Microsoft unveils Maia 300, third-generation AI chip built to displace Nvidia in the cloud

Microsoft announced Maia 300 today, the third generation of its homegrown AI processor, and designated it the "primary inference system" for its products and services. The new chip not only promises improved cost and efficiency for running AI workloads in the cloud but was also architected from the ground up to handle current reasoning models and extended agentic tasks. Production deployment is slated for 2027, when the company will begin rolling the silicon into real production workloads.

Scott Guthrie, executive vice president of the Cloud group, said Maia 300 is expected to deliver 40% more tokens per megawatt compared with leading GPU-based systems. In a full configuration of 96 accelerators per rack, throughput increases sevenfold over Maia 200. Those gains are intended to let Microsoft rely primarily on its own silicon instead of chips from Nvidia and AMD, with a plan to reach multi-gigawatt deployments as early as 2027.

The efficiency leap comes from a new chiplet architecture built on TSMC's N2P process. The design separates compute units, memory controllers, and I/O functions, a move that expands the silicon area dedicated to compute performance and lets each function use the process technology best suited to it. Azure's hardware team calls that approach the key to the critical efficiency gains.

The previous generation, Maia 200, was announced only in January this year and is already running production workloads across Microsoft, including Copilot, GitHub Copilot, Foundry, and the company's own models. Guthrie noted that Maia 200 is the highest-performing AI infrastructure in Microsoft's fleet when measured against real-world AI service metrics. Maia 300 was designed as a complete AI system for the new workload types: long-context reasoning, latency-sensitive conversational experiences, and multi-step agentic tasks.

Mustafa Suleyman, chief executive of Microsoft AI, tied the effort directly to the ambition of closing the gap with Google DeepMind, OpenAI, and Anthropic. At the required scale, he said, the company must own a full-stack platform co-developed with its models. Running Microsoft's AI models on Maia silicon is defined as a critical focus for his team, with the goal of controlling every layer down to the silicon itself.