Microsoft targets Maia 300 for primary inference role

Microsoft unveiled Maia 300 at the OCP Global Summit on October 12, 2026, positioning its custom AI accelerator as the future primary inference system for Microsoft products and services. Production deployments are targeted for 2027, when the company also plans to begin expanding Azure Maia to multi-gigawatt scale.

Microsoft expects the accelerator to deliver 40% more tokens per megawatt than leading GPU-based systems. It also plans to fit 96 chips into each rack, twice the Maia 200 rack count, with rack-level token throughput expected to reach up to seven times that of its predecessor. These remain company projections rather than independently verified benchmarks. Microsoft has not named the competing GPU systems used for its efficiency comparison or disclosed pricing, detailed chip specifications, or the methodology behind its throughput target.

The strategy focuses on inference rather than replacing GPUs across all workloads. Microsoft says it will continue buying Nvidia and AMD processors, while training remains heavily dependent on GPU platforms. Maia 200 already supports workloads involving OpenAI models, Foundry and Copilot, giving Microsoft an existing production foundation. Manufacturing yields, memory packaging, software compatibility and available data-center power remain potential constraints on the planned rollout.