In summary
- Microsoft plans to scale its Azure Maia AI accelerator family to multi-gigawatt deployments starting in 2027, positioning Maia as its primary inference system.
- Azure Maia 200 is already in production across the United States, Europe and Australia, delivering up to 1.7 times more tokens per watt than leading GPU infrastructure.
- The successor Maia 300, built on TSMC's N2P process with 96 accelerators per rack, is set for production deployments in 2027 and promises 40% more tokens per megawatt.
Microsoft has announced it will scale its Azure Maia AI accelerator family to multi-gigawatt deployments starting in 2027, with the company describing Maia as the primary inference system for its products and services as it targets what it calls the era of "Super Intelligence".
The Azure Maia 200 is already in production, serving workloads including Microsoft Copilot and supporting the company's MAI model development team. GitHub Copilot and Microsoft Foundry are expected to follow by the end of the year.
Microsoft said Maia 200 is being deployed across the United States, Europe and Australia, and delivers up to 1.7 times more tokens per watt for AI capabilities in Teams, Copilot and Cowork experiences compared with leading GPU infrastructure.
The successor chip, Maia 300, is designed for long-context reasoning, latency-sensitive conversational workloads and multi-step agentic tasks. Microsoft said it expects the system to deliver 40% more tokens per megawatt compared with leading GPU-based systems, and up to seven times the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration.
The company said Maia 300 uses an inference-first architecture built around a software-defined dataflow design and AI-assisted kernel optimisation, intended to reduce data movement and improve performance per watt.
Each rack system brings together 96 accelerators within a 167-kilowatt power envelope, delivering 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth. Eight HBM4 stacks provide 18.0 terabytes per second of peak memory bandwidth within a 950-watt TDP, according to Microsoft.
Compute chiplets in the Maia 300 are built on TSMC's N2P process node, and a new multi-chiplet architecture separates compute, memory-controller and I/O functions.
The system also supports advanced low-precision formats including enhanced MXFP4, which Microsoft said reduces memory and bandwidth requirements for computation and KV-cache storage - enabling larger context windows at the same power envelope.
Maia 300 is scheduled to begin production deployments in 2027.
.jpg&h=142&w=230&c=1&s=1)
