Microsoft Announces New Maia 300 AI Chip to Reduce Dependence on Nvidia and AMD
Microsoft has introduced the third-generation Maia 300 chip, which will serve as the primary inference system for its products and services, aiming for multi-gigawatt production starting in 2027.
Microsoft has officially announced the Maia 300, the third member of its in-house AI chip family, which will serve as the core inference system across its products and services.
Next-Generation Chip and Architecture
Microsoft introduced its new Maia 300 chip, developed to increase the cost efficiency and performance of AI workloads in the cloud environment. The new architecture achieves efficiency by separating computing, memory controller, and I/O functions from one another.
Built on TSMC's N2P process node, this new structure allows silicon area to focus more heavily on computational performance.
Efficiency and Performance Goals
Scott Guthrie stated that they expect the Maia 300 chip to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems.
In a full 96-accelerator configuration, this system can deliver up to seven times the rack-scale token throughput compared to the Maia 200 model.
Production Timeline and Market Impact
Microsoft plans to scale the Azure Maia family with multi-gigawatt deployments starting in 2027, thereby reducing its reliance on Nvidia and AMD hardware.
The Maia 300 production roll-out is projected to begin in 2027, becoming the primary AI inference infrastructure within the company's own products and services.
Current Infrastructure and Future Vision
Announced in January, the Maia 200 model is already successfully running production workloads such as Microsoft Copilot and GitHub Copilot.
Mustafa Suleyman emphasized that scaling up requires an end-to-end owned platform designed co-actively with the models.