Microsoft Aims to Build Over One Million Custom AI Accelerators by 2027

Microsoft is making significant strides in custom AI silicon development, joining major cloud providers like Google, Amazon, and Meta. The company's Maia custom AI accelerator has reached production deployments with its Maia 200 model, and it's already moving forward with the next-generation Maia 300.

The Maia accelerators are designed to handle AI inference tasks efficiently, leveraging lower power consumption compared to general-purpose GPUs like those from NVIDIA and AMD. The Maia 200 chip features over 140 billion transistors, built using TSMC's 3-nanometer process, with a memory subsystem including 216GB of HBM3e and providing up to 7TB/s of memory bandwidth.

Microsoft is reportedly negotiating with TSMC for production capacity that could support more than 300,000 Maia 300 chips in 2027, with longer-term plans potentially reaching over one million accelerators. The company aims to use these custom silicon components exclusively within its own data centers, further solidifying its commitment to in-house development.

The success of the Maia 300 will depend on Microsoft's ability to secure sufficient TSMC capacity, HBM supply, advanced packaging, and software support for large-scale deployment. This marks a significant milestone in Microsoft's custom AI silicon journey, with the potential to become a substantial part of its Azure compute strategy.