Microsoft has for the first time detailed the design of Maia 300, its third-generation AI accelerator that will power Copilot, GitHub Copilot and Foundry. The chip is built from multiple chiplets on TSMC N2P, carries 288 GB of HBM4 memory and supports sub-3-bit precision formats. This architecture lets Microsoft's system handle more concurrent requests and run larger workloads, raising the overall capacity of the datacenter. While earlier Maia generations never fully replaced NVIDIA systems, Microsoft now plans to scale its Azure infrastructure on Maia 300 to several gigawatts starting in 2027, making the new chips the primary inference platform for the company's products and services.