Microsoft to deploy Maia 300 chips, targeting 2027 production start

Microsoft announced it will primarily rely on its third-generation Maia AI accelerator to process incoming requests, a shift from its previous heavy dependence on Nvidia graphics processing units. The company aims to bolster its artificial intelligence stack by integrating this homegrown silicon into its Azure cloud infrastructure and services like Microsoft 365 Copilot.

Microsoft has long used Nvidia GPUs as the standard for AI workloads, filling its existing data centers and assisting in the training of models from OpenAI and Anthropic. However, the software giant is now expanding its reliance on in-house chips to meet unrelenting demand for advanced processors. Finance chief Amy Hood stated at a Goldman Sachs conference last month that utilizing its own chips has contributed to margin improvement over the past decade. For general-purpose computing, the company already uses central processing units from Intel and Advanced Micro Devices alongside its own hardware.

The new Maia 300 chip is designed to increase efficiency in data center operations. Microsoft expects to pack 96 Maia 300 chips into each rack, double the capacity of the Maia 200. This configuration allows the system to process up to seven times more AI tokens per rack compared to the previous generation. Additionally, the Maia 300 delivers 40% more tokens per megawatt than GPU-based systems. Alistair Speirs, a general manager at Microsoft, noted that power usage translates directly into operational costs, making new architectures essential for better performance and resource utilization.

Production of the Maia 300 is scheduled to begin in 2027. In a blog post, Executive Vice President Scott Guthrie stated that Microsoft is targeting multi-gigawatt deployments for the Maia chip family. However, Speirs declined to provide a specific timeline for full deployment, citing the rapid changes in the industry. He said the company reserves optionality to ensure it deploys the right silicon for the right models at the appropriate time. Jim Kleewein, a technical fellow at Microsoft, indicated that parts of Microsoft 365 Copilot should experience speed improvements with the new hardware.

Evercore ISI analysts estimated in a client note that Maia chips will represent approximately 10% of Microsoft's spending on AI chips by the first quarter of 2027. The firm recommends buying Microsoft shares. Despite the push for in-house solutions, Nvidia remains a significant partner. CEO Jensen Huang recently spoke at an event where Microsoft announced a Surface laptop containing an Nvidia processor. Speirs confirmed that Microsoft continues to partner with both Nvidia and AMD for model training.

Unlike tensor processing units from Google or Amazon's Inferentia and Trainium chips, Microsoft's accelerators are not available for hourly rental by external companies. Patrick Moorhead, CEO of advisory firm Moor Insights and Strategy, noted that no outside parties have yet tested these chips.