Intel's new AI chip skips the costly memory Nvidia relies on

Scarcity has a way of making people forget there was ever another option.

When one ingredient becomes the thing everyone needs at once, the price stops behaving like a price and starts behaving like a ransom. We have watched this happen with oil, with shipping containers, with baby formula, and lately with a type of computer memory most people have never heard of and will never see.

For the past two years, the story of artificial intelligence has been told almost entirely in superlatives, with each new model and data center bigger than the last.

The companies selling the picks and shovels of the AI gold rush have been rewarded with some of the largest market caps in history, and the assumption baked into it all was that the recipe never changes. You needed a specific kind of chip, wrapped in a specific kind of expensive memory, and you paid whatever the suppliers asked.

That assumption took a public hit on June 1.

At Computex 2026 in Taipei, Intel (INTC) laid out fresh details on a data center chip, code-named Crescent Island, that quietly walks away from the most expensive part of the standard AI hardware recipe.

Why the AI memory shortage is squeezing everyone

Start with the part that touches your wallet.

The reason your next laptop, phone, or graphics card may cost more has little to do with the screen or the processor and almost everything to do with the memory chips inside.

More AI:

• Micron sits at the center of a red-hot chip rally
• IBM CEO sends blunt message on AI and quantum computing
• Anthropic CEO makes shocking admission about AI

Modern AI accelerators are surrounded by high-bandwidth memory (HBM), a fast, dense, expensive type of chip that sits right next to the processor. Nvidia (NVDA) builds its most powerful systems around stacks of the stuff, and so does most of the industry.

HBM is the scarce ingredient in the AI recipe, and right now there is not enough of it to go around.

Memory makers Samsung, SK Hynix and Micron (MU) have spent two years shifting production toward HBM and other AI-grade memory, because that is where the fat margins are. That choice left far less capacity for the ordinary memory that goes into consumer gadgets.

The squeeze shows up in a few hard numbers:

• Data centers are on track to consume as much as 70% of the memory chips made worldwide in 2026, Everstream Analytics noted.
• Memory now makes up about 20% of the hardware cost of a laptop, up from 10% to 18% in the first half of 2025, according to CNBC.
• Prices for DRAM, the everyday memory in phones and PCs, surged in early 2026 as AI demand outran supply, IDC shared.

TrendForce analyst Tom Hsu described the run-up in memory prices as "unprecedented," according to CNBC. When the people who track this market for a living reach for that word, it is worth paying attention.

Kajdi Szabolcs / Getty Images

Inside Intel's Crescent Island bet on cheaper memory

This is where Intel saw an opening. At Computex 2026, the company laid out new details on Crescent Island, a data center graphics processing unit (GPU) built for AI inference, the everyday work of running models rather than training them. The headline choice is what it leaves out.

Instead of HBM or graphics memory (GDDR), Crescent Island uses LPDDR5X, a low-power memory more common in phones and laptops than in high-end accelerators. Intel's reference design carries 160GB of it, and the company says partners can build cards with up to 480GB, according to Tom's Hardware.

That is an unusual call, and a deliberate one. LPDDR5X is cheaper, more available, and does not force Intel to fight Nvidia and everyone else for scarce HBM or for limited advanced-packaging capacity.

Related: AMD and Intel lead 2026 gains as AI guard changes

The chip is a PCI Express add-in card with a 350W power target, close to Nvidia's RTX Pro 5000 Blackwell, and it is air-cooled, so it can drop into the standard servers many companies already run. Intel says the design is meant to "efficiently handle large, token-intensive workloads while reducing total cost of ownership," according to Intel.

When I ran Intel's own spec sheet against the going rate for HBM, the trade-off was easy to see. You give up raw memory bandwidth. Crescent Island offers about 684 GB/s, far below what an HBM accelerator delivers. In exchange, you get sheer capacity and a much smaller bill.

For inference work that needs to keep a big model close at hand rather than move data at top speed, that can be a smart swap. Stack eight of these cards in one box and you get 3.8 TB of local GPU memory, enough to keep a very large model or a swarm of smaller AI agents resident in a single server.

What Intel's memory gamble means for the Nvidia trade

Here is why this matters beyond the spec sheet. Nvidia's dominance rests partly on a hardware recipe only a few suppliers can feed, and that scarcity has helped it command premium prices and rich margins. Anything that offers a cheaper substitute, even for a slice of the market, chips at that story.

Intel is not pretending to match Nvidia's fastest training chips. Crescent Island is aimed squarely at inference, the part of the AI workload growing fastest as companies move from building models to running them at scale. Intel CTO Sachin Katti framed the shift as AI moving "to real-time, everywhere inference," according to Data Center Dynamics.

The catch is software. Intel's oneAPI toolkit is far less established than Nvidia's CUDA, the platform most AI developers already know, according to Tom's Hardware. Cheaper hardware no one wants to code for is just cheaper hardware.

For Intel (INTC), the stakes are existential. Its last serious run at AI accelerators, the Gaudi line, never hit its modest $500 million revenue target and was effectively lapped by Nvidia and AMD (AMD), according to Data Center Dynamics. A cheaper, easier-to-produce inference chip is a more realistic place to plant a flag.

My read is that Crescent Island is less a Nvidia killer than a wedge. It does not need to win the data center. It only needs to give cost-conscious buyers, especially companies running AI on their own premises, a reason to slot an Intel card into the rack instead of paying up for memory they cannot easily get.

Customer sampling is slated for the second half of 2026, so the real test is still months away. Watch whether board partners actually ship cards near that 480GB ceiling, and whether anyone publishes benchmarks pitting LPDDR5X against HBM on real inference jobs.

If those numbers hold up, the lesson for investors is older than any chip. The most expensive ingredient is only essential until someone proves it is not.

Related: Nvidia's latest product is a game-changer

The Arena Media Brands, LLC THESTREET is a registered trademark of TheStreet, Inc.

This story was originally published June 3, 2026 at 3:03 AM.