Microsoft Unveils Maia 300 AI Inference Accelerator

Microsoft has unveiled Maia 300, an inference-focused AI accelerator built with four computing chiplets on TSMC's two-nanometer node and 288GB of HBM4 memory. Microsoft claims it delivers 40% more tokens per megawatt than leading GPU-based systems. A dataflow architecture minimizes data movement by using on-chip SRAM caches and compressed data formats like MXFP4. Deployed in 96-accelerator liquid-cooled racks, Maia 300 is expected to deliver seven times the token throughput of its predecessor, Maia 200. Microsoft plans to begin production deployment in 2027, grouping chips into 12,288-accelerator superpods.