Microsoft Challenges OpenAI with First In-House AI Models

Microsoft has announced two in-house AI models, MAI-Voice-1 and MAI-1-preview, signalling a move beyond reliance on OpenAI while aligning with a broader plan to build a proprietary AI ecosystem. MAI-Voice-1 is described as Microsoft's first natural speech generation model and is already integrated into Copilot Daily for Samsung smart TVs and podcasts, while MAI-1-preview is the first fully homegrown foundation model, currently available for public testing on LMArena.

According to Semafor, Mustafa Suleyman, head of Microsoft’s AI division, emphasised efficiency and cost-effectiveness, noting that MAI-Voice-1 can run on a single GPU and MAI-1-preview was trained on approximately 15,000 NVIDIA H100 GPUs, compared with xAI’s Grok, which reportedly required more than 100,000 GPUs. Copilot remains primarily powered by OpenAI’s GPT technology, with the in-house models positioned as a long-term strategic capability rather than an immediate replacement.

Suleyman also outlined a five-year roadmap for major AI investments with quarterly funding boosts to accelerate the evolution of Microsoft's proprietary models, a move framed as ensuring autonomy should the competitive landscape shift.