Shenzhen Zhilian Weichuang Technology Co., Ltd.

Microsoft's self-developed chip Maia 200 goes into production: performance per watt of token improved by 1.7 times

Publication Date: 2026-11-19 Source:China Power Grid Key words: Microsoft chips View count:

At the Ignite 2026 event held on November 17, Microsoft announced that its second-generation self-developed AI accelerator, Maia 200, has officially entered large-scale production and has begun running Microsoft's largest cutting-edge AI large model in data centers.


Compared to the "leading GPU infrastructure," this accelerator delivers up to 1.7 times the processing per watt of tokens in built-in AI features like Microsoft Teams, Copilot, and Cowork.


12.png


Note: The Maia 200 is a custom chip designed by Microsoft for AI inference, using TSMC's 3nm process, featuring FP4 and FP8 tensor cores, 216GB HBM3e memory, and 272MB on-chip SRAM; A single chip can deliver over 10 petaFLOPS of computing power at FP4 precision and can scale via Ethernet to a cluster of up to 6,144 accelerators.


Microsoft stated that the chip has been deployed in the US, Europe, and Australia, serving production workloads such as Microsoft Copilot, and will gradually support GitHub Copilot and Microsoft Foundry. The company also plans to scale Azure Maia systems to multi-gigawatt deployments by 2027, making it the primary inference infrastructure for Microsoft products and services.