Microsoft has announced Maia 300, the third generation of its custom AI chip. The company says it will serve as the main inference system behind its products and services. The chip aims to lower the cost and energy needed to run AI in the cloud, and it is better suited to the latest reasoning models and agent-style workloads. Production rollout begins in 2027.
Efficiency and Performance Gains
Scott Guthrie, executive vice president of Microsoft’s cloud group, said Maia 300 should deliver about 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator setup, rack-scale token throughput is expected to reach up to seven times that of Maia 200.
Microsoft plans to scale the Azure Maia family to multi-gigawatt deployments starting in 2027, which would let it lean mainly on its own silicon rather than chips from Nvidia and AMD.
Maia 200, announced only in January, already runs production workloads across Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company’s own AI models. Guthrie called it the best-performing AI infrastructure in Microsoft’s fleet based on real-world service metrics. He described Maia 300 as a complete system built for emerging workloads such as long-context reasoning, low-latency conversational experiences, and sustained multi-step agentic tasks.
A Chiplet Design and Microsoft’s Model Ambitions
To achieve these gains, Maia 300 uses a chiplet architecture built on TSMC’s N2P process. Compute, memory controller, and I/O functions are separated, which frees up more silicon area for compute and lets each function use the technology that suits it best, according to Microsoft’s Azure hardware and infrastructure team.
The chip also supports Microsoft’s push to catch up with Google DeepMind, OpenAI, and Anthropic in AI models. Microsoft AI CEO Mustafa Suleyman said that at this scale the company needs a platform it owns and can co-design with its models, so running Microsoft AI models on Maia is a key priority.
