Microsoft announced the launch of its third-generation self-developed AI chip, Maia300, which is planned to enter production workloads in 2027, with the goal of becoming the primary AI inference system for Microsoft's products and services. The new chip focuses on improving the performance and energy efficiency of cloud-based AI inference, and is optimized for long-context reasoning, low-latency conversations, and multi-step agent tasks.

Scott Guthrie, Executive Vice President of Microsoft Cloud Group, stated that compared to leading GPU systems, Maia300 is expected to deliver 40% more tokens per megawatt; with a full configuration of 96 accelerators, its rack-level token throughput will reach seven times that of the previous generation Maia200. Microsoft plans to expand the deployment scale of the Azure Maia series to tens of billions of watts starting in 2027, gradually increasing the proportion of self-developed chips in its internal AI infrastructure.

Chip

Maia300 is the successor to Maia200, announced in January this year. Maia200 has been used in production-level AI workloads such as Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft's self-developed AI models. Guthrie said that Maia200 is currently one of the best infrastructures in Microsoft's chip cluster in terms of actual AI service metrics, and Maia300 will further adapt to the needs of next-generation AI applications.

As the demand for computing power from inference models and agent tasks continues to grow, Microsoft is enhancing inference efficiency, reducing infrastructure costs, and strengthening control over the AI computing power supply chain through self-developed chips.