This spring, you’ll notice something strange if you walk into a suburban Virginia computer parts store. Compared to earlier, the DRAM aisle appears thinner. Basic memory stick prices have increased in a way that seems disproportionate to anything that is happening on CNBC. If you ask the clerks, they shrug and say the same thing: “It’s the data centers.”
They are correct, but the majority of investors are still unaware of the true implications. Everyone’s narrative about AI revolves around GPUs. Jensen Huang wearing a leather jacket on stage. valuations of trillions of dollars. Every new server rack in Virginia, Texas, and Oregon is equipped with Nvidia chips. Beneath that headline, however, is a more subdued and peculiar shortage that has less to do with Nvidia and more to do with the tiny memory rectangles that are piled on top of every AI accelerator that is currently on the market. Furthermore, the market might not have factored it in at all.
The area of an AI chip that no one takes pictures of is called high-bandwidth memory, or HBM. It is positioned next to the processor and provides it with data quickly enough to keep up with the calculations. The most costly GPU in the world is a paperweight without HBM. Samsung, SK Hynix, and Micron are the three companies that make up the industry. That’s all. Industry insiders believe that this concentration, rather than the Taiwanese factories that receive all the attention, is the true bottleneck of the AI boom.
The extent of the squeeze was embarrassingly evident during Micron’s most recent earnings call. Investors were informed by management that the company can fill between half and two-thirds of its orders. All of 2026’s production has already been sold out. The Idaho fab, which was scheduled to go online in late 2027, is being accelerated. Suddenly, five-year supply contracts are being signed, which is unprecedented in the memory industry, which has always operated on one-year handshakes. As this develops, it’s difficult to ignore how different the tone is from that of the memory industry ten years ago, which was unable to go six quarters without a glut.
The less evident but possibly cleaner beneficiary is Broadcom. The ASICs it creates for Google, Meta, and OpenAI are part of its custom AI accelerator business, which reported $8.4 billion in revenue last quarter, more than doubling year over year. Currently, the AI backlog exceeds $73 billion. These are not the figures of a business that is following a trend. It is more difficult to write off CEO Hock Tan’s advice as hype because he has always been perceived as being too cautious to make promises he cannot fulfill.
The majority of American investors have never directly owned SK Hynix and are unlikely to do so in the future. It supplies Nvidia’s most cutting-edge chips and dominates the majority of the worldwide HBM market. Purchasing the stock is difficult because it is listed in Seoul. The opportunity still exists in part because of this inconvenience.
There are good reasons to maintain skepticism. Investors have previously been humiliated by memory; those who experienced the 2018 crash still remember it. Micron fell 18% in a single month as a result of Google’s recent TurboQuant work, which suggests a future in which software lowers memory demand. In a quarter, this entire transaction could be undone by an abrupt decrease in data center expenditures. Orders continue to come in, though. The factories continue to operate. And the chip that keeps everything afloat is the one that hardly anyone discusses.