Microsoft says it will mainly use its third-generation Maia AI accelerator for incoming AI requests, as it increases use of homegrown chips and reduces reliance on Nvidia GPUs for some workloads.