Microsoft has officially introduced the Azure Maia 300, marking the third generation of its custom-built in-house AI accelerator family. Production rollouts for the Maia 300 are slated to begin in 2027, with plans to make it the primary inference system across Azure and Microsoft products. The company claims the new chip design delivers up to 40% more tokens per megawatt compared to leading GPU-based systems. Built using a chiplet architecture, a full 96-accelerator configuration of the Maia 300 yields up to seven times the rack-scale token throughput of its predecessor. The move highlights Microsoft’s broader strategy to scale in-house semiconductor capabilities and lessen its total dependence on external hardware providers like Nvidia.

Full report : Microsoft unveils its third-generation in-house accelerator, promising higher token throughput and major energy efficiency gains ahead of a 2027 rollout.