Microsoft Releases Maia 300 Chip, to Begin Multi-Gigawatt Deployment in 2027

en.Wedoany.com Reported - On October 12, Microsoft announced the launch of its third-generation in-house AI inference chip, Maia 300, and plans to begin deploying multi-gigawatt-scale Azure Maia computing systems in global data centers starting in 2027, positioning it as the primary inference platform for its AI products and services. Microsoft expects that Maia 300 will process 40% more tokens per megawatt than leading GPU systems, with full-rack token throughput reaching up to 7 times that of the previous-generation Maia 200.

Maia 300 adopts a multi-chiplet architecture that separates compute, memory controller, and input/output functions. The compute chiplet is manufactured using TSMC's N2P process, and the chip integrates 8 stacks of HBM4 high-bandwidth memory, achieving a peak memory bandwidth of 18.0TB/s with a thermal design power of 950 watts. The chip is optimized for long-context inference, memory-intensive computing, and multi-step AI agent tasks, and supports the enhanced MXFP4 low-precision compute format.

In terms of server system configuration, Microsoft's Maia 300 rack is equipped with 96 AI accelerators, with a total rack power of approximately 167 kilowatts, integrating 27.6TB of high-bandwidth memory and achieving an aggregate memory bandwidth of 1.7PB/s. The chips communicate with each other through a dedicated high-speed interconnect network, employing an Ethernet-based transmission architecture that supports scaling from a single rack to data-center-level computing clusters.

Microsoft has already deployed the previous-generation Maia 200 in data centers in the United States, Europe, and Australia, and uses it for Microsoft Copilot and internal AI model inference services. The company states that Maia 200 achieves up to 1.7 times the token processing efficiency per watt of leading GPU infrastructure when handling related AI workloads. Products such as Microsoft Teams already use the related inference capabilities, and GitHub Copilot and Microsoft Foundry are planned to be further integrated by the end of 2026.

Maia 300 will begin production environment deployment in 2027, primarily providing computing power for Microsoft's own AI applications, covering Copilot, GitHub, Microsoft Foundry, and next-generation AI model inference tasks. Rani Borkar, head of Microsoft's data center hardware, stated that the company currently has no plans to directly rent out the underlying computing power of Maia chips to cloud customers, and will prioritize meeting internal AI service needs at this stage.