Microsoft to rely more on its Maia chips for AI inference, reducing Nvidia dependence

Microsoft says its third-generation Maia 300 accelerator will mainly handle incoming AI requests, reducing reliance on Nvidia GPUs. The chip fits 96 units per rack and runs up to seven times more AI tokens per rack than GPUs, with 40% more tokens per megawatt.

Production starts in 2027, and Nvidia and AMD remain central to model training, while Maia chips are not rented to outside customers.