Intel has introduced "Crescent Island" a new data-center GPU that is specifically created to perform AI inference tasks.
The company unveiled the chip in the 2025 Open Compute Project Global Summit and planned to start customer sampling in the second half of 2026.
The launch is a strategic rebranding of the accelerator business of Intel, which has not been successful with the Gaudi AI chips and the cancellation of the Falcon Shores product.
With the new CEO Lip-Bu Tan coming to the helm in March 2025, the company has now embarked on an aggressive and open approach to be able to compete with other market players such as Nvidia and AMD.
Intel has set Crescent Island as an inference-optimized graphics card, that is, it is designed to execute AI models, when they are trained, but not the training process itself.
The chip is based on the Intel Xe3P microarchitecture, and has 160GB of LPDDR5X memory and is intended to be mounted in air-cooled enterprise servers.
The major focus is providing a high performance-per-dollar for large-scale inference workloads, which is a critical and expanding segment of the AI market.
The main feature of the new strategy of Intel is the promise of new data-center AI chips annually and a modular, open architecture.
This would enable the customers to intermix and interchange parts of various vendors, which would be a direct answer to the more closed eco-system of its competitors.
Although Intel has not yet released the performance numbers and the process node of Crescent Island, the announcement was a clear favorite that it is entering the high-stakes AI accelerator race again.
The company is betting that a focus on cost-effective interference and an open, yearly roadmap will win back customers and establish a foothold in a market it has so far struggled to dominate.