Overview
- Microsoft unveiled Maia 300 on Monday, describing a third‑generation, inference‑first accelerator it plans to make the primary system for serving AI requests across its products and services.
- The Maia 300 is a chiplet‑based design built with TSMC N2P compute chiplets, eight HBM4 stacks for high memory bandwidth, and a 96‑accelerator rack configuration that fits in a roughly 167‑kilowatt power envelope.
- Microsoft says Maia 300 targets about 40% more tokens per megawatt than leading GPU systems and up to seven times the rack token throughput of Maia 200, claims that the company says come from system‑level co‑design of hardware and software.
- Microsoft plans production deployments starting in 2027 and to scale the Maia family to multi‑gigawatt capacity, with reporting that manufacturing orders for 2027 could range from hundreds of thousands to over one million units.
- Maia 200 is already running production workloads such as Microsoft 365 Copilot and GitHub Copilot, Microsoft will continue to use Nvidia and AMD for training and some offerings, and external access plus independent benchmarks for Maia 300 have not yet been made available.