Microsoft and NVIDIA unveil unified agent stack

At Microsoft Build on 2 June, NVIDIA and Microsoft set out a partnership that links Windows devices, Microsoft’s cloud and on-premise servers into one stack for AI agents, announced in a post by NVIDIA. NVIDIA chief executive Jensen Huang joined Microsoft chief executive Satya Nadella’s keynote via livestream from Taipei. The companies framed five threads: a refresh of Windows hardware, a secure runtime for autonomous agents, more open models on Microsoft’s hosted AI platform, acceleration in Microsoft’s data warehouse, and the next generation of Microsoft’s large-scale AI factories.

What the announcement contains

The hardware thread covers new laptops and deskside machines aimed at developers building local agents. RTX Spark laptops arrive in autumn 2026; the deskside DGX Station follows in Q4. Both run OpenShell, NVIDIA’s new secure runtime for autonomous agents.

OpenShell sandboxes every agent in its own container, checks every outbound call against a policy file that lives in your code repository, and ships as open source. It is now integrated with GitHub Copilot, GitHub’s coding assistant.

On the model side, NVIDIA’s open reasoning model, alongside models for speech recognition and content safety, lands this month on Foundry, Microsoft’s hosted AI platform. Anthropic’s Claude models now run on NVIDIA’s latest server chips on Azure, with customer availability in the weeks ahead.

In the data layer, Microsoft says its analytics warehouse runs markedly faster on the new setup, and NVIDIA’s open-weights model family is being used to extend Foundry into sovereign, on-premises and hybrid deployments for manufacturing, energy and latency-sensitive workloads.

On infrastructure, Microsoft’s AI factory in Wisconsin is now live, running hundreds of thousands of NVIDIA systems as a single factory, with a second in Georgia. The companies have validated NVIDIA’s next-generation platform for Azure deployment.

6× faster on Microsoft’s data warehouse against a CPU baseline, and up to 7× against three other leading cloud warehouse products, per Microsoft’s own benchmark.

What to weigh up for a small UK firm

The announcements are framed for enterprise buyers, with partner-sold the only route carrying public pricing. Three pieces, though, are within reach of a small UK team this quarter, worth a look before any partner conversation sets the terms.

• If you build with GitHub Copilot. OpenShell is the first credible safety layer for autonomous agents that ships open source and lives in your repo. Read its policy file format before your next agent touches a customer record. A stub policy today beats hand-rolling guardrails.
• If your agents pull numbers from a warehouse. Benchmark Microsoft’s data warehouse on the new setup against your current pipeline. Microsoft’s headline speedup comes from its own benchmark and will vary by workload, but the order-of-magnitude is real. The governance toolkit Microsoft published earlier in 2026 gives you the audit frame to compare.
• If you are planning a hardware refresh. Windows laptops built for always-on local agents arrive this autumn from the major manufacturers. Deferring a Q3 purchase decision by six weeks beats committing to a non-accelerated laptop today.

The stack is real, the partner route is the realistic one for a small firm, and the open-source pieces deserve a closer look.