Overview
- Microsoft published technical specs and performance claims for Maia 300 on Monday, describing an inference‑first, multi‑chiplet accelerator it plans to begin deploying at scale in 2027.
- The chip uses TSMC 2nm compute chiplets plus 3nm I/O and controller chiplets and pairs those with eight HBM4 stacks that Microsoft says provide a peak 18 TB/s memory bandwidth inside a 950 W thermal envelope.
- Microsoft says a rack will hold 96 Maia 300 accelerators in a roughly 167 kW power envelope with 27.6 TB of HBM and 1.7 PB/s aggregate HBM bandwidth, and the company claims the design yields about 40% more tokens per megawatt than leading GPU systems.
- The company plans multi‑gigawatt rollouts starting in 2027 and has said it may deploy more than a million Maia chips next year, though those scale and efficiency outcomes remain company‑reported and unverified in independent tests.
- Microsoft expects Maia 300 to speed and expand inference for services like Copilot, GitHub and Teams, and the move follows a wider cloud trend of custom silicon to lower costs and reduce reliance on external GPU suppliers while depending on TSMC capacity for production.