Intel says it will begin limited customer shipments of its new AI inference accelerator, Crescent Island, by late 2026. The chip uses LPDDR5X memory and an air-cooled design, with support for up to 480GB of memory and a 350W TDP. The pitch is direct: challenge Nvidia and AMD in inference deployments with a lower-cost alternative to HBM-based, liquid-cooled systems.
The company’s last notable push in AI chips was Gaudi, which struggled commercially, and its follow-up products were canceled last year. This time Intel is not trying to fight for the training market. It is aiming at inference, the stage where trained models respond to user requests in real time.
After Gaudi, Intel narrows its focus to inference
Kevork Kechichian, head of Intel’s data center division, told the Financial Times that the company is rebuilding “from the fundamentals.” Kechichian, who joined Intel from Arm last year, said the company decided to rebuild its AI capability without specifically targeting training.
That choice reflects the market reality. AI training remains dominated by Nvidia’s H100, H200 and Blackwell products, backed by an entrenched software stack and developer base. Intel is trying to avoid the hardest front of the battle and enter where enterprise demand is spreading more broadly.
Crescent Island is based on the Intel Xe3P architecture and comes as a PCIe add-in card. Intel said development has taken 18 months, with customer sampling set for the second half of 2026 and limited shipments by year-end.
LPDDR5X instead of HBM
The main differentiator is memory. Nvidia’s Blackwell lineup and AMD’s top inference cards rely on HBM, which offers very high bandwidth but comes at a much higher cost. Crescent Island uses LPDDR5X, a memory type more commonly found in premium smartphones and laptops, where production volume is larger and pricing is lower.
That decision also changes the cooling story. HBM paired with high-performance processors often pushes board power to 700W or more, making liquid cooling common in advanced deployments. Crescent Island, at 350W, stays within the range of standard air-cooled racks. For many mid-sized enterprises, that means AI inference hardware could be installed without overhauling an existing data center.
Intel is clearly selling total cost of ownership, not just a card. Hardware is only one part of the bill. Power, cooling and maintenance continue long after purchase. If air cooling can keep deployment inside current facilities, it removes a major infrastructure hurdle for customers that are priced out of HBM systems.
Large memory capacity, but bandwidth remains the open question
On paper, 480GB of LPDDR5X is meaningful for large-model inference. More memory capacity allows more model parameters to remain resident, which can reduce latency and help in long-context inference or multi-workload settings. The obvious question is whether LPDDR5X bandwidth can keep up well enough against HBM-based competitors.
That is also where Intel’s positioning becomes more precise. Crescent Island is not being framed as a direct replacement for Nvidia at the top end. It is aimed at customers that want inference hardware but cannot justify HBM pricing or the added complexity of liquid-cooled infrastructure. Mid-range deployment scenarios appear to be the intended target.
First clear AI infrastructure move under Lip-Bu Tan
This is Intel’s first explicit push into the profit-heavy AI infrastructure market since CEO Lip-Bu Tan took charge. His predecessor, Pat Gelsinger, exited after the company’s turnaround strategy came under pressure, and Intel has had little presence in the current AI wave. Tan’s opening move is not a broad attack across every segment. It is a narrower attempt to secure a position in inference first.
Whether that works will depend on customer testing once samples arrive in the second half of 2026. Intel has laid out the specifications and the cost argument. The next step is whether real-world performance can turn that pitch into actual demand.