- Microsoft will use its Maia chips as the main AI inference system from 2027 onward
- Maia 200 chips are currently deployed powering Microsoft Copilot and OpenAI models
- Maia 300, launching in 2027, offers 40% more tokens per megawatt than top GPU systems
Microsoft said on Monday that its in-house Maia chips will become the primary inference system across its products and services, with deployments scaling to multiple gigawatts from 2027. The move could reduce the company's reliance on GPU makers such as Nvidia.
“As we move into the era of Super Intelligence, Microsoft is entering a new phase of infrastructure,” Microsoft executive vice president Scott Guthrie wrote in a blog post. “We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.”
Maia 200 Already in Use
The current-generation Maia 200 is already live in Microsoft's production systems. It powers Microsoft Copilot and is being used by the company's MAI team as it works on its next AI models. GitHub Copilot and Microsoft Foundry will move onto it by the end of the year. Models from OpenAI run on the chip, as do Microsoft's in-house MAI models and well-known open-source models.
ALSO READ | Microsoft CEO Satya Nadella Calls For 'Emergency Brake' On Advanced AI
Microsoft is rolling out Maia 200 across the US, Europe and Australia. For built-in AI features in Teams, Copilot and Cowork, the company says it generates up to 1.7 times more tokens per watt than leading GPU infrastructure.
Maia 300 Coming in 2027
The third-generation Maia 300 will enter production deployments in 2027. Microsoft says it is built primarily for inference and designed for newer kinds of AI work, such as AI agents handling multi-step jobs, reasoning over very long inputs and conversations where response speed matters. The company expects it to produce 40 per cent more tokens per megawatt than leading GPU-based systems.
Each chip carries eight HBM4 memory stacks, delivering 18 TB/s of peak memory bandwidth at a 950-watt TDP. A new chiplet design splits compute, memory control and input/output into separate parts, with the compute chiplets made on TSMC's N2P process.
ALSO READ | 'AI Models Must Be Governed Like Insider Threats': Satya Nadella Outlines Seven Principles To Contain Risk
Rack-Scale Gains
Microsoft packs 96 Maia 300 accelerators into a single rack that draws up to 167 kilowatts. Together, they offer 27.6 TB of HBM and 1.7 PB/s of total memory bandwidth. In this setup, Microsoft says it delivers up to seven times the token throughput of a Maia 200 rack.
Reducing data movement inside the system was a key design goal. “Moving data efficiently, rather than raw compute power, is increasingly the bottleneck in large inference systems,” Guthrie wrote.
ALSO READ | Call For Kill Switch: Why Microsoft's Satya Nadella Backs Human-Controlled 'Emergency Brake' On AI
Not a Single-Chip Bet
Guthrie added that Microsoft sees choice in models and hardware as central to its strategy. “The future will not be defined by a single model or a single piece of silicon or a single innovation,” he wrote.
Microsoft plans to share more about its AI infrastructure strategy at Ignite 2026, scheduled for next month in San Francisco and online.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.