Maia 300 is the third generation of Microsoft's accelerator and moves to a chiplet design, with compute dies on TSMC's N2P node. It uses eight HBM4 stacks for 18 TB/s of peak bandwidth inside a 950W envelope. Microsoft says a full 96-accelerator configuration delivers up to 7x the rack-scale token throughput of Maia 200. Production deployments start in 2027, and the Maia family is meant to become Microsoft's main architecture for running models. Maia 200 already serves Microsoft Copilot and runs OpenAI models, Microsoft's own MAI models and open-source families. GitHub Copilot and Foundry are due to move onto it by the end of 2026. Google with TPUs and Amazon with Trainium have taken the same route of shifting inference off Nvidia GPUs. For builders, more hyperscaler inference on custom silicon could push Azure per-token prices down. It could also mean more variation between clouds in which models and low-precision formats, such as MXFP4, run best.