TLDR Tech / Launches
Microsoft's Maia 300 AI chip to power multi-gigawatt data centers
Microsoft has unveiled its third-generation inference accelerator, the Maia 300, designed for workloads requiring long contexts and intensive reasoning. The company plans to deploy these chips extensively, starting in 2027, as its primary system for inference, with multi-gigawatt deployments anticipated. This decode-first chip is equipped with 288 GB of HBM4 memory, offering 18 TB/s of bandwidth, and features a software-defined dataflow architecture capable of scaling to fabrics of up to 12,288 accelerators. Executives at Microsoft state that the current Maia 200 chip already outperforms leading GPUs in performance-per-watt for production workloads.
Read the full story