Microsoft has officially launched its third-generation custom AI accelerator, the Maia 300, designating the Maia series as its primary architecture for running AI models. Scheduled for production in 2027, the chip aims to progressively reduce reliance on Nvidia GPUs. Each data center rack will support 96 chips, delivering a 7x increase in AI token processing per rack and 40% higher efficiency per megawatt compared to current GPU solutions. The company plans to deploy multi-gigawatt Maia clusters, with Microsoft 365 Copilot set to be the first beneficiary of the performance upgrades. Evercore ISI analysts project that Maia chips will account for approximately 10% of Microsoft’s AI chip spending by Q1 2027. The accelerators remain strictly for internal use, while Nvidia and AMD will continue to support the company's model training operations.