Microsoft Unveils Maia 300 to Cut Reliance on Nvidia GPUs

Microsoft unveiled its third-generation Maia 300 AI accelerator on Oct. 12, outlining a long-term effort to make its own chips the primary architecture for AI models and reduce reliance on Nvidia GPUs.

Maia 300 is scheduled to enter production in 2027. Each data center rack will support up to 96 chips, with Microsoft targeting seven times more AI token processing per rack than its previous setup.

The company also expects Maia 300 to deliver 40% higher token-processing efficiency per megawatt than GPUs. Microsoft plans to deploy Maia chips in clusters spanning multiple gigawatts, a scale intended to support its expanding internal AI infrastructure.

Microsoft 365 Copilot will be among the first products to benefit from the performance gains. The Maia accelerators will initially be used inside Microsoft rather than offered for external rental, while Nvidia and AMD chips will continue to support model-training operations.

Evercore ISI analysts estimate Maia chips could account for about 10% of Microsoft’s AI chip spending by the first quarter of 2027. Production is the next major milestone for the new accelerator.