Microsoft unveils Maia 300, its most powerful AI chip yet - and it might make using Teams, Copilot and Cowork much better at last

Microsoft has unveiled Maia 300, its latest AI chip, designed to enhance the capabilities of its Teams, Copilot, and cowork products. This inference-first chip promises significant advancements in efficiency and power compared to its predecessor, the Maia 200. Set for deployment in 2027, Maia 300 will be deployed in multi-gigawatt data centers, aiming to deliver 40% more tokens per megawatt than other leading GPU-based systems.

Specifically tailored for inference, long-context reasoning, and responsive conversational AI, Maia 300 is expected to improve data generation, agentic coding in GitHub, and large-scale model serving. Its architecture optimizes data movement, memory usage, and networking resources to minimize latency. With an eight-stack HBM4 memory configuration and a peak bandwidth of 18TB/s, Maia 300 incorporates a multi-chiplet architecture with TSMC's N2P process.

The chip's software-defined dataflow and AI-assisted kernel optimization ensure efficient resource utilization. Networking capabilities are integrated into the dataflow, enabling tensor parallelism and custom AI transport based on Ethernet infrastructure. The launch of Maia 300 builds upon the success of Maia 200, which already powers Microsoft 365 Copilot and models developed by the company's AI team, offering up to 1.7 times more tokens per watt than leading GPU infrastructure. Microsoft Foundry and GitHub Copilot are expected to integrate Maia 300 by the end of the year.