Microsoft has launched its third-generation Maia AI chip, positioning it as the primary AI inference system for its products and services. The Maia 300 is engineered to enhance the cost and efficiency of running AI workloads in the cloud, while also supporting advanced reasoning models and agentic tasks. The chip is set to begin rolling out to production AI workloads in 2027, with Microsoft aiming to scale its Azure Maia family to multi-gigawatt deployments during that period.
Scott Guthrie, Microsoft's executive vice president of cloud group, highlighted the chip's efficiency, stating that it is expected to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems. In a full 96-accelerator configuration, the Maia 300 can provide up to seven times the rack-scale token throughput of its predecessor, the Maia 200. These improvements are anticipated to reduce Microsoft's dependency on external GPU providers like Nvidia and AMD.
The Maia 300 is built using a new chiplet architecture and TSMC's N2P process node, allowing for better separation of compute, memory controller, and I/O functions. This design increases the silicon area dedicated to compute performance and enables each function to use the most suitable technology. The chip is also integral to Microsoft's AI model development, helping the company compete with industry leaders like Google DeepMind, OpenAI, and Anthropic.
Microsoft's AI CEO, Mustafa Suleyman, emphasized the importance of owning and co-designing the platform for its AI models. He noted that running Microsoft's AI models on the Maia chip is a critical focus, as it allows the company to work across the full stack, from software to silicon. This strategic move is expected to strengthen Microsoft's position in the AI landscape and support its long-term goals.