Microsoft today announced its third-generation in-house Maia AI chip, designed to be the “primary AI inference system” for Microsoft products and services. Maia 300 not only improves the cost and efficiency of running AI workloads in the cloud, but it is also more capable of handling the latest reasoning models and agentic tasks.
The Maia 300 will begin production in 2027 and has been designed with efficiency in mind. “We expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems,” says Scott Guthrie, executive vice president of Microsoft’s cloud group. “And in a full 96-accelerator configuration, it delivers up to 7x the rack-scale token throughput of the Maia 200.”
These efficiency improvements will soon allow Microsoft to rely primarily on its in-house AI chips, instead of those from Nvidia and AMD. “We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the leading AI inference system for Microsoft products and services,” says Guthrie.
The Maia 300 is the successor to the 200, only announced in January. Maia 200 already runs on production AI workloads at Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft’s own AI models.
“Maia 200 is the highest performing AI infrastructure in our fleet, measured in real-world metrics for AI services,” says Guthrie. “Maia 300 is a comprehensive AI system designed for the new types of AI workloads emerging today: long-context reasoning, latency-aware conversational experiences, and sustained multi-step agentic tasks. »
Microsoft is using a new chipset architecture to achieve its very significant efficiency gains. Built on TSMC’s N2P process node, the latest Maia 300 architecture separates compute, memory controller, and I/O functions. “This increases the silicon surface area dedicated to compute performance and allows each function to use the technology best suited to its role,” explains the Azure hardware and infrastructure team.
Maia 300 is also a key part of Microsoft’s AI model work and could help the company achieve its goal of catching up with Google DeepMind, OpenAI and Anthropic. “We need to work across the entire stack, right down to the silicon,” says Mustafa Suleyman, CEO of Microsoft AI. “At this scale, we need a platform that we own and co-design with our models. That’s why running our Microsoft AI models on the Maia chip is a critical goal for our team.”
Gn bussni
