Microsoft Enters Mass Production for Maia 200 Custom AI Chip
Microsoft has officially started mass production of its second-generation custom AI accelerator, the Maia 200, which is now powering large-scale workloads across its data centers. The new accelerator delivers a 1.7x increase in tokens-per-watt efficiency compared to leading GPU infrastructure for features like Microsoft Teams, Copilot, and Cowork. This development marks a significant step for major tech companies seeking independence from third-party hardware vendors like Nvidia by utilizing custom silicon. By drastically improving token generation efficiency and planning multi-gigawatt deployments by 2027, Microsoft can scale its AI services more sustainably and cost-effectively. Built on TSMC's 3nm process, the Maia 200 features FP4 and FP8 tensor cores, 216GB of HBM3e memory, and 272MB of on-chip SRAM, offering over 10 petaFLOPS of compute power in FP4 precision per chip. It can be scaled via Ethernet into clusters of up to 6,144 accelerators and is slated to support GitHub Copilot and Microsoft Foundry.
## BACKGROUND
AI accelerators are specialized hardware designed to handle the heavy computational demands of training and running machine learning models much faster than standard CPUs. High Bandwidth Memory (HBM) stacks memory chips vertically to achieve ultra-high data transfer speeds, which is crucial for feeding massive amounts of data to modern AI models.