Microsoft plans Maia family multi-gigawatt scale-up alongside Maia 300 production in 2027

Microsoft says Maia 300 production deployments will begin in 2027, and it plans to scale the Maia family to multi-gigawatt deployments starting that year.[1] The company plans to make Maia its primary inference system for its products and services.[1] Its cited workloads include long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks.[1] Bloomberg reported that Microsoft expects to stock multiple gigawatts of Maia chips in its server farms, with Maia 300 starting to roll out next year.[4] Those plans describe a future deployment, not a claim that Maia 300 is already in production.

A Maia 300 rack system contains 96 accelerators within a 167-kilowatt power envelope, with 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth, according to Microsoft.[1] Each accelerator uses eight HBM4 stacks, with up to 18 terabytes per second of peak memory bandwidth within a 950-watt TDP.[1] Microsoft says Maia 300 uses a multi-chiplet architecture that separates compute, memory-controller and I/O functions.[1] Its memory organization shifts from uniform access to locality-aware non-uniform access across chiplets.[3]

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.[1] The company says a 96-accelerator Maia 300 configuration can deliver up to seven times Maia 200’s rack-scale token throughput.[1] Microsoft also says Maia 300 offers 2.5 times Maia 200’s memory bandwidth and 33% more capacity, and approximately twice the HBM bandwidth per watt.[3] These are company-reported comparisons; the evidence does not provide independent test results or details of the configurations used for those comparisons.

Microsoft says Maia 200 is already serving production workloads, including Microsoft Copilot, and runs OpenAI models, its own MAI models and leading open-source model families.[1] The company expects GitHub Copilot and Microsoft Foundry to follow Maia 200 by the end of the year.[1] Microsoft also says Maia 200 delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Teams, Copilot and Cowork than leading GPU infrastructure.[1] The Maia 200 rollout offers context for Microsoft’s plans, but it does not establish Maia 300’s production performance.