Microsoft builds in-house MAI stack and Foundry IQ to cut OpenAI dependence

At and around Build 2026 (June 6), Microsoft laid out a clear strategy to lessen its dependence on OpenAI by building a full in-house AI stack. It moved the Agent 365 SDK to general availability, announced Work IQ APIs (launching June 16) for building agents over Microsoft 365 data, and unified its knowledge layer around Foundry IQ — spanning Work IQ, Fabric IQ, Azure SQL, file search, and a Web IQ component for live grounding — plus guided guardrails, enhanced tracing, and agent governance.

The model news is the bigger signal: Microsoft released four distinct in-house MAI model series — Alon-1.0, MAI-Thinking-1, MAI-Image-2.5, MAI-Voice-2 and MAI-Code-1-Flash — separate from the OpenAI models powering Copilot. Microsoft AI CEO Mustafa Suleyman touted that MAI-Transcribe-1.5 is 'in a league of its own' on Artificial Analysis benchmarks. A PCMag hands-on offered a more 'brutal truth' assessment, suggesting the MAI models are competitive in spots but not uniformly frontier-class.

Commercially, Microsoft 365 Copilot SKUs become permanent on July 1 at $23.50 and $32 per user per month, and a new GitHub Copilot desktop app manages tasks and pull requests. Satya Nadella highlighted real-world traction, noting NHS England is scaling Microsoft 365 Copilot to more than 500,000 staff who saved an average of 43 minutes per day in early trials.

Strategically this mirrors a wider hedging trend — Amazon adding OpenAI models to Bedrock, Apple leaning on Gemini — as every hyperscaler diversifies model supply. The open question is whether Microsoft's MAI models can match OpenAI quality enough to justify shifting Copilot workloads in-house, or whether they remain a negotiating lever and a fallback.