Intel Plans AI Inference Chips Shipments This Year, Betting on Low-Cost Memory and Air Cooling Solutions
N.R. Finch
Intel plans to ship Crescent Island, an AI inference chip using air cooling and cheaper memory, by year-end — the first AI infrastructure product under new CEO Lip-Bu Tan.
What is this chip actually trying to do?
Crescent Island targets inference — the stage where a model generates an answer after a user's prompt — not training.
This means → Intel is sidestepping Nvidia's strongest territory and picking a lane the rival guards less tightly.
Intel data-center head Kevork Kechichian told the Financial Times: the company is not rushing to challenge Nvidia head-on in training, but wants to rebuild its AI chip capability first.
What makes it "low cost"?
Two key design choices: air cooling instead of liquid cooling, and LPDDR5 memory instead of HBM — high-bandwidth memory used in Nvidia's Blackwell chips.
In plain terms = HBM is expensive and demands liquid-cooled data centers. Crescent Island skips both, pulling the cost floor significantly lower.
Intel also plans to eventually move production back to its own fabs, further narrowing the cost gap with competitors that rely on TSMC.
Why did Intel end up here?
Its previous AI training GPU, Gaudi, sold poorly; the planned successor was cancelled last year.
This reflects a hard lesson: going toe-to-toe with Nvidia on training didn't work, so Intel is trying a different entry point.
Kechichian's own words: "We decided to start rebuilding our capability in AI, but given our past experience, we're not specifically targeting training."
What has changed since the new CEO took over?
Lip-Bu Tan replaced Pat Gelsinger as CEO last year, then cut costs and tightened spending on manufacturing projects.
Intel's stock has risen more than 200% year-to-date.
Crescent Island is Tan's first AI infrastructure product to reach the market — this means → the chip is both a product and a scorecard for the new management team.
Is there a China angle?
Kechichian said Intel is evaluating whether a version of Crescent Island can be sold in China within U.S. export-control rules.
His words: "Some of those tiers of the chip may be appropriate there… that particular market has demand at this particular price point."
Put simply = a low-cost inference chip naturally fits price-sensitive markets, and China is an obvious potential buyer — but whether it can be sold depends on where export controls draw the line.
We decided to start rebuilding our capability in AI, but given our past experience, we're not specifically targeting training.
Kevork Kechichian
Intel Head of Data Center
(Interview with the Financial Times)
Content is for reference only, not financial advice.