Overview
- On Monday Intel introduced Crescent Island, a data-center GPU built for inference workloads rather than model training and planned for limited customer shipments by the end of 2026.
- The chip uses lower-cost LPDDR5/LPDDR5X memory and an air-cooled 350W PCIe card reference to avoid expensive HBM memory and liquid-cooling infrastructure used by top-end rivals.
- Intel plans to produce Crescent Island in its own fabs as part of CEO Lip-Bu Tan’s turnaround and is evaluating China-compliant variants that would meet U.S. export controls.
- NVIDIA’s same-day push into PC AI with the RTX Spark Superchip heightened competitive pressure and helped drive Intel shares down roughly 4–5% in premarket trading.
- The move follows Intel’s Gaudi training missteps and recent stronger quarterly results, and it signals a focused, cost-first strategy that could reshape where customers place inference capacity and how data centers are provisioned.