Microsoft Unveils Maia 300 Chip to Reduce Nvidia Reliance
Microsoft is deploying multi-gigawatt scales of its proprietary Maia 300 AI chips to power Copilot services and lower dependence on external GPU providers.
Microsoft unveiled the Maia 300, a third-generation in-house AI accelerator designed to serve as the primary inference system for its products and services. Announced at the OCP Global Summit in San Jose, the chip utilizes a multi-chiplet architecture manufactured by TSMC using 2-nanometer and 3-nanometer process nodes. The company plans to scale the Maia family to multi-gigawatt deployments, with some reports indicating a rollout starting next calendar year and full production scaling by 2027.
The Maia 300 is optimized for long-context reasoning and agentic tasks, such as those used in Microsoft 365 Copilot and GitHub Copilot. Microsoft claims the hardware delivers 40% more tokens per megawatt than leading GPU-based systems and offers seven times the rack-scale token throughput of the previous Maia 200. The chips will be deployed in liquid-cooled racks of 96 accelerators, which can be grouped into superpods of up to 12,288 units.
This strategic shift aims to reduce the company's reliance on Nvidia and AMD for AI workloads, although Microsoft will continue to partner with both firms for model training. The announcement coincided with Microsoft's market capitalization surpassing $4 trillion. While the Maia 300 prepares for deployment, the Maia 200 is already active in the United States, Europe, and Australia.
