Microsoft unveils its own Maia 300 SI accelerator and plans multi-gigawatt deployments from 2027
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A major company is building its own SI hardware to depend less on a single supplier; more chip options mean wider access to compute.
Maia 300 is a third-generation accelerator built on TSMC's N2P process and designed mainly for inference, that is, running already trained models. According to Microsoft, a rack holds 96 accelerators at 950 W each (167 kW per rack), 27.6 TB of HBM memory and 1.7 PB/s of aggregate bandwidth. Production deployments begin in 2027, and Maia is meant to run Copilot, GitHub Copilot, Foundry and the MAI team's work on OpenAI, Microsoft and open-weight models. CNBC and The Verge headlines describe it as a step toward reducing dependence on Nvidia GPUs.