SAN JOSE, USA, Oct 12, 2026 /Commerce Wire/
Microsoft announced at the Open Compute Project Global Summit that its Azure Maia family of custom AI accelerators will scale to multi-gigawatt deployments, starting in 2027. Maia is set to become the primary inference hardware stack for Microsoft’s cloud and consumer services.
Scott Guthrie, Executive Vice President of Microsoft Cloud + AI, outlined the roadmap on Microsoft’s official blog. The Maia lineup spans Maia 100, production-ready Maia 200 and newly unveiled Maia 300. Maia 300 is scheduled for rollout in 2027. Each rack hosts 96 accelerators, double the count of Maia 200 racks, delivering up to 7x higher token throughput.
Microsoft claims Maia 300 achieves a 40% improvement in token throughput per megawatt compared with leading GPU systems for AI inference. Built on TSMC N2P chiplet design, the chip carries 288GB HBM4 memory and delivers 13.8 PFLOPS FP4 performance. It is optimized for large language model inference, powering Microsoft Copilot, Azure AI Foundry and OpenAI workloads.
Industry observers note that the multi-gigawatt Maia rollout aims to reduce reliance on third-party GPUs and gain better control over AI infrastructure supply chain and economics. Focused on inference workloads, Maia complements training silicon to sustain Microsoft’s massive cloud AI expansion.
Microsoft has not released a precise timeline for reaching multi-gigawatt capacity, saying the ramp-up will adapt to AI demand and industry technology cycles.