Microsoft has introduced the Maia 300, its new inference chip named "Clea." The company says the processor was designed with a focus on bandwidth—not just raw power—to meet the demands of OpenAI models and the AI features integrated into its productivity applications.
The chip is aimed both at inference for models like ChatGPT and at training Microsoft's own AIs. The company claims the Maia 300's architecture includes changes that allow inference to run faster and scale more efficiently, without relying solely on increasing the number of processing units.
The decision to prioritize bandwidth reflects a known bottleneck: most large language models are limited by memory access speed, not computation. The chip addresses this directly, with a structure that keeps data flowing without interruptions.
For AI infrastructure operators, the arrival of a specialized chip from Microsoft means another option outside the Nvidia ecosystem—and a sign that the company plans to scale its services without relying exclusively on external suppliers. The paper evaluates that bandwidth as a design priority is a technical choice that could influence the entire next generation of AI accelerators.