Microsoft said on October 12 that Maia 200 was serving production AI workloads, including Microsoft Copilot, and announced plans to begin deploying Maia 300 in 2027. The company intends to expand the Maia family to multigigawatt deployments starting next year.
The announcement distinguishes current use from future adoption. Microsoft said Maia 200 supports its MAI model-development team, while GitHub Copilot and Microsoft Foundry were expected to follow by the end of 2026. Those later deployments were plans in the published account, not confirmed completed rollouts.
Maia 300 is designed primarily for inference: running trained AI models to produce outputs. Microsoft identifies long-context reasoning, responsive conversations and extended agent tasks as priorities. Its architecture emphasises reducing data movement between processing, memory and networking resources rather than relying only on raw computing power.
The published specification lists eight HBM4 memory stacks and peak memory bandwidth of 18 terabytes per second within a 950-watt thermal design power. Microsoft says compute chiplets use TSMC's N2P process, with other functions separated into a multi-chiplet architecture.
At rack scale, Microsoft describes 96 accelerators in a 167-kilowatt power envelope. It expects 40 percent more tokens per megawatt than leading GPU-based systems and claims up to seven times Maia 200's rack token throughput. These are company comparisons, not independently reproduced measurements here.
The distinction matters because useful efficiency depends on the workload and complete system, not one chip specification. Microsoft's report explains its intended design and deployment schedule, but does not establish a public price or universal performance advantage. Production deployment of the new generation remains scheduled for 2027.