October 12, 2026
Microsoft makes its own chip the main inference engine: Maia 300 delivers 40% more tokens per megawatt
On Oct 12, Microsoft announced Maia 300: 40% more tokens per megawatt than leading GPU systems. The chip is designed as the main inference system for Microsoft's products and services, which reduces the company's dependence on Nvidia GPUs.
Maia 200, its predecessor, is built on TSMC's 3 nm process with 216 GB of HBM3e and already runs in production in the US, Europe and Australia. Copilot runs on it now. Maia 300 gets eight stacks of HBM4 and 18 TB/s of memory bandwidth. According to Microsoft, a rack of 96 accelerators delivers up to 7 times the token throughput of a Maia 200 rack.
The figures are still vendor claims. The comparison is against “leading GPU systems” without naming models. The material says nothing about pricing for Azure customers.
Production deployments of Maia 300 and a scale of several gigawatts are promised from 2027.