Intel Bets on Heterogeneous Infrastructure for Agentic AI

According to AI TIMES, Intel said on November 12 that it is difficult to operate data centers with GPUs alone as agentic AI spreads, and unveiled a strategy that combines CPUs, GPUs, and dedicated accelerators. The inference GPU it announced that day targets demand for handling multiple models at once or processing long context, with a 350-watt design and up to 480GB of memory scalability. Intel said that because agents do not stop at generating answers but repeatedly perform search, tool execution, and result validation, it has become important to divide resources according to the nature of each task. In such an environment, Intel said the key is a structure in which Xeon CPUs handle coordination tasks such as request classification, resource allocation, and execution environment management, while GPUs and accelerators support them. It also emphasized that the efficiency of the entire system should be considered rather than boosting the performance of a specific device. Ultimately, the announcement shows that the focus of competition in AI infrastructure is shifting from securing training GPUs to how to balance the cost and efficiency of actual inference operations.