NVIDIA and Microsoft Expand a Unified Agentic-AI Stack Across Windows, Azure, and Local Deployments

NVIDIA and Microsoft used the Microsoft Build stage to push a single, accelerated stack for agentic AI, aiming to let developers build and run autonomous software agents across three environments: Windows PCs, Azure cloud, and on-premises or hybrid deployments. The thrust of the announcement is that delivering practical “agents” requires more than frontier model quality, including low-latency hardware, secure execution, and data systems that can respond fast enough for agents that continuously query and reason.

At Build, NVIDIA founder and CEO Jensen Huang joined Microsoft chair and CEO Satya Nadella’s keynote via livestream to describe the expanded partnership. NVIDIA positioned the collaboration as an end-to-end path from local development to enterprise scale, spanning NVIDIA RTX Spark Windows systems, NVIDIA DGX Station for Windows deskside supercomputers, Microsoft Fabric and Foundry services, and NVIDIA software components for secure agent execution and optimized inference.

On the Windows device side, NVIDIA introduced RTX Spark as a new class of PCs “purpose-built” for personal AI agents. NVIDIA says RTX Spark systems deliver 1 petaflop of AI performance and up to 128GB of unified memory, with full AI and graphics performance intended to stay available even when unplugged. NVIDIA also said systems with RTX Spark will arrive this fall through partners including Microsoft Surface, ASUS, Dell, HP, Lenovo, and MSI.

For enterprise developers who need more sustained compute close to their data, NVIDIA described DGX Station for Windows as a deskside AI supercomputer designed for building and running agents inside Windows enterprise workflows. NVIDIA’s DGX Station for Windows is powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, with up to 748GB of coherent memory and up to 20 petaflops of FP4 performance, and it is positioned for always-on enterprise agents running frontier models up to 1 trillion parameters. NVIDIA said DGX Station for Windows systems are expected from partners in Q4 and will run NVIDIA OpenShell, described as a secure-by-design runtime for autonomous agents.

A central software component in the push is NVIDIA OpenShell, which NVIDIA says is designed to give agent workflows “real capability without real credentials.” In the Build presentation, NVIDIA and Microsoft described OpenShell as isolating each agent in its own sandboxed container and evaluating every outbound call against policy before access is granted to files, networks, or credentials. NVIDIA added that policies are written as code in versioned repositories and can be updated, and it says OpenShell is open source under the Apache 2.0 license and is model-agnostic across on-premises, hybrid, and cloud environments. NVIDIA also said OpenShell is integrated into GitHub Copilot.

To address the data side of agentic AI, NVIDIA said GPU-accelerated computing is built into Microsoft Fabric Data Warehouse, enabling faster SQL execution for high-concurrency workloads. The Nvidia blog attributes benchmarking results to Microsoft, citing up to 6x faster SQL execution versus CPU baselines and up to 7x faster than three other leading cloud data warehouse providers for demanding concurrency scenarios. The companies framed the goal as letting enterprise data layers keep pace with agents that continuously query and reason over internal systems, not just applications that generate a single response.

On Azure and “local” enterprise deployment paths, NVIDIA said it is expanding Foundry capabilities for agentic systems, including hosted agents in the Foundry Agent Service that can run models such as Anthropic’s Claude alongside NVIDIA and OpenAI models. NVIDIA also said Anthropic’s Claude models run natively on NVIDIA GB300 Blackwell Ultra systems on Azure, with customer availability expected in the weeks ahead. For on-premises and hybrid environments, Microsoft said it is pairing Foundry Local on Azure Local with NVIDIA’s RTX PRO 6000 Blackwell Server Edition, and NVIDIA said Foundry Local now supports multinode deployments and the vLLM runtime for scaling inference in latency-sensitive settings.

Beyond the core platform, NVIDIA highlighted tooling intended to help developers assemble agents into production workflows. NVIDIA mentioned an NVIDIA Agent Toolkit and NVIDIA NemoClaw blueprints for building production agents on Foundry, along with “domain-specific skills” exposed to agents through NVIDIA CUDA-X libraries such as cuDF, cuOpt, AI-Q, and NeMo. It also described a broader model ecosystem, including Nemotron 3 Ultra, positioned as an open frontier reasoning model for long-running agents, as well as content safety and speech-recognition variants in the Nemotron 3.5 family.

The announcement also pointed to infrastructure progress meant to improve throughput and reduce cost per unit of agent output. NVIDIA said Microsoft’s Fairwater Wisconsin AI factory is live and connected to a similar facility in Georgia, using Grace Blackwell systems and infrastructure optimizations intended to improve token economics. It also referenced the Vera Rubin platform being in full production for deployment across Azure data centers, with claims of higher inference throughput per megawatt and lower cost per agentic token, plus software infrastructure such as NVIDIA Dynamo for accelerating inference and orchestration components like NVIDIA Grove. Still, the companies did not disclose pricing, contractual terms, or independent performance validation, and some product availability was described only at a high level by quarter or “weeks ahead,” leaving details for later.

What to watch next is how quickly developers can assemble the stack end-to-end and how security and governance features perform under real enterprise workloads. The most immediate milestones are the arrival timelines for RTX Spark PCs in the fall and DGX Station for Windows systems in Q4, plus the rollout window for Claude availability on NVIDIA GB300 systems in Azure. After that, attention will likely shift to whether the OpenShell-and-identity approach meaningfully reduces the operational risk of autonomous actions in production, and whether Fabric acceleration and Foundry Local multinode scaling meet the performance targets beyond vendor benchmarks.