Microsoft unveils Maia 300 AI accelerator, with deployment planned for 2027

Microsoft has unveiled Maia 300, its next-generation accelerator for AI workloads in Azure. The company says deployment is planned to begin in 2027.

Microsoft says Maia 300 is designed primarily for inference, including long-context reasoning, conversational AI and multi-step agent tasks. It claims the chip will deliver 40% more tokens per megawatt than other leading GPU-based systems. A rack with 96 Maia 300 accelerators is expected to provide up to seven times the rack-scale token throughput of a Maia 200 rack.

The design combines eight stacks of HBM4, with claimed peak bandwidth of 18TB/s, and a 950W thermal design power. Microsoft says a Maia 300 rack will contain 96 accelerators within a 167kW power envelope, with 27.6TB of HBM capacity and aggregate HBM bandwidth of 1.7 petabytes per second. Compute chiplets will use TSMC’s N2P process.

Microsoft expects Maia 300 to support AI services including GitHub, Microsoft Foundry, Copilot, Cowork and Microsoft Teams. The company says Maia 200 is already running production workloads, including Microsoft 365 Copilot and models developed by its Microsoft AI team. It claims Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry are expected to follow by the end of this year.