Microsoft Unveils Third-Generation Maia AI Chip, Reduces Reliance on Nvidia's GPUs

Microsoft has announced its third-generation Maia in-house AI chip, designed to be the primary AI inference system for Microsoft’s products and services. The new chip, called Maia 300, is a successor to the previous generation, Maia 200, which was only introduced in January this year. According to Scott Guthrie, executive vice president of Microsoft’s cloud group, Maia 300 has been designed with efficiency in mind.

The new Maia 300 chip will start rolling out to production AI workloads in 2027 and is expected to deliver significant improvements over its predecessor. In a full 96-accelerator configuration, it delivers up to 7x the rack-scale token throughput of Maia 200. Guthrie also stated that Maia 300 will provide 40 percent more tokens per megawatt compared to leading GPU-based systems.

The efficiency gains achieved by Maia 300 are expected to enable Microsoft to rely mainly on its in-house AI chips, rather than those from Nvidia and AMD. This shift is part of Microsoft’s plan to scale its Azure Maia family to multi-gigawatt deployments starting in 2027. Guthrie emphasized that this will make Maia the primary AI inference system for Microsoft products and services.

Maia 300 uses a new chiplet architecture, built on TSMC’s N2P process node, which separates compute, memory controller, and I/O functions. This design allows each function to use the technology best suited to its role, increasing the silicon area dedicated to compute performance. The Azure hardware and infrastructure team notes that this approach enables more efficient processing of AI workloads.

Microsoft is also using Maia 300 as a key component in its own AI model development. According to Mustafa Suleyman, Microsoft AI CEO, running their models on top of the Maia chip is crucial for achieving their goal of catching up with industry leaders like Google DeepMind and OpenAI. Suleyman emphasized that owning and co-designing a platform at this scale requires working across the full stack, from silicon to model development.