Microsoft today releases two in-house AI models — MAI-Voice-1 and MAI-1-preview — in what the company frames as a deliberate step to build its own foundation stack alongside its longstanding OpenAI partnership.

MAI-Voice-1, a speech generation model, can produce a full minute of natural audio in under a second on a single GPU. It is already powering Copilot Daily and Podcasts, and a new Copilot Labs experience called Audio Expressions lets users paste text, select a voice and style, and download the result. MAI-1-preview is Microsoft’s first end-to-end foundation model, a mixture-of-experts system pretrained on roughly 15,000 Nvidia H100 GPUs. The model is now available for public testing on the LMArena leaderboard and will roll out to select text tasks inside Copilot in the coming weeks.

Microsoft AI CEO Mustafa Suleyman has made no secret of the company’s “off-frontier” strategy — building models that are strong but not necessarily state-of-the-art, optimized for real consumer companion use rather than benchmark records. The announcement arrives amid public strain between Microsoft and OpenAI, including the loss of Microsoft’s exclusive cloud-provider status for OpenAI and complaints from Sam Altman about compute shortfalls. The subtext is not subtle: after years as OpenAI’s biggest backer and cloud landlord, Microsoft is quietly building the tools to need it less.