetson-specific tasks which are included in JetPack 7.2.
NVIDIA describes Jetson Skills for Linux customization, memory optimization, model inferencing and benchmarking, diagnostics, package recommendations, and model-serving recipes.
That maps to real Jetson work: BSP customization, clock settings, fan profiles, nvpmodel modes, memory audits, model benchmarks, diagnostics, and container recommendations.
NemoClaw on Jetson
NemoClaw is NVIDIA’s stack for running OpenClaw-style agent workflows with privacy and security controls.
JetPack 7.2 is now NemoClaw-ready out of the box. We’ll have to take a look at this in the near future.
Open Models
NVIDIA lists current open model support across Nemotron, Qwen, Gemma, and MiniMax.
That is useful, but the model name is only the start. On Jetson, the practical question is how to get it to work on a given Jetson. Which model, how much memory, which server and so on.
A small model may be the right choice for latency. A larger model may be better for instruction following. Quantization may decide whether the model fits at all. The inference server may matter as much as the model when prompts get long or multiple users are involved.
Jetson AI Lab has added useful resources for understanding which models to use under different circumstances:
https://www.jetson-ai-lab.com/
VSS Is the Application Example
VSS, or Video Search and Summarization, is NVIDIA’s application example for agentic edge AI.
The workflow combines video streams, recorded video, vision models, language models, retrieval, alerts, reports, and analytics.
That is a good Jetson example because it is not just a single model demo. A video analytics system has to deal with cameras, stored video, computer vision, language models, retrieval, user requests, and deployment details.
Here’s the VSS Blueprint repository:
https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization.git
And the VSS skills catalog:
https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization/tree/main/skills
Conclusion
For Orin developers, JetPack 7.2 is the main event: Ubuntu 24.04, Linux kernel 6.8, and CUDA 13 move onto the Orin platform. That enables a lot of the other new features and relationships in the Jetson community.
AGX Orin 32GB Super Mode, official Yocto support, memory optimization skills, NemoClaw, open model support, and VSS fill out the rest of the announcement. Some of those pieces are immediately practical. Some need hands-on testing. The big take away is that Orin moves onto the JetPack 7 software generation.
The post JetPack 7.2 Announcement – GTC 2026 Taipei appeared first on JetsonHacks.
