Microsoft and NVIDIA used their October 7 event in San Francisco to announce RTX Spark-powered Windows PCs and new tools for running AI agents locally. The lineup is led by the Surface Laptop Ultra, while Windows updates focus on securely managing agents and letting Copilot choose between on-device and cloud models.
Key takeaways
Microsoft’s new hardware targets developers, creators and gamers who want substantial AI compute on a Windows PC. The announcements also extend beyond laptops: compact RTX Spark desktops and a forthcoming Windows version of NVIDIA’s DGX Station are aimed at always-on agents and larger model workloads.
- Surface Laptop Ultra starts at $2,599 and is built around NVIDIA RTX Spark.
- RTX Spark configurations offer up to 128GB of unified memory and up to one petaflop of FP4 AI performance.
- Microsoft Execution Containers are generally available to help isolate and govern Windows agents.
- Laptop preorders opened October 7, with availability scheduled for October 16; compact desktops are expected in November.
Surface Laptop Ultra leads the RTX Spark lineup
Microsoft’s 15-inch Surface Laptop Ultra pairs a premium notebook design with RTX Spark, a platform combining an NVIDIA Grace CPU and Blackwell RTX GPU. The top configuration is described as having up to 20 CPU cores, 6,144 GPU cores and 128GB of unified memory, with the components connected at 600GB/s. Microsoft says the system can deliver up to one petaflop of AI compute for running larger models locally.
The laptop starts at $2,599, according to event coverage. Microsoft also showed it running GitHub Copilot and Gears of War: E-Day at 60 frames per second. RTX Spark systems are also planned from Acer, ASUS, Dell, HP, Lenovo, MSI and Gigabyte, spanning laptops and compact desktops. NVIDIA says the compact desktop versions are designed for continuous operation, a fit for developers who want local agents running in the background.
Windows adds controls for local AI agents
Microsoft announced general availability of Microsoft Execution Containers (MXC), Windows infrastructure that lets agents run in isolated environments under operating-system control. The company says the tools are intended to make agents easier for users and IT teams to secure, observe and govern. That matters as agents gain access to files and applications, rather than simply responding to prompts.
Microsoft also outlined a hybrid approach for Copilot: local models can handle work where privacy or reduced cloud usage is a priority, while more demanding tasks can use cloud models. Demonstrations included Copilot finding files across folders, preparing an archive and drafting an email for user approval. Microsoft said Windows ML will provide a runtime for deploying models across CPUs, GPUs and NPUs.
DGX Station brings larger workloads to Windows
NVIDIA previewed DGX Station for Windows, a deskside system based on its GB300 Grace Blackwell Ultra platform. NVIDIA says it will offer up to 748GB of coherent memory and up to 20 petaflops of FP4 performance, aimed at enterprise teams developing and running large models without leaving their Windows environment. Linux AI toolchains will remain accessible through Windows Subsystem for Linux. The system is expected later in 2026, while RTX Spark compact desktops are slated to go on sale in November.
Via NVIDIA Blog