Microsoft unveils Maia 300 for 2027 AI deployments

Azure Maia 200 is now in production, and Azure Maia 300 is due to begin production deployments in 2027.

Maia 200 is already serving Microsoft production workloads, including Microsoft Copilot. It is also supporting the company's MAI team as it develops next-generation models, with GitHub Copilot and Microsoft Foundry set to follow by the end of the year. Across those workloads, Maia 200 runs OpenAI models, Microsoft's own MAI models and open-source model families.

The announcement outlines Microsoft's plan to expand the Maia line as part of its AI infrastructure strategy. The company intends to scale the Azure Maia family to multi-gigawatt deployments from 2027 and make Maia its main inference system for Microsoft products and services.

Current rollout

According to Microsoft, Maia 200 is being deployed across the United States, Europe and Australia. The system delivers up to 1.7 times more tokens per watt for built-in AI features in Microsoft Teams, Copilot and Cowork experiences than leading GPU infrastructure.

Lessons from running AI models and services on Maia 200 shaped the design of Maia 300, Microsoft said. The two generations are 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.

Next generation

Maia 300 is the third generation of the Azure Maia programme and is designed with what Microsoft described as an inference-first architecture. It is aimed at long-context, memory-intensive and decode-heavy AI workloads, including long-context reasoning, conversational services that require low latency, and sustained multi-step agent tasks.

According to Microsoft, Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, it delivers up to seven times the rack-scale token throughput of Maia 200.

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

The hardware uses eight HBM4 stacks and provides 18.0 TB/s of peak memory bandwidth within a 950-watt thermal design power envelope, according to Microsoft. It also uses specialised kernel support with on-die SRAM as a software-managed scratchpad to reduce pressure on high-bandwidth memory.

Maia 300 adopts a multi-chiplet design that separates compute, memory-controller and input-output functions, Microsoft said. Compute chiplets are built on TSMC's N2P process node.

Rack design

At rack level, Maia 300 systems combine 96 accelerators within a 167-kilowatt power envelope, according to Microsoft. That setup provides 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth across a high-speed scale-up domain.

Networking is also part of the system design. 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 said.

Scott Guthrie, Executive Vice President of Cloud and AI at Microsoft, said the company sees Azure Maia as a central part of its wider AI platform strategy. "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.