Microsoft has officially launched its second-generation custom AI chip Maia200, aimed at further improving the AI inference efficiency of Azure cloud services and reducing reliance on expensive NVIDIA GPUs.
This Maia200 chip, manufactured by TSMC, has already begun shipping to Microsoft's data centers in Iowa and will subsequently be deployed in the Phoenix area. Scott Guthrie, Microsoft's corporate vice president for cloud and AI, said the chip is the "most efficient inference system" Microsoft has ever deployed. It will first be delivered to Microsoft's internal super AI team to generate high-quality data for training next-generation models and to power enterprise-facing Copilot assistants and the latest models from OpenAI.
Microsoft emphasized that the Maia200 outperforms similar custom semiconductors from Google and Amazon Web Services in specific AI tasks. To ensure continued leadership in the computing power race, Microsoft revealed that it is already designing the next-generation chip Maia300. In addition, as part of its collaboration agreement with OpenAI, Microsoft retains the right to use the startup's early chip designs if its own development is hindered, providing a "double insurance" for its underlying computing architecture.
