The development

What happened

Microsoft says its Maia 200 accelerators are now serving production AI workloads, while the next-generation Maia 300 will begin deployment in 2027. The October 12 announcement sets out a broader plan to make Maia the primary inference system for Microsoft’s products and services, with multi-gigawatt expansion starting next year. The change combines current deployment of custom silicon with a longer-term capacity and efficiency roadmap.

Key facts

  • Maia 200 serves Microsoft Copilot and MAI model development. Microsoft plans GitHub Copilot and Microsoft Foundry support by year-end.

  • Microsoft says Maia 200 deployments span the U.S., Europe and Australia and run OpenAI, MAI and open-source model families.

  • Maia 300 targets long-context, memory-intensive inference and sustained agent tasks. Production deployment is scheduled for 2027.

  • Microsoft projects 40% more tokens per megawatt than leading GPU systems and up to seven times Maia 200 rack throughput with 96 accelerators. These are company comparisons, not independent measurements.

  • The announced Maia 300 rack combines 27.6 TB of high-bandwidth memory with a 167 kW power envelope, emphasizing system-level memory and power requirements.

The context

Why it matters

TideReach assessment: Bringing hardware design and large internal AI workloads closer together could give Microsoft more control over serving costs and constrained datacenter capacity. The important test will be whether Maia 300 delivers its projected efficiency on production workloads at the planned scale. A higher tokens-per-megawatt figure alone does not establish lower customer prices or total deployment costs, and the roadmap does not imply that Microsoft will stop using GPUs.