Microsoft says its second-generation Azure Maia 200 chip is now running production workloads, including Microsoft Copilot, and is being deployed across the United States, Europe and Australia, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. In a blog post, cloud and AI chief Scott Guthrie writes that Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for its products and services, and that Maia 300 arrives in 2027. Microsoft claims Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI in Teams, Copilot and Cowork, and that Maia 300 is expected to deliver 40 percent more tokens per megawatt than leading GPU-based systems and, in a 96-accelerator rack, up to 7 times the token throughput of Maia 200. Maia 300 pairs eight HBM4 stacks, giving 18.0 TB/s of peak bandwidth within a 950-watt TDP, with a multi-chiplet design built on TSMC's N2P node.
Our takeThese are vendor benchmarks, so treat the multiples as claims, not measurements, since Microsoft does not disclose the GPU baseline it compares against. The strategic signal is clearer than the numbers: a hyperscaler putting a custom accelerator at the centre of its own inference fleet is aiming to cut its dependence on Nvidia for the highest-volume, most commoditised part of AI. That is a real cost play. The risk is software, because custom silicon only wins if the tooling keeps pace with CUDA's ecosystem.