Microsoft Details Maia 300 AI Chip and Multi-Gigawatt Plan

· Tech · United States

Microsoft unveiled architecture details for Maia 300, its third-generation Azure AI accelerator, and said it will scale its in-house Maia chips to multi-gigawatt deployments starting in 2027. Maia will become the primary AI inference system for Microsoft's own products and services, including Microsoft 365 Copilot, GitHub Copilot and Microsoft Foundry, though the company will keep using Nvidia and AMD GPUs as well. Maia 300 puts 96 accelerators per rack with a two-tier network linking up to 12,288 accelerators into a single superpod, uses HBM4 memory and is mostly liquid-cooled, built for cloud-scale agentic inference. Production deployments begin next year, with the design based on TSMC's N2P process.

Why it matters：Microsoft's move to make in-house silicon the main inference platform for its own AI services is a major test of whether hyperscalers can reduce reliance on Nvidia and AMD GPUs at scale.

Key points

  • Maia 300 packs 96 accelerators per rack and links up to 12,288 into one superpod
  • The design uses HBM4 memory and is mostly liquid-cooled for cloud-scale agentic inference
  • Microsoft will still use Nvidia and AMD GPUs alongside Maia

Sources (original reporting)

More stories from this day