Microsoft has introduced Maia 300, its third-generation internally developed AI accelerator, scheduled for data center use in 2027. The chip prioritizes inference to process AI models for answers and tasks, supporting services such as Copilot and Azure. Each server rack is designed to hold 96 Maia 300 accelerators within about 167 kilowatts of power. Microsoft claims the configuration can generate about 40% more tokens per megawatt than leading GPU systems and about seven times the token output of the previous Maia 200 generation. The chip is built on TSMC’s 2-nanometer process and includes eight high-bandwidth memory stacks. Microsoft plans to expand the Azure Maia family to multi-gigawatt-scale deployments and position the hardware as a foundation for its AI ecosystem. Analysts have estimated Maia chips could represent a portion of the company’s AI hardware budget by early 2027, though that is not an official company target.
Microsoft has introduced Maia 300, its third-generation internally developed AI accelerator, scheduled for data center use in 2027. The chip prioritizes inference to process AI models for answers and tasks, supporting services such as Copilot and Azure. Each server rack is designed to hold 96 Maia 300 accelerators within about 167 kilowatts of power. Microsoft claims the configuration can generate about 40% more tokens per megawatt than leading GPU systems and about seven times the token output of the previous Maia 200 generation. The chip is built on TSMC’s 2-nanometer process and includes eight high-bandwidth memory stacks. Microsoft plans to expand the Azure Maia family to multi-gigawatt-scale deployments and position the hardware as a foundation for its AI ecosystem. Analysts have estimated Maia chips could represent a portion of the company’s AI hardware budget by early 2027, though that is not an official company target.