NVIDIA and Microsoft unveil unified agentic AI stack across Windows devices, Azure, and local deployments

NVIDIA and Microsoft used Microsoft Build on June 2, 2026 to announce a coordinated agentic-AI stack that spans Windows PCs, Azure cloud, and local enterprise deployment, plus a concrete hardware roadmap that includes RTX Spark, DGX Station for Windows, GB300 Blackwell Ultra on Azure, and Vera Rubin systems. Anthropic's Claude models were named as a launch partner running natively on the new Blackwell Ultra Azure systems.

What's new

NVIDIA frames the moment plainly: "The agentic AI moment has arrived, but delivering on its promise requires more than good models." The companies used Build to ship a coordinated stack across silicon, OS, cloud, and data platform. Concrete product disclosures include:

• RTX Spark. NVIDIA describes it as "a new beginning, powering the world's first Windows PCs purpose-built for personal agents, with 1 petaflop of AI performance, up to 128GB of unified memory."
• DGX Station for Windows. Positioned as "the most powerful deskside AI supercomputer for building and running agents on Windows enterprise applications and workflows."
• Anthropic on GB300 Blackwell Ultra. "Anthropic's Claude models now run natively on NVIDIA GB300 Blackwell Ultra systems on Azure, with customer availability in the weeks ahead."
• Microsoft Fabric acceleration. "NVIDIA accelerated computing is now built into Microsoft Fabric Data Warehouse, with Microsoft's internal benchmarking delivering SQL execution up to 6x faster than the CPU-powered baseline."
• Vera Rubin roadmap. "Vera Rubin slots in alongside Blackwell with no retrofits, delivering up to 10x inference throughput per megawatt."

Context

The announcement reads as the deepest NVIDIA-Microsoft alignment to date, covering hardware (RTX, DGX, GB300, Vera Rubin), cloud (Azure), data platform (Fabric), and developer surface (Windows, VS Code, agents). Both companies have been visibly racing to consolidate developer mindshare around an end-to-end agent stack ahead of competitors — AWS/Anthropic/Bedrock and Google/Gemini API/Vertex are the obvious counterweights, and Anthropic's named inclusion here despite its parallel AWS bets is the most strategically loaded detail of the post.

Why it matters

Two numbers in the post deserve scrutiny. The "6x SQL execution" claim against a CPU baseline in Fabric Data Warehouse is the kind of benchmark that, if it survives independent replication, reshapes data-warehouse procurement economics — historically a Snowflake-dominated category — by undercutting CPU-only competitors on raw query cost. The "10x inference throughput per megawatt" Vera Rubin claim, if accurate, is the kind of perf-per-watt step-function that data-center operators model their next three years of capex around, and is consistent with NVIDIA's pattern of one major generational leap per year. The Anthropic-on-Blackwell-Ultra disclosure also matters strategically: it puts Claude inference at performance parity on Azure in the same week Anthropic confidentially submitted its S-1, removing a hyperscaler-lock-in narrative that competitors had been pushing against the company.