With new in-house models, Microsoft lays the groundwork for independence from OpenAI (arstechnica.com)

Microsoft has unveiled two new AI models trained entirely in-house, signaling a strategic move toward reducing reliance on OpenAI despite its deep investment in the company. Among the newly announced models is MAI-Voice-1, designed to generate high-quality, expressive speech for both single and multi-speaker interactions, highlighting Microsoft's bet on voice becoming a primary interface for AI in the future. The second, MAI-1-preview, is a large foundational language model specifically developed to power Microsoft's Copilot chatbot, representing a significant leap in scale and capability compared to the smaller, locally runnable models Microsoft previously explored. MAI-1-preview was trained using approximately 15,000 Nvidia H100 GPUs, yet it can perform inference efficiently on just one GPU, showcasing advances in model optimization and deployment. This internal model marks a shift from Copilot’s historical dependence on OpenAI’s technology, aiming to give Microsoft greater control over its AI ecosystem. While Microsoft and OpenAI remain closely linked, recent tensions suggest divergent priorities that could drive Microsoft to invest more heavily in self-sufficiency. This move is crucial for the AI/ML community as it underscores how major players are balancing collaboration and independence in foundational AI development, potentially reshaping the competitive landscape and accelerating innovation in scalable, efficient AI architectures.