According to CoinMeta, Microsoft has announced plans to expand the deployment scale of its self-developed Maia chips to several gigawatts starting from 2027, making them the primary inference platform for its AI products. Reuters reports that Microsoft may deploy over 1 million Maia 300 units next year, further reducing its reliance on NVIDIA GPU. Currently, the previous generation of Maia 200 chips is already in use in Microsoft Copilot, supporting OpenAI models, as well as Microsoft's own and open-source models. The new generation of Maia 300 chips is optimized for long-context inference and multi-step agent tasks, using TSMC's 2nm manufacturing process. Each chip is equipped with 288GB of HBM4 video memory, and the memory bandwidth reaches 18TB/s. A single cabinet can accommodate 96 chips, with a power consumption of about 167kW, which is twice that of the previous generation. Microsoft claims that the throughput of the new cabinets with token chips is up to 7 times that of the Maia 200 chips, and the amount of token that can be generated per megawatt of power is 40% higher than that of leading GPU systems.