Intel: Our upcoming AI chips are cheaper than NVIDIA and AMD, with air cooling. Crescent Island focuses on inference.

Intel plans to begin small-volume shipments of the new inference accelerator “Crescent Island” to customers by the end of 2026. It will use LPDDR5X memory and an air-cooling design, supporting up to 480GB of memory capacity, with a TDP of 350W. Positioned to directly challenge Nvidia and Supermicro (AMD) solutions that focus on HBM combined with liquid cooling.
(Backgrounder: Challenging Nvidia and AMD! Intel unveils a new affordable AI chip “Crescent Island,” emphasizing air cooling and low cost)
(Additional background: Nvidia partners with China-based startup Unitree! Huang Renxun promotes a “humanoid robot AI platform,” targeting the tens of billions of dollars physical AI market)

Table of Contents

Toggle

• From the failed Gaudi to air-cooled Crescent Island
• LPDDR5X vs HBM: Could affordability be a moat?
• Chen Liwu’s first move, and Intel’s situation

The last time Intel left a mark in the AI chip market was a processor called Gaudi. It sold poorly, and its successor was canceled last year. This time, Intel has chosen to return to the fight from a different angle—going after inference only, not training.

From the failed Gaudi to air-cooled Crescent Island

Kevork Kechichian, head of Intel’s data center division, told the Financial Times that the company is now “starting from the basics.” Kechichian joined Intel last year from chip design firm Arm, and he said:

“We decided to rebuild our AI muscles…… but based on past experience, we are not specifically targeting the training market.”

That line is worth reading carefully. Gaudi’s failure back then was partly because it tried to take on Nvidia head-on in the training market. By then, that market’s ecosystem, software toolchain, and developer habits were already deeply tied to CUDA.

This time, Intel is choosing inference. In plain terms, it’s the stage where users submit a question and the model replies immediately, rather than the training stage where the model learns from massive amounts of data.

The two have completely different positions in the market. Training is dominated by Nvidia’s H100, H200, and Blackwell series, with extremely high entry barriers. Inference, however, is exploding in a more dispersed way thanks to the rapid adoption of enterprise-side AI applications—and sensitivity to cost is far higher than in academic research scenarios.

The new chip “Crescent Island” is based on Intel’s Xe3P architecture and is offered in the form of a PCIe add-in card: TDP 350W, air-cooled, supporting up to 480GB of LPDDR5X memory.

It has been in development for 18 months. It is expected to begin sampling to customers in the second half of 2026, and to enter small-volume shipments by the end of the year.

LPDDR5X vs HBM: Could affordability be a moat?

Intel’s most core differentiation this time lies in its choice of memory.

For Nvidia’s Blackwell and Supermicro’s flagship inference cards, the memory used is HBM (high bandwidth memory). It offers extremely high bandwidth, but at a very high cost—cost per GB is far higher than that of standard memory.

Crescent Island instead uses LPDDR5X, which is low-power memory originally used in high-end smartphones and laptops. With large-scale production and much lower pricing, it is significantly cheaper.

At the same time, this choice also changes the cooling solution. The combination of high-compute processors with HBM often pushes the card’s overall TDP to 700W or even higher, requiring liquid cooling pipelines to conduct heat directly away from the chips. The cost of infrastructure upgrades can easily run into millions of dollars.

With its 350W TDP, Crescent Island falls within the range acceptable for standard air-cooled server racks.

Air cooling versus liquid cooling is not only a difference in heat dissipation methods—it also reflects a clash between two data center construction philosophies. For many mid-sized enterprises, bringing in liquid cooling means having to do major surgery on their existing data halls. Crescent Island’s selling point is that AI inference can run in existing facilities without needing to tear down and rebuild the data center.

This logic directly targets total cost of ownership (TCO). In plain terms: buying the equipment is just the first expense; afterwards, the costs for electricity, cooling, and maintenance are where the money is burned. Intel is trying to build an advantage on this dimension.

From a pure specification standpoint, 480GB of LPDDR5X is indeed compelling for large-model inference: a larger memory capacity allows more model parameters to remain resident in memory, reducing read latency, which is especially critical for long-text inference and multi-task parallel scenarios. But whether its bandwidth can keep up with HBM remains the key question for this approach.

Chen Liwu’s first move, and Intel’s situation

This marks the first time that, after CEO Chen Liwu officially took over, the company has clearly moved to target the AI infrastructure profit market. His predecessor, Gelsinger, was forced to step down last year after widespread doubts about the company’s transformation strategy, leaving Intel in a situation where it was losing its foundry advantage and was nearly absent from the AI era.

Chen Liwu’s strategic direction, at least for now, appears to be: not trying to win every battlefield at once, but instead finding a corner where Nvidia and Supermicro are least able to pay attention—mid- to low-end inference deployment scenarios—using cheaper memory and lower infrastructure barriers to attract mid-sized enterprise customers who are kept out by HBM pricing and the demand for liquid cooling.

Kechichian’s remarks emphasize “rebuilding the foundation,” not “disrupting the market.” This posture may be the most honest assessment Intel can make at the moment: Crescent Island is not a Nvidia killer; it is a product trying to find a foothold in the inference boom.

Whether it can succeed depends on the real-world performance numbers after the samples are sent in the second half of 2026, when enterprise customers actually get the hardware and start running it.