Microsoft has announced its third-generation Maia in-house AI chip, the Maia 300, designed to serve as the "primary AI inference system" for the company's vast array of products and services. The new silicon aims to improve both the cost and efficiency of running AI workloads in the cloud, specifically targeting the requirements of modern reasoning models and agentic tasks.

Efficiency and Performance Metrics

Scheduled to roll out to production in 2027, the Maia 300 focuses heavily on energy efficiency. Scott Guthrie, executive vice president of Microsoft’s cloud group, stated that the chip is expected to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems. In a full 96-accelerator configuration, it reportedly delivers up to seven times the rack-scale token throughput of its predecessor, the Maia 200.

These efficiency gains are central to Microsoft’s strategy to transition away from external suppliers. Guthrie noted that the company plans to scale the Azure Maia family to multi-gigawatt deployments starting in 2027, effectively making it the backbone of Microsoft’s AI infrastructure over Nvidia and AMD hardware.

Technical Architecture and Vertical Integration

The Maia 300 utilizes a new chiplet architecture built on TSMC’s N2P process node. This design separates compute, memory controller, and I/O functions, maximizing the silicon area dedicated to pure compute performance. According to the Azure hardware team, this allows each specific function to utilize the technology best suited for its role.

Mustafa Suleyman, Microsoft AI CEO, emphasized that owning the hardware is essential for the company's competitive goals against rivals like Google DeepMind and OpenAI. "We have to work across the full stack, all the way down to the silicon," Suleyman said, highlighting that co-designing hardware alongside AI models is now a critical focus for the team.

"Maia 300 is a complete AI system designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks." — Scott Guthrie