From rented cloud to desktop, Nvidia and Microsoft move AI agents to Windows

For decades, building AI agents and developing advanced models has been tied to rented cloud accounts and escalating bills, while desktop computers inside organizations have been unable to run weights beyond traditional memory limits. In San Francisco, Jensen Huang, CEO of Nvidia, and Satya Nadella, CEO of Microsoft, set a limit to this separation, announcing a joint redesign of hardware and software that lets AI agents run continuously locally under the Windows operating system directly.

The shift began deep within the operating system, as Microsoft announced the public availability of the Microsoft Execution Containers (MXC) architecture, an OS-level infrastructure that provides agents with a secure, continuously running background environment under direct compute management, integrated with Microsoft security tools and the Entra 365 suite. Just as the DirectX interface once transformed the graphics software industry, this infrastructure establishes an operational standard that enables monitoring independent agents, controlling their behavior, and securing them without leaving the desktop environment.

The hardware equation for individual and developer machines is embodied in the RTX Spark chip.This hardware combines a Blackwell-based GPU with up to 6,144 cores and a Grace CPU with up to 20 cores, linked by a 600 GB/s bandwidth, delivering a compute capacity of one petaflop in FP4 format with unified memory of up to 128 GB. For the first time on integrated laptops and desktops such as the Surface Laptop Ultra and devices from Dell, HP, Lenovo and Asus, these specs enable running massive local models like the 125-billion-parameter Qwen 3.8 Flash-Next with 51 billion tokens, without any cloud inference usage or sending sensitive work data to the cloud.

The biggest breakthrough in institutional barriers comes in the teaser announcement of the DGX Station for Windows.Historically, DGX stations required Linux as a single environment to run heavy AI workloads, forcing developers in large Windows-based companies to manage two separate, complex environments. The new station arrives with a GB 300 Grace Blackwell Ultra Desktop chip, offering 748 GB of unified memory and 20 petaflops of FP4 compute power, sufficient to run and train models up to a trillion parameters right beside the desk, while keeping Linux tools available through the Windows Subsystem when needed.

This shift reshuffles operational priorities in tech offices, Gulf banks, and startups in Egypt and the Levant. Financial institutions and government agencies that faced strict data-localization regulations and hesitated to route their sensitive records through international cloud interfaces can now move multi-task agents to their closed office devices without recurring inference costs. The hardware also frees regional development teams from the costly transition between runtime environments, allowing agent automation to be managed with the same governance and security tools already approved in the enterprise infrastructure.

Moving AI agents from distributed clouds to office-installed processors forces technology leaders to re-evaluate budgets, balancing capital expenditure on high-end local hardware against monthly subscriptions for external inference services. Tracking the availability of these devices, from mid-October for laptops through November for integrated computers and desktop stations, represents the first step in calculating the true cost of on-premise AI for your work environment.