Microsoft says Maia will become the primary inference platform for its own products and services, scaling to multi-gigawatt deployments from 2027. The company also introduced Maia 300, which it claims delivers 40% more tokens per watt than leading GPU-based systems, while the previous generation Maia 200 is already running in the US, Europe, and Australia and produces up to 1.7x more tokens per watt in Teams and Copilot. At gigawatt scale tokens per watt, not peak teraflops, sets the cost of each Copilot answer, and an in-house chip cuts dependence on Nvidia while giving Microsoft control over the whole stack from silicon to model, a path Google took with TPU and Amazon with Trainium. Microsoft reportedly sought TSMC capacity for roughly 300,000 Maia 300 chips, but says it will keep deploying the best GPUs from the market alongside them.

Related: Microsoft Azure Launches Cobalt 200 VMs, Maia 200 Accelerator, and HorizonDB, Microsoft and Anthropic Adopt AMD AI Chips

Microsoft to expand Azure Maia system to multi-gigawatt deployments