Microsoft unveils Maia 300 AI inference chip

Microsoft announced its third-generation Maia 300 AI accelerator, which will become the primary inference system across its products and services. Production starts in 2027 with over 1 million units planned that year; Microsoft claims 40% more tokens per megawatt than leading GPU-based systems.

• Eight HBM4 stacks deliver 18 TB/s of memory bandwidth
• A rack of 96 accelerators consumes about 167 kW
• Token throughput up to 7x higher than Maia 200
• Compute chiplets use TSMC's 2nm-class N2P process