Microsoft plans to make its self-developed chips the primary workhorses for AI inference, with over a million Maia 300 chips expected to be deployed next year.

Beating AI News Flash: Microsoft announced plans to scale up deployment of its in-house developed Maia chips to several gigawatts starting in 2027, positioning them as the primary inference platform for its AI products. According to Reuters, Microsoft may deploy over 1 million Maia 300 chips next year, further reducing its reliance on NVIDIA GPUs. The previous-generation Maia 200 is already in operation with Microsoft Copilot, supporting OpenAI, Microsoft’s own, and open-source models. The new-generation Maia 300 is optimized for long-context inference and multi-step agents, with its compute chiplets fabricated using TSMC’s 2nm process. Each chip is equipped with 288GB of HBM4 memory, delivering a memory bandwidth of 18TB/s. A single Maia 300 rack holds 96 chips, with a power draw of approximately 167kW – twice the number of chips in a previous-generation rack. Microsoft claims the new rack delivers up to 7 times the token throughput of the Maia 200, and generates 40% more tokens per megawatt of power than leading GPU systems.