Microsoft는 내년에 자체 개발한 Maia 300 AI 칩을 배치할 계획이며, 메가와트당 토큰 출력을 40% 늘리고 Nvidia GPU 의존도를 낮추는 것을 목표로 하고 있다.

Microsoft plans to deploy its next-generation in-house AI chip, Maia300, in data centers starting next year, expanding the Maia series to a gigawatt scale and reducing its reliance on Nvidia GPUs. Designed specifically for inference, the Maia300 is projected to deliver 40% more token generation per megawatt than mainstream GPU systems. Its predecessor, the Maia200, already powers services such as Copilot. Unlike Amazon and Google, however, Microsoft currently has no plans to rent Maia compute capacity to external cloud customers.

The Maia chips are built for AI inference, the process of using existing models to perform tasks such as summarizing meetings and answering research questions. Microsoft expects the Maia300 to generate 40% more tokens per megawatt than mainstream GPU systems.

Rani Borkar, the executive who heads Microsoft's data center hardware and chip design team, said, "We are confident that we have built the best product for these workloads." She added that Microsoft will back this conviction with action by deploying Maia on a gigawatt scale starting next year.

Most of Microsoft's AI inference workloads still run on Nvidia GPUs. By customizing chips for its own data centers, the company hopes to improve inference efficiency and cut a portion of its spending on Nvidia GPUs.

Microsoft began deploying the Maia200 earlier this year. Borkar said the chip already powers AI software such as Copilot, running models from OpenAI, Microsoft's AI team, and open-source model families. The Maia200 has been deployed in data centers in the U.S., Europe, and Australia, supporting products such as Microsoft Teams and Copilot.

Custom AI accelerators have become an important route for major cloud providers to reduce their dependence on external hardware. Amazon and Google, which started earlier than Microsoft, are considered leaders in custom chips. Microsoft's expanded deployment aims to accelerate the use of in-house hardware for its internal AI services.

Microsoft does not currently plan to rent out the underlying compute power of Maia chips to cloud customers. Borkar noted that this differs from Google and Amazon, which already offer compute capacity from their in-house AI chips to customers.