At the OCP Global Summit, Microsoft unveiled Maia 300, its third-generation in-house AI accelerator, which it plans to deploy at scale from 2027 as the main inference system for Microsoft products. It is a key step in moving Microsoft’s own chips from a “supporting role” to the “mainstay.”
Key points:
Top-end specs, built for inference. Maia 300 uses a TSMC 2nm chiplet design with 288GB of HBM4 high-bandwidth memory, integrates 96 accelerators per rack, and delivers rack-level token throughput 7 times that of the previous-generation Maia 200.
Energy efficiency is the main selling point. Microsoft says Maia 300 delivers about 40% more effective compute per megawatt than mainstream GPU systems. At a time when inference cost determines the gross margin of AI products, energy efficiency translates directly into competitiveness.
In-house chips become the mainstay. Maia 200 already runs production workloads such as Copilot, while Maia 300 is meant to carry Microsoft’s entire AI product line. Cloud providers’ own chips are moving from “spare tire” to “mainstay,” posing a real challenge to Nvidia’s share of the cloud inference market.
Note that Microsoft has not published independent third-party benchmarks, and the reference system for the energy-efficiency comparison is not named, so actual performance will have to be verified after deployment in 2027.
Source: read the original. This text is a summary for reference only.
