AI News June 2 2026 - Microsoft Cuts the OpenAI Cord at Build, NVIDIA Drops Its Biggest Open Model Yet

1. Microsoft Build: Windows Is Now an Agent Platform - and Project Polaris Is the Surprise

Satya Nadella opened the Build 2026 keynote at Fort Mason Center, San Francisco at 9:30 AM PT with a clear thesis: Windows is no longer a platform for human users only. Agents are now first-class citizens in the runtime, the tooling, and the distribution model. The pre-announced items all shipped: Windows Agent Framework 1.0 reached general availability, the Windows Agent Store launched with 85% developer revenue share (early partners Adobe and Zoom), Azure Agent Mesh was announced for federated agent execution across Windows 365 and Azure Arc (GA Q4 2026), and WSL 3 shipped with paravirtualized GPU and NPU access for near-native Linux AI performance on Windows.

The surprise was Project Polaris. Microsoft unveiled its own in-house AI coding model - a mixture-of-experts architecture with specialized sub-modules tuned for different programming languages and frameworks. Polaris replaces GPT-4 Turbo as the default model for all GitHub Copilot subscribers starting August 2026. The migration is automatic; Microsoft is offering a three-month fallback period for teams that want to stay on GPT-4 Turbo. This is strategically significant: Microsoft is cutting its dependency on OpenAI's models for its most important developer product, while maintaining the broader OpenAI partnership for other services. Internal benchmarks show Polaris outperforming GPT-4 Turbo on HumanEval and MBPP, particularly for low-resource languages like Rust and Haskell.

Copilot Workspace also exited beta at Build - the feature that gives developers an AI-powered project planning environment inside GitHub, allowing full task breakdowns, file-level editing plans, and multi-agent PR workflows from a single interface. GitHub Copilot CLI reached GA in March 2026; Build expanded its agentic coding capabilities with multi-agent support inside VS Code and deeper GitHub-Azure integration. The Windows Agent Runtime Insider preview - text-based agents only initially, with vision-based agents targeting 2027 - is available to Insiders from today.

Full Build preview: Windows Agent Framework, WSL 3, Azure Agent Mesh, and Windows Agent Store explained ->  ·  Deep dive: Project Polaris - what changes, when, and who is affected ->

2. NVIDIA Nemotron 3 Ultra - The Most Capable US Open Model Yet, Still Behind China

Jensen Huang announced Nemotron 3 Ultra at the Computex 2026 keynote in Taipei on June 1, positioning it as the flagship of NVIDIA's open-weights model family. The model ships June 4 on Hugging Face, ModelScope, OpenRouter, and as an NVIDIA NIM microservice on build.nvidia.com. At 550 billion total parameters with only 55 billion active per token (mixture-of-experts architecture), it delivers over 300 output tokens per second - three to six times faster than comparable Chinese open models - while keeping inference costs significantly below a dense model of equivalent size.

The benchmark numbers: Intelligence Index score of 48 in US open-weights rankings, outperforming Gemma 4 31B and all prior Nemotron generations. The gap over Nemotron 3 Super (120B parameters, launched March 2026) is 12 index points - a large jump in this benchmarking landscape. NVIDIA has been transparent with the release: weights, training recipes, and data are published on the NVIDIA-NeMo/Nemotron GitHub repository and Hugging Face, including 2.5 trillion pre-training tokens for the Nano variant.

The honest caveat: Kimi K2.6 from Moonshot AI still leads the open-weight intelligence ranking overall. As Decrypt noted, Chinese open-source models jumped from roughly 1.2% of global open-model usage in late 2024 to around 30% by end of 2025. Nemotron 3 Ultra is the most visible result of NVIDIA's bet on becoming a full-stack AI platform company - not just a chipmaker - but it has not yet closed the gap with the leading Chinese open models. NVIDIA also confirmed Nemotron 4 is under development via the Nemotron Coalition, a group of eight AI labs including Mistral AI and Perplexity assembled in March 2026 to co-develop open frontier models on DGX Cloud infrastructure.

Full review: Nemotron 3 Ultra architecture, benchmarks, access methods, and how it compares ->

3. Pentagon Signs $422M AI Enterprise Deal Built on Azure Agent Mesh

The US Department of Defense announced a landmark enterprise software agreement at Build 2026, built on the Azure Agent Mesh platform announced earlier the same day. The deal is expected to save the Pentagon $422 million annually through AI-driven process automation across logistics, procurement, and administrative workflows. It is one of the largest government AI deployments announced on public infrastructure to date.

The agreement reflects a broader pattern: Azure Agent Mesh's announcement at Build was accompanied by a joint demo with Siemens and Rockwell Automation showing a "digital shift supervisor" agent adjusting assembly line speeds in real time, and the Pentagon deal extending the same federated agent execution model to government operations. Windows IoT Enterprise and Azure Stack HCI both support the same agent framework, allowing factory-floor and DoD-facility agents to run inference locally and sync with a cloud orchestrator - a sovereign-computing architecture that addresses the data residency requirements of both defense and regulated industry customers.

The timing is deliberate. Azure Agent Mesh reaching GA in Q4 2026 means the DoD agreement was structured around a forward commitment - the Pentagon is signing on the platform's roadmap, not just its current state. For other government buyers evaluating enterprise AI infrastructure, the DoD deal provides the reference deployment that procurement teams typically need before committing to a new platform at scale.

Full breakdown: Pentagon $422M Azure Agent Mesh deal - what it covers and why it matters ->