Nvidia unveiled its new RTX Spark personal AI computers and the OpenShell runtime alongside Microsoft security primitives at GTC Taipei at COMPUTEX, addressing critical privacy and hardware bottlenecks for local AI developers. According to Nvidia, the hardware lineup features 1 petaflop of AI compute and 128GB of unified memory designed to run personal agents securely on device.
NVIDIA RTX Spark and DGX Station Hardware Architecture
The RTX Spark hardware class brings 1 petaflop of AI compute and 128GB of unified memory to slim Windows laptops and ultra-efficient desktop PCs, according to Nvidia. These systems provide the necessary resources to run complex AI workflows locally without relying exclusively on cloud infrastructure. Alongside the consumer-focused Spark machines, Nvidia introduced the NVIDIA DGX Station for Windows, a deskside supercomputer for professionals equipped with a data-center-class GPU and CPU for local inference. According to Nvidia, these systems integrate Microsoft Windows for enterprise manageability, security, and compatibility.
Security and Privacy Infrastructure via OpenShell
Running personal agents locally requires robust security boundaries to protect user data and system applications. To meet this demand, the NVIDIA OpenShell runtime is coming to Windows, utilizing Microsoft’s new security primitives for agents. According to Nvidia, the OpenShell runtime provides policy capabilities that let users define specific permissions for agents, intelligently route queries to local models based on privacy thresholds, and disguise personal information before sending queries to cloud models. Open-source agent developers, including Hermes Agent and OpenClaw, are integrating OpenShell and Microsoft security primitives directly into their new Windows applications.
Ecosystem Expansion and Software Updates
The NVIDIA NemoClaw blueprint is expanding across the company’s full local AI lineup, spanning GeForce RTX, RTX PRO, RTX and DGX Spark, and DGX Station units, alongside new streamlined installers and support for Hermes Agent. According to Nvidia, users can expect double the inference performance on top agentic models via multi-token prediction in llama.cpp and vLLM, alongside multi-GPU optimizations for ComfyUI. Furthermore, H Company is releasing computer-use tools optimized for RTX and DGX PCs, including new models and an upcoming desktop agent harness. Adobe is rearchitecting its Photoshop and Premiere applications, Blender is adding NVIDIA DLSS 4.5 Ray Reconstruction, and NVIDIA is introducing RTX Video Frame Generation to ComfyUI. All announced updates are scheduled to arrive this fall alongside RTX Spark.
