Microsoft outlines multi-gigawatt rollout for Azure Maia silicon, previewing Maia 300 chiplet design
Maia 200 handles live Copilot workloads as Microsoft targets a 40 percent efficiency gain per megawatt for its custom AI hardware.
Technology news
Microsoft has disclosed architectural and deployment roadmaps for its custom Azure Maia AI accelerator family, confirming that Maia 200 is now running live production workloads across datacenters in the United States, Europe, and Australia. Executive Vice President Scott Guthrie announced that Microsoft will scale custom Maia systems into multi-gigawatt installations starting in 2027.[3]
According to Microsoft, Maia 200 currently serves inferences for Microsoft Copilot and internal MAI foundation models, delivering up to 1.7 times more tokens per watt than comparable commercial GPU clusters. Support is expanding to GitHub Copilot and the Microsoft Foundry enterprise platform before the end of the year, providing native execution for OpenAI models and open-source architectures.[3]
The company also previewed technical specifications for its third-generation Maia 300 processor, scheduled for 2027 production. Built using TSMC’s N2P process node on a multi-chiplet architecture, Maia 300 incorporates eight HBM4 memory stacks delivering 18.0 TB/s of bandwidth within a 950-watt TDP. In a full 96-accelerator rack, Microsoft projects a sevenfold increase in token throughput over Maia 200 and a 40 percent efficiency improvement per megawatt.[3]
Our take: custom datacenter silicon is critical for hyper-scalers seeking to rein in power consumption and reduce dependence on NVIDIA’s hardware supply chain. However, Microsoft’s performance projections reflect optimized internal inference workloads under low-precision MXFP4 formats rather than general-purpose training benchmarks. Real-world enterprise cost savings will depend on broad developer compiler maturity.