Microsoft plans to expand the Azure Maia systems family to operate in multi-gigawatt deployments starting in 2027, transforming these systems into the core inference platform for its products and services. The company confirmed that Azure Maia 200 has entered production and is currently running live workloads, including Microsoft Copilot.
Maia 200 uses OpenAI models, Microsoft’s MAI models, and families of open-source models. The system also supports the MAI team in developing subsequent models, with expansion to GitHub Copilot and Microsoft Foundry planned before the end of 2026. Current Maia 200 deployments are spread across the United States, Europe, and Australia.
What is Microsoft launching?
The company presented Maia 300 as an integrated third-generation intelligent system designed for long-context inference workloads, latency-sensitive conversational experiences, and multistep agentic tasks. According to Microsoft’s figures, the system targets delivering 40% more tokens per megawatt than systems based on leading GPU units, while a full configuration containing 96 accelerators can achieve up to seven times the rack-level token rate compared with Maia 200.
Production deployments of Maia 300 are scheduled to begin in 2027. Microsoft did not provide a more detailed timeline or commercial pricing for the system in the material.
A design focused on inference and energy efficiency
Maia 300 is based on an inference-first architecture that balances compute capabilities, memory bandwidth, networking, and power consumption. Microsoft says its software-defined dataflow architecture, together with AI-assisted core optimization, aims to reduce data movement and increase utilization of computing, memory, and networking resources.
The system includes eight HBM4 units providing peak bandwidth of 18.0 terabytes per second within a 950-watt thermal design limit. It also uses a multi-chip architecture that separates compute functions from memory-control and input/output units, with the compute chips manufactured using TSMC N2P technology. Low-precision formats, including enhanced MXFP4, support reducing memory and bandwidth requirements for models and KV-cache memory.
What changes in practice?
The Maia 300 rack system combines approximately 96 accelerators within a 167-kilowatt power envelope, with 27.6 terabytes of HBM capacity and total bandwidth of 1.7 petabytes per second. Microsoft is counting on this density to run more simultaneous AI interactions and larger workloads per rack, which could increase data centers’ ability to scale.
The company also designed the scale-out network so that dataflow operations extend across compute, memory, and networking as a single system, with high-bandwidth links between accelerators and dedicated AI transport based on an Ethernet architecture. This architecture targets scaling from a single rack to the data-center level.
Why does this matter?
The move shows that Microsoft does not view inference merely as a matter of purchasing separate accelerators, but as a system encompassing silicon, memory, networking, software, and live service workloads. The company says that running Maia 200 inside Teams, Copilot, and Cowork demonstrated up to 1.7 times as many tokens per watt compared with a leading GPU architecture, and that this experience directly influenced the design of Maia 300.
According to the material, the new systems will be used to run Microsoft services, GitHub Copilot, Microsoft Foundry, MAI models, and Dragon Copilot. However, the performance and efficiency figures presented here are estimates and results announced by Microsoft, while the material does not include independent testing or complete comparative details against competing systems; therefore, practical verifiability remains tied to actual testing after deployments begin in 2027.