MSFT is scaling its Maia chips as the primary inference system for its products and services, with multi-gigawatt deployments planned and Maia 300 targeting 40% more tokens per watt than leading GPU systems. The post says this could reduce NVDA and AMD’s role in inference, though they remain used for training; Maia 200 is reported to cut costs 30–40% versus comparable NVIDIA hardware.