Microsoft plans to begin deploying its Maia 300 accelerator in 2027 and make the Maia family the primary inference platform across its products and services. The strategy would reduce its dependence on Nvidia and AMD hardware for workloads such as Copilot, GitHub Copilot and Microsoft’s own models.
The company says Maia 300 is designed for long-context reasoning, low-latency conversations and sustained agent tasks. It claims the chip will deliver 40 percent more tokens per megawatt than leading GPU-based systems, while a 96-accelerator rack could produce up to seven times the token throughput of a Maia 200 rack.
Microsoft is using a chiplet design built on TSMC’s N2P process, separating compute, memory control and input-output functions. The approach is intended to dedicate more silicon to computation while matching each supporting function to an appropriate manufacturing technology.
These performance figures are Microsoft’s projections and have not yet been validated in broad production use. Maia 200 is already serving workloads in several regions, but the newer chip will not enter production until next year. The important strategic change is vertical integration: Microsoft wants to co-design models, systems and silicon at multi-gigawatt scale rather than treat accelerators as interchangeable cloud components.