NVIDIA and Microsoft unify the stack for agentic AI on Windows, cloud, and on-premises

What happened?

At Microsoft Build 2026, Jensen Huang (CEO of NVIDIA) and Satya Nadella (CEO of Microsoft) revealed a significant expansion of their partnership to bring agentic AI to developers. The collaboration spans multiple fronts: specialized hardware (RTX Spark and DGX Station for Windows), a secure runtime called NVIDIA OpenShell integrated into GitHub Copilot, open models from NVIDIA in Microsoft Foundry, and GPU acceleration of Microsoft Fabric. The goal is for developers to build, fine-tune, and run AI agents natively on Windows, in Azure cloud, or on-premises, with a consistent and optimized stack.

Why is it important?

Agentic AI—autonomous systems that plan, reason, and execute long-running tasks—requires more than good models: it needs fast hardware, secure runtimes, responsive data layers, and models fine-tuned for extended reasoning. NVIDIA and Microsoft offer for the first time a complete stack covering from the user's device (PC with RTX Spark) to the desktop supercomputer (DGX Station) and the cloud. This democratizes agent development, allowing companies of all sizes to experiment and deploy autonomous AI solutions without relying exclusively on cloud infrastructure.

Key details of the announcement

RTX Spark and DGX Station for Windows

RTX Spark is the first line of Windows PCs designed specifically for personal agents, with 1 petaflop of AI performance, up to 128 GB of unified memory, and all-day battery. It arrives this fall from manufacturers including Microsoft Surface, ASUS, Dell, HP, Lenovo, and MSI. Meanwhile, DGX Station for Windows is a desktop supercomputer with the GB300 Grace Blackwell Ultra chip, up to 748 GB of coherent memory, and 20 petaflops of FP4 performance, capable of running models up to 1 trillion parameters. It will be available in the fourth quarter through ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro.

NVIDIA OpenShell: secure runtime for agents

OpenShell is an open-source runtime designed with security by default, integrated directly into GitHub Copilot. It allows agents to run in an isolated environment, protecting data and systems. This approach addresses growing concerns about security in autonomous agents, which can access sensitive systems and data.

Microsoft Fabric acceleration and models in Foundry

Microsoft Fabric, the unified analytics platform, is now accelerated with NVIDIA GPUs, improving performance for data and AI workloads. Additionally, NVIDIA's open models will be available in Microsoft Foundry, making it easier for developers to access models optimized for reasoning and planning.

Historical context

This is not the first collaboration between NVIDIA and Microsoft, but it is the most ambitious in the realm of agentic AI. In previous years, both companies worked together on integrating CUDA into Azure, developing the Azure AI platform, and optimizing models like GPT on NVIDIA hardware. However, the 2026 announcement marks a qualitative shift: it is no longer just about accelerating models, but about offering a complete ecosystem for autonomous agents, from client hardware to the cloud.

Impact and consequences

For developers and businesses, this unified stack reduces friction when moving from local experimentation to production deployment. AI agents will be able to run on high-performance PCs without relying on an internet connection, which is crucial for applications with latency or privacy requirements. For the market, the alliance strengthens Windows' position as an AI development platform against alternatives like macOS (with the M4 chip) or Linux solutions. Furthermore, the integration of OpenShell into GitHub Copilot suggests that Microsoft is betting on security as a key differentiator in enterprise adoption of agents.

However, success will depend on actual developer adoption and the availability of applications that leverage these capabilities. The price of the equipment (especially DGX Station) could limit its reach to large enterprises or research centers.

What readers should know

• RTX Spark PCs will be available this fall, while DGX Station will arrive in the fourth quarter of 2026.
• OpenShell is open source and integrates with GitHub Copilot, offering a secure runtime for agents.
• Microsoft Fabric accelerated with NVIDIA GPUs is already available for Azure customers.
• NVIDIA's open models (such as Nemotron) can be tested in Microsoft Foundry.
• This collaboration positions Windows as a serious platform for agentic AI development, directly competing with Linux or macOS-based alternatives.

"The moment of agentic AI has arrived, but fulfilling its promise requires more than good models. It also requires fast hardware, secure runtimes, a responsive data layer, and models fine-tuned for long-duration reasoning. NVIDIA and Microsoft are bringing that complete stack to developers." — NVIDIA Blog

Conclusion

The NVIDIA-Microsoft alliance represents a milestone in the evolution of agentic AI, offering a unified stack that spans from device to cloud. By providing specialized hardware, a secure runtime, and optimized models, both companies aim to accelerate enterprise adoption of autonomous agents. It remains to be seen how competitors will respond and whether the value proposition will justify the investment in high-end hardware.