NVIDIA and Microsoft Launch RTX Spark Superchip for Local AI Agent Integration

Announced at GTC Taipei, the NVIDIA RTX Spark superchip combines an NVIDIA Blackwell RTX GPU with a 20-core NVIDIA Grace CPU. The GPU features 6,144 CUDA cores and fifth-generation Tensor Cores utilizing FP4 precision, while the CPU was developed through a collaboration with MediaTek. These components are linked via the NVLink-C2C interconnect, supporting up to 128GB of unified memory to manage large-scale AI and graphical datasets.

To address privacy and security concerns associated with AI agents, NVIDIA and Microsoft have introduced new Windows security primitives alongside the NVIDIA OpenShell runtime. This software stack provides identity management, containment, and policy enforcement, allowing users to control agent actions and route queries to local models. Developers such as Nous Research and the OpenClaw Foundation are integrating these security layers into their respective Hermes Agent and OpenClaw applications.

The hardware is designed to handle intensive computational tasks locally, including the execution of 120-billion-parameter large language models with a 1-million-token context window. For creative professionals, the chip supports rendering 3D scenes larger than 90GB and editing 12K 4:2:2 video. Gaming performance reaches over 100 frames per second at 1440p resolution in AAA titles, supported by DLSS 4.5 Ray Reconstruction and Reflex technologies.

Adobe is currently rebuilding its flagship Photoshop and Premiere applications to run natively on the RTX Spark architecture, targeting a 2x improvement in AI and graphics efficiency. Other software partners, including Blender and Blackmagic Design, are also optimizing their tools for the platform. Hardware availability is scheduled for the fall of 2026, with initial systems coming from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, followed by releases from Acer and GIGABYTE.