Microsoft has formally introduced its third-generation AI acceleration platform, Azure Maia 300. This cutting-edge system is scheduled to commence production in 2027, forming a pivotal element of Microsoft's extensive multi-gigawatt AI infrastructure initiative. Notably, the Azure Maia 300 boasts an inference-centric architecture, meticulously crafted to handle emerging workloads like long-context reasoning. It comes equipped with eight stacks of HBM4 memory and leverages a multi-chip design built on TSMC's advanced N2P process technology.

When pitted against conventional GPU systems, the Azure Maia 300 demonstrates a remarkable 40% surge in token output per megawatt. Moreover, the token throughput of a fully-loaded 96-accelerator rack soars to seven times that of its predecessor, the Maia 200. The system is further enhanced with a suite of optimized designs, compatibility with low-precision formats, and the flexibility to scale from individual racks to entire data center deployments.