Intel unveiled “Crescent Island” a new data center GPU specifically designed to handle the world’s transition from AI training to real time AI inference. As agentic AI expands, the demand for processing mass volumes of data tokens efficiently will soar.

The chip will use the Xe3P microarchitecture, which focuses on maximizing performance per watt to keep energy consumption low. It features a substantial 160GB of LPDDR5X memory, providing the high bandwidth necessary for complex inference tasks. The chip is optimized for standard air cooled servers, making it easier and cheaper for companies to integrate their chips with their existing tech stack, without specialized cooling.

Intel’s strategy focuses on heterogenous computing, the idea that different AI tasks require different types of silicon. By combining these GPUs with their Xeon 6 processors and an open source software stack, Intel plans to make AI deployment more scalable and developer friendly.

Developers are already able to test and experiment software for this ecosystem, using Intel’s Arc Pro B-Series GPUs. Customer sampling of the new data center is expected to begin in the latter half of 2026.