Intel plans Crescent Island AI GPU launch

According to the Financial Times, Intel plans to begin shipping its new "Crescent Island" AI graphics processing unit in limited quantities by the end of 2026, following an 18‑month development effort, Kevork Kechichian, head of Intel's data centre group, told the FT. The FT reports Crescent Island is an air‑cooled design that uses LPDDR5 memory rather than the high‑bandwidth memory (HBM) used in many training GPUs, and that the chip is optimized for inference rather than model training. The FT also reports that an earlier Intel training GPU effort, "Gaudi," had weak sales and that a planned successor was cancelled. The FT frames Crescent Island as a first major AI infrastructure product under chief executive Lip‑Bu Tan.

Industry-pattern observations: using LPDDR5 and air cooling is a clear cost-and-deployment tradeoff compared with HBM plus liquid cooling. LPDDR5 offers substantially lower cost per GB and simpler board and thermal design, but it provides lower sustained memory bandwidth than HBM. For many inference workloads, peak memory bandwidth is less critical than for large‑model training, which is why lower‑cost memory can make sense for inference-focused accelerators. Observers evaluating Crescent Island will compare end‑to‑end inference throughput, latency, and power efficiency against incumbents that use HBM and liquid cooling.

Industry-pattern observations: Nvidia and AMD currently dominate training and high‑end inference segments largely because of high memory bandwidth platforms and mature software ecosystems. Public reporting notes Intel's prior training‑focused GPU effort did not gain traction, and Crescent Island represents a materially different product focus centered on inference and lower deployment cost. The FT frames this launch as part of a broader attempt to re-enter AI infrastructure under new leadership; the reporting documents Lip‑Bu Tan's recent appointment and positions Crescent Island as an early product in that effort.

observers should monitor:

• •independent benchmarks for inference throughput and latency versus incumbent HBM-based cards
• •power efficiency and cooling requirements in typical data‑centre racks
• •customer trials and announced partnerships or purchase commitments
• •how software and ecosystem support (compilers, runtimes, frameworks) are provisioned for the new GPU. Public reporting from FT, Seeking Alpha, and The Information provides the initial timeline and design intent; further vendor disclosures or benchmark tests will be required to judge market impact