Intel’s Crescent Island GPU Targets Cheaper, Air-Cooled Inference

Intel announced Crescent Island, an AI data-center GPU built for inference. The chip supports up to 480GB of LPDDR5X memory and is designed to run in air-cooled server racks. Customer sampling is planned for the second half of 2026.

The design uses lower-cost LPDDR5X instead of the high-bandwidth memory found in many top-tier accelerators. Intel projects that choice will lower upfront hardware costs and reduce energy use. The product targets inference workloads rather than large-scale model training and emphasizes performance per dollar and token economics. The company plans an open, modular approach that allows operators to mix Intel GPUs with hardware from other vendors.

Intel reported more than $18 billion in fresh funding tied to its AI efforts, including $11.1 billion from the U.S. government, $5 billion from Nvidia and $2 billion from SoftBank. The company says the funds will support production scaling and partner integration.

Operators that convert power-rich facilities into AI compute centers are a target market. Some bitcoin miners have shifted capacity into AI workloads because revenue per megawatt for AI can exceed returns from mining. One operator expanded its AI footprint around a 1GW campus after securing $3 billion in backing and winning $12.8 billion in AI contracts.

Air-cooled chips with lower power draw reduce demand on electrical infrastructure and can avoid the need for liquid cooling and costly power upgrades. Intel positions Crescent Island as a lower-cost option for dense, always-on inference capacity that can be deployed without specialized cooling.

Near-term milestones include customer sampling in H2 2026, independent performance and efficiency tests, and early design wins. Nvidia and AMD continue to serve the high-performance segments of the market; the forthcoming tests and design wins will show how widely Crescent Island is adopted for inference workloads.