Microsoft unveils Maia 300 chip for AI computing
The company aims to reduce dependence on Nvidia and AMD by integrating its own accelerator into services starting in 2027
Microsoft has introduced the Maia 300 chip, which will serve as the foundation for the company's AI services. According to Scott Guthrie, Microsoft Vice President, the new accelerator is 40% more efficient than competitors in terms of tokens per megawatt of energy. In server configurations with 96 chips, performance will increase sevenfold compared to the previous model—the Maia 200.
The Maia 200 is already used in Microsoft products, including Copilot, GitHub Copilot, and the company's own AI models. The Maia 300 is optimized for tasks involving long context, logical reasoning, and AI agent operations. This will allow the company to gradually reduce its reliance on Nvidia and AMD graphics processors.
The chip is built on an architecture with separated functions: computing, memory management, and input-output are implemented on distinct chiplets. Production uses TSMC's N2P process technology, increasing the die area for computational blocks.
Microsoft plans to scale the Azure Maia platform to systems with capacities of several gigawatts. Mustafa Suleyman, Head of AI at Microsoft, noted that running Microsoft AI models on Maia chips will be a priority for competing with Google DeepMind and OpenAI.
