An hour south of Seoul, in Pyeongtaek, there is a fabrication plant where wafers are produced at a speed that was previously calibrated for gaming PCs and smartphones. That calibration has changed recently. The majority of what is currently produced there is going somewhere else entirely, such as a data center in Northern Virginia, perhaps Oregon, or an under-construction campus in Abu Dhabi that no one outside the industry has yet heard of. The cost has changed along with it.
About 30% of all hyperscaler capital expenditures in 2026 will go toward memory, according to SemiAnalysis. Compared to 2023, when the percentage was around 8%, that represents a four-fold increase. That’s the kind of move that deserves attention in an industry that typically moves in slow, predictable cycles. In general, it hasn’t.
Together, the hyperscalers—Amazon, Alphabet, Meta, and Microsoft—have invested roughly $650 billion in capital expenditures this year. According to a recent report by Bloomberg, this amount surpasses the combined 2026 budgets of twenty-one other significant companies in a variety of industries, including defense contracting and automotive manufacturing. They referred to it as an unparalleled boom this century. It’s difficult to disagree.
However, the real destination of the money is being obscured by the headlines. Not precisely GPUs. Not the silicon being photographed for magazine covers. Recall. the dull stuff. The high-bandwidth chips, DRAM and HBM, are bolted onto Nvidia’s accelerators in a vertical stack. DRAM prices are predicted by SemiAnalysis to more than double this year and to increase by double digits in 2027. Since early 2025, the cost of LPDDR5 contracts has already tripled.
Some of this has a subtle absurdity. Nvidia receives “Very Very Preferred” pricing on DRAM, which is significantly less than what hyperscalers and the general market pay, according to SemiAnalysis. The severity of the supply shortage for everyone else is concealed by this preferential treatment. AMD is on the wrong side of that calculation because it carries more memory per accelerator and lacks Nvidia’s purchasing power. So does any business that manufactures phones, laptops, or servers.
This brings us to the point that no one wants to focus on during these earnings calls. Significant shortages are anticipated through at least 2027, according to a warning issued by Samsung’s memory chief in April. A week prior, SK Hynix stated essentially the same thing. These three businesses, along with Micron, hold more than 90% of the global DRAM market. It is no longer truly a warning when two of the three largest suppliers issue multi-year shortage warnings in the same month. It’s a prediction.
During a recent earnings call, Jeff Clarke, COO of Dell, used the word “unprecedented” to describe how quickly component costs are rising. Dell doesn’t usually target that language. By the end of this year, DDR5 64GB server modules may be twice as expensive as they were in early 2025, according to Counterpoint Research.
Customers are beginning to take notice. Because the chips that would have gone into everyday devices are instead being routed to data centers in Virginia and Loudoun County, a report from the UK earlier this spring suggested that the cost of these devices could increase by 15 to 30%. Walking through any electronics store these days gives you the impression that something has changed beneath the prices on the shelf, even though the employees who stock those shelves are unable to pinpoint exactly what.
The industry may correct itself, as it usually does. Memory has always fluctuated between glut and famine. However, analysts are beginning to believe that this cycle is unique. Consumer upgrades are not the source of the demand. It comes from a few trillion-dollar corporations that are developing infrastructure more quickly than anybody else. It’s still unclear if that results in a smooth landing or something more messy.
