Microsoft Unveils Maia 300 AI Chip for 2027 Deployments

Microsoft has announced Azure Maia 300, a third-generation AI inference chip designed to power the company's AI services with improved efficiency when it enters production deployment in 2027. The announcement signals Microsoft's strategic expansion of purpose-built silicon for AI workloads across its product portfolio.

Performance Gains for AI Inference

The Maia 300 chip delivers 40% more tokens per megawatt compared to GPU-based systems and achieves seven times the rack-scale token throughput of the current Maia 200 generation, according to Microsoft. The architecture employs eight HBM4 memory stacks with 18.0 TB/s peak memory bandwidth and operates at a 950-watt thermal design power.

Each rack configuration houses 96 accelerators within a 167-kilowatt power envelope, providing 27.6 TB HBM capacity and 1.7 PB/s aggregate HBM bandwidth. The compute chiplets are manufactured on TSMC's N2P process node using a multi-chiplet design that separates compute, memory-controller, and I/O functions.

Scott Guthrie, Executive Vice President of Cloud and AI at Microsoft, positioned the chip as an inference-first architecture optimized for long-context reasoning, low-latency conversational services, and sustained multi-step agent tasks. "Microsoft is unique because we are able to optimize from silicon to systems to agent on an open, heterogenous and model diverse platform our customers can count on for their AI transformation, today and in the years ahead," Guthrie said.

Current Generation Already in Production

The Maia 200 chip is currently in production serving Microsoft Copilot and AI model development across data centers in the United States, Europe, and Australia. Microsoft reports that Maia 200 delivers 1.7x more tokens per watt than leading GPU infrastructure for Microsoft Teams, Copilot, and other AI features.

Deployments for GitHub Copilot and Microsoft Foundry are planned by the end of 2026, establishing the infrastructure foundation that Maia 300 will expand upon. The Maia 200 runs OpenAI models, Microsoft's own MAI models, and open-source model families across these workloads.

Architectural Design and Deployment Strategy

Microsoft said Maia 300 uses a software-defined dataflow architecture intended to reduce data movement and improve utilization of compute, memory, and networking resources. The design incorporates AI-assisted kernel optimization and treats dataflow as a central design principle across the system.

The hardware uses specialized kernel support with on-die SRAM as a software-managed scratchpad to reduce pressure on high-bandwidth memory. High-bandwidth interconnects between Maia accelerators are tuned for tensor parallelism, while a custom transport layer uses Ethernet-based networking infrastructure to link accelerators and racks in a unified topology.

Microsoft intends Maia 300 to serve as the primary inference system for its products and services, with multi-gigawatt deployments planned from 2027. The company stated that lessons from running AI models and services on Maia 200 shaped the design of Maia 300, with both generations intended to support a wider range of AI services including multi-step agent systems for GitHub coding tools, cloud-scale model serving for Microsoft Foundry, inference and data generation for the MAI team, and latency-sensitive interactions for Dragon Copilot.

The Maia 300 is designed for long-context, memory-intensive, and decode-heavy AI workloads. Organizations building on Azure infrastructure may see operational changes as Microsoft transitions from GPU to Maia-based inference infrastructure during the 2027 rollout.

Sources

1. Microsoft unveils Maia 300 for 2027 AI deployments — IT Brief New Zealand