Microsoft announced today its third-generation Maia 300 processor, positioning it as the central inference engine for the company’s cloud services. The chip is slated to enter production AI workloads in 2027, marking a shift toward internal silicon for Azure. By moving away from Nvidia and AMD GPUs, Microsoft aims to control more of its AI stack.
Maia 300 promises lower operating costs and higher efficiency while supporting the most recent reasoning and agentic AI models. It follows the Maia 200, introduced earlier this year, which already serves workloads such as Microsoft Copilot, GitHub Copilot, the Foundry platform, and Microsoft-developed models. The new silicon is intended to broaden those capabilities across the Azure portfolio.
The processor uses a chiplet architecture fabricated on TSMC’s N2P node, separating compute, memory control, and input-output sections onto distinct dies. This layout frees additional silicon area for raw processing power and lets each block employ the manufacturing technology best suited to its function, the Azure hardware and infrastructure team explained. “This increases the silicon area,” the team said, underscoring the efficiency focus.
Guthrie said the company will expand the Azure Maia line to multi-gigawatt scale beginning in 2027, positioning Maia as the default inference platform for Microsoft services. “We will scale our Azure Maia family,” he stated, indicating a strategic move to lessen dependence on third-party GPU providers. This roadmap may reshape the economics of AI compute within the firm’s cloud offering.
If the Maia 300 achieves its projected efficiency, Microsoft could lower the cost of running large language models and other AI workloads while gaining tighter integration with its software stack. The shift may also pressure Nvidia and AMD to accelerate their own AI-focused silicon strategies, as a major cloud provider prepares to rely primarily on its own custom chips for inference.