- Microsoft and Nvidia expand local AI processing on Windows PCs.
- Microsoft introduces AI agent security and enterprise management controls.
Microsoft and Nvidia are expanding local AI processing on Windows PCs, allowing developers and businesses to run larger AI models and autonomous agents without relying entirely on cloud services.
The companies announced the updates at a Windows and Surface event in San Francisco. The announcements include new AI hardware, model-routing capabilities, and security controls for managing AI agents.
Microsoft calls the approach hybrid intelligence, which combines local processing with access to cloud-based models depending on the workload.
Running larger AI models locally
Nvidia’s RTX Spark platform is among the hardware introduced for local AI processing. It combines a Blackwell RTX GPU with a Grace CPU, supporting up to 128GB of unified memory and one petaflop of theoretical FP4 AI performance using sparsity, according to Nvidia.
The platform supports AI inference, model development, and workloads involving autonomous agents. Developers can run supported models directly on their devices without sending every request to cloud services.
Microsoft also detailed how it is optimising larger AI models for local processing. Its MAI Code 1.1 Flash coding model contains 137 billion total parameters and 6.8 billion active parameters, with 3-bit quantisation reducing its size by nearly 80%.
The optimised model supports a local context window of 256,000 tokens. Microsoft is also preparing support for an Nvidia Nemotron model with more than 70 billion parameters, using 2-bit quantisation to reduce its memory requirements to just over 20GB.
These optimisations are part of Microsoft’s efforts to reduce the computing resources needed to run AI models locally. The company said its hybrid intelligence approach is intended to help customers manage cloud computing budgets while retaining access to cloud-based models.
To support this, Microsoft is extending GitHub’s HydraFusion technology, which currently routes tasks between cloud-based models, to include models running directly on Windows devices.
Experimental support is scheduled for GitHub Copilot, GitHub Copilot CLI, and Visual Studio Code later in October, allowing supported development tools to route workloads between local and cloud-based models.
Separately, Microsoft announced support for llama.cpp in Windows ML, allowing developers to run open-source models locally. Windows ML supports inference across GPUs, neural processing units, and CPUs.
Managing security risks from AI agents
Microsoft also announced the general availability of Microsoft Execution Containers (MXC) on Windows 11, introducing runtime security controls for AI agents operating on workplace computers.
MXC provides an isolated environment for AI agents, allowing developers and IT administrators to restrict access to files, applications, and networks. These policies are enforced independently of the agent, preventing it or its generated code from granting itself additional permissions.
The system supports several isolation methods, including process containers, Windows sessions, and Windows Subsystem for Linux containers. Windows session containers run agents in separate operating system sessions, isolating their desktop, clipboard, and input from the employee’s active session.
MXC process containers on Windows also support three operating modes for managing agent permissions. Enforcement blocks unauthorised operations, while Learning blocks and records access attempts to help administrators identify the resources an agent requires. Permissive records operations that would otherwise be blocked without enforcing those restrictions.
Microsoft said several AI development tools already support MXC, including OpenAI Codex, GitHub Copilot, OpenClaw, and Nvidia OpenShell. Support from Anthropic Claude Code and other providers is planned.
Beyond containment, Microsoft is preparing identity controls through Entra and Agent 365 to distinguish AI agent activity from employee activity. These capabilities are intended to help administrators identify individual agents, investigate their actions, and restrict agent access without automatically blocking the employee using the device.
The company is also developing Microsoft Intune policies for managing MXC process containers across Windows 11 devices. These policies will allow IT administrators to control container creation requests and enforce resource restrictions across managed systems.
MXC’s containment capabilities are generally available, while the additional Entra, Agent 365, and Intune integrations are still planned.
Bringing local AI to workplace devices
Microsoft introduced the Surface Laptop Ultra, powered by Nvidia RTX Spark, as part of its new Windows hardware lineup. Higher-end configurations support up to 128GB of unified memory, allowing the device to run supported AI models exceeding 120 billion parameters locally.
The laptop starts at US$2,599 and is scheduled to become available on October 16. Microsoft also introduced the Surface RTX Spark Dev Box, priced at US$5,999, with shipments in the United States expected to begin in November.
Designed for desktop AI development, the Dev Box supports local inference, model evaluation, and testing. It comes with development tools and environments including Visual Studio Code, Git, GitHub Copilot, and Windows Subsystem for Linux.
Although Microsoft has positioned local processing as a way to manage cloud computing expenditure, it has not provided a direct comparison of the total costs of running equivalent AI workloads locally and through cloud services.
Other manufacturers, including ASUS, Dell, HP, Lenovo, and MSI, are also introducing RTX Spark-powered Windows PCs, with initial laptop shipments scheduled for October 16.
For larger workloads, Microsoft and Nvidia previewed DGX Station for Windows, powered by Nvidia’s GB300 Grace Blackwell Ultra Desktop Superchip. The system supports up to 748GB of coherent memory and delivers up to 20 petaflops of FP4 AI computing performance.
According to Microsoft, DGX Station can run AI models with up to one trillion parameters and support 32 or more simultaneous AI agents. The system is intended to provide shared computing resources for development, engineering, and research teams.
DGX Station for Windows is expected to become available later this year, although Microsoft has not announced a specific release date.
Microsoft also announced updates to Copilot that will bring additional local AI capabilities to Copilot+ PCs. With user permission, the assistant will be able to access relevant local files and recent activity, perform tasks such as organising files and troubleshooting devices, and use models running directly on the computer.
Microsoft said the features are expected to begin rolling out over the coming months, with availability varying by device, market, and hardware platform.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

