Microsoft unveils Maia 300 chip for AI inference workloads

what happened

REDMOND (dpa-AFX) — Newsmetry API reports that Microsoft Corp. (MSFT) has introduced Maia 300, the third generation of its Maia AI chip. The company designed it for artificial-intelligence workloads across its products and services and says it can generate 40% more tokens per watt than leading systems built around graphics processors.

The chip was developed around the workloads Microsoft runs in practice. The latest stage of the Azure Maia program is aimed at inference, or using trained models to produce responses and results. Maia 300 is intended for applications involving substantial memory requirements, lengthy inputs and demanding computation.

An advanced low-precision format, MXFP4, reduces the memory bandwidth needed for calculations and for storing key-value cache data. Microsoft says this lets models accept longer inputs, serve more requests and use computing capacity more efficiently without raising power use. A software-defined dataflow design, together with AI-assisted kernel tuning, is intended to improve transfers among compute, memory and networking systems.

Maia 300 uses a chiplet architecture that places computing, memory control and input-output functions in separate components, allowing the design to be adapted to different workloads. MSFT shares traded at $542.08 on Nasdaq, up 1.31%.

why it matters

The launch highlights Microsoft’s push to develop in-house AI hardware and improve efficiency in inference workloads. Investors will focus on the stated performance gains and the 1.31% move in MSFT shares.

instruments