Microsoft talk on October 12 say Maia 200 don dey serve production AI workloads, including Microsoft Copilot. E also announce plan to begin deploy Maia 300 for 2027. The company intend expand Maia family enter multigigawatt deployments starting from next year.
The announcement separate current use from future adoption. Microsoft say Maia 200 dey support im MAI model-development team, while GitHub Copilot and Microsoft Foundry suppose follow before 2026 end. Those later deployments na plans for the published account; company no say dem don finish.
Maia 300 mainly target inference: running AI models wey dem don train to produce answers or other outputs. Microsoft identify long-context reasoning, responsive conversations and long agent tasks as priorities. The architecture focus on reducing data movement between processing, memory and network, no be only raw computing power.
The published specification list eight HBM4 memory stacks and peak memory bandwidth of 18 terabytes per second inside 950-watt thermal design power. Microsoft say compute chiplets use TSMC N2P process, while other functions dey separate for the multi-chiplet architecture.
For rack level, Microsoft describe 96 accelerators inside 167-kilowatt power envelope. E expect 40 percent more tokens per megawatt than leading GPU systems, and claim up to seven times Maia 200 rack token throughput. Na company comparisons be those; this report no independently reproduce the measurements.
That difference matter because useful efficiency depend on workload and complete system, no be one chip specification alone. Microsoft's report explain intended design and deployment schedule, but no establish public price or performance advantage for every situation. Production deployment of the new generation still dey scheduled for 2027.