Microsoft announces Maia 200 is in production and details Maia 300

In its October 12 announcement, Microsoft says Maia 200 now serves production workloads, including Microsoft Copilot, and supports the work of its MAI team. The company specifies that this infrastructure runs OpenAI models, its own MAI models and families of open models.

Microsoft reports deployment in the United States, Europe and Australia. For AI features integrated into Teams, Copilot and Cowork, it claims up to 1.7 times as many tokens per watt as the reference GPU infrastructures it cites. The available document does not allow this gain to be independently verified or extended to all inference workloads.

The announcement also presents Maia 300 as a system designed for long contexts, memory-intensive tasks and token generation. Microsoft describes a multi-chiplet architecture separating computing, memory controllers and input/output. It also emphasizes reducing data movement and coordinating computing, memory and networking.

The technical appeal of this approach lies in optimizing the entire system. If the announced gains are confirmed under the intended operating conditions, they could improve the amount of processing achieved for a given level of energy consumption. At this stage, the reported move of Maia 200 into production is separate from the announced specifications for Maia 300.