Microsoft has introduced its first fully in-house artificial-intelligence models—MAI-Voice-1 for speech and the MAI-1-preview large language model—signaling a move to lessen its reliance on OpenAI technology. The work is led by Mustafa Suleyman, the former DeepMind co-founder who now heads the Microsoft AI division. MAI-Voice-1 can generate a full minute of high-fidelity audio in under a second using a single graphics processor. The model already powers Copilot Daily and Microsoft’s podcast features and is available in Copilot Labs for developers to test expressive, multilingual speech synthesis. MAI-1-preview, trained on roughly 15,000 Nvidia H100 GPUs, is Microsoft’s first end-to-end foundation text model. It is being benchmarked on the public LMArena site, and Microsoft plans to integrate it into select Copilot text tasks in the coming weeks while offering early access to external developers. The rollout underscores Microsoft’s dual strategy of advancing proprietary AI while maintaining its $13 billion partnership with OpenAI. Suleyman said the company has a five-year roadmap that includes expanding its data-center capacity with Nvidia’s next-generation GB200 chips, aiming to reach billions of users through Microsoft products even as competition with OpenAI intensifies.