To accelerate AI operations and reduce power consumption, there's a growing trend among companies to design their own cloud computing infrastructure. On October 12, Microsoft announced plans to expand its proprietary AI accelerator "Azure Maia" to a multi-gigawatt scale starting in 2027, and to deploy its third-generation AI chip, "Maia 300," in the same year.
The currently operational "Maia 200" is already being utilized in Microsoft Copilot. According to Microsoft, it is being deployed in the US, Europe, and Australia, and for built-in AI features like Microsoft Teams and Copilot, it can process up to 1.7 times more tokens per watt compared to existing GPU-based infrastructures.
The Maia 300 is a new model designed for inference involving long contexts, conversations where response speed is critical, and multi-stage AI agent processing. Microsoft states that it can increase the number of processed tokens per megawatt of power by 40% compared to previous GPU-based systems. In configurations equipped with 96 accelerators, Microsoft also claims that the token processing volume per rack will be up to 7 times greater than that of the Maia 200.
Microsoft intends to leverage its proprietary chips as its primary inference infrastructure, vertically integrating everything from AI models to software and hardware, thereby enhancing power efficiency and expanding AI processing capabilities.
Made with AI
