Microsoft is taking a significant leap by launching two innovative AI models crafted entirely in-house. MAI-Voice-1 and MAI-1-preview represent the company’s commitment to expanding its own AI capabilities and reducing its reliance on external partners like OpenAI.
MAI-Voice-1 has already impressed many with its speed and clarity. It can generate a minute of audio in under a second using a single GPU—a performance that stands out even for seasoned developers. Integrated into Microsoft’s Copilot Daily and featured across various podcasts, this model is now available for testing through Copilot Labs.
MAI-1-preview marks Microsoft’s debut in large-scale foundation modelling. Despite being trained on about 15,000 NVIDIA H100 GPUs—far fewer than the over 100,000 GPUs that powered xAI’s Grok—this model focuses on smart data selection and optimal resource use. Mustafa Suleyman, Head of Microsoft AI, explains that the success of MAI-1-preview hinges more on strategic efficiency than on raw computing power.
Currently open for public testing on the LMArena platform, MAI-1-preview is gradually being integrated into various Copilot features. Although detailed benchmarks weren’t released, the model holds the 13th ranking on LMArena, with developers invited to apply for early API access.
This launch aligns with Microsoft’s broader plan to develop a dedicated AI infrastructure. With a new compute cluster powered by Nvidia’s GB200 chips, the company is poised to roll out specialised models for a range of applications across Windows, Office, and Azure.
An important part of the strategy is refining how these models behave post-training. Microsoft aims to prevent the AI from giving the misleading impression of having emotions or intentions—an issue that Mustafa Suleyman has openly addressed.
Despite remaining OpenAI’s largest backer with investments up to $13 billion, Microsoft is clearly working to diversify its approach. This initiative signals not a break, but a balanced strategy aimed at fostering both independence and ongoing collaboration.
