Microsoft plans to scale its Azure Maia AI infrastructure to multi-gigawatt deployments beginning in 2027, positioning the Maia family as the company’s primary inference system across Microsoft products and services.
The expansion represents a major step in Microsoft’s strategy to build vertically optimized AI infrastructure spanning custom silicon, systems, networking, software, and AI models.
Azure Maia 200 is already running production workloads, including Microsoft Copilot, and is supporting Microsoft’s MAI team as it develops next-generation models.
Microsoft plans to extend Maia 200 to GitHub Copilot and Microsoft Foundry by the end of 2026.
Across these workloads, Maia 200 supports OpenAI models, Microsoft’s own MAI models, and leading open-source model families.
Microsoft said its ability to operate AI services at enormous scale gives its hardware and software teams data that can be used to jointly optimize silicon, systems, and applications for real-world enterprise workloads.
The next generation, Maia 300, is scheduled to begin production deployments in 2027.
Microsoft designed Maia 300 specifically for emerging AI workloads including long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic tasks.
The company expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.
In a full 96-accelerator configuration, Maia 300 is also expected to provide up to seven times the rack-scale token throughput of Maia 200.
Maia 200 is currently being deployed across the U.S., Europe, and Australia.
Microsoft said the system is delivering up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot, and Cowork compared with leading GPU infrastructure.
Experience running production AI models and services on Maia 200 directly influenced the development of Maia 300.
Microsoft expects the combined Maia generations to support a broader range of workloads, including multi-step agentic coding experiences for GitHub, cloud-scale model serving through Microsoft Foundry, inference and data generation for the MAI organization, and latency-sensitive Dragon Copilot healthcare workflows.
Maia 300 uses an inference-first architecture optimized for long-context, memory-intensive, and decode-heavy AI workloads.
The platform uses a software-defined dataflow architecture and AI-assisted kernel optimization to reduce data movement and improve utilization across compute, memory, and networking resources.
Microsoft said the system delivers industry-leading high-bandwidth memory performance per watt, with eight HBM4 stacks providing 18 TB per second of peak memory bandwidth within a 950-watt thermal design power envelope.
The Maia 300 architecture separates compute, memory-controller, and I/O functionality across multiple chiplets.
Its compute chiplets are manufactured using TSMC’s N2P process technology, which Microsoft said helps improve performance per watt, scalability, and overall cost efficiency.
Maia 300 also supports advanced low-precision formats, including enhanced MXFP4, reducing memory and bandwidth requirements while enabling larger context windows and increased model throughput.
Microsoft designed the networking architecture to scale from individual racks to data center-scale deployments.
High-bandwidth interconnects connect Maia accelerators for tensor-parallel workloads, while Microsoft’s custom AI transport uses Ethernet-based networking infrastructure to scale communication across accelerators and racks.
A Maia 300 rack combines 96 accelerators within a 167-kilowatt power envelope, providing 27.6 TB of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth.
Microsoft said this density will allow its data centers to handle more simultaneous AI interactions and larger workloads from each rack, improving both capacity and the economics of scaling AI services.
The Maia program is becoming a foundational component of Microsoft’s broader Frontier Transformation strategy, which emphasizes an open and heterogeneous infrastructure capable of supporting multiple models and silicon architectures.
Microsoft plans to disclose additional details about its AI infrastructure strategy, technology adoption, and Frontier Transformation initiatives at Microsoft Ignite 2026.
