The cost of computer memory has quietly become one of the biggest stories in technology this year, and the numbers are hard to ignore. A 32GB DDR5 module that Samsung listed at $149 earlier in the year now carries a $239 price tag, a 60 percent jump in a matter of months. On the contract side, where the big buyers negotiate, prices for DDR5 have more than doubled, climbing to roughly $19.50 per unit from about $7 in 2025. For anyone building a PC, speccing a server, or shopping for a workstation, the sticker shock is real.
The cause is not a factory fire or a natural disaster. It is demand, and specifically the appetite of artificial intelligence. The three companies that make almost all the world’s memory chips, Samsung, SK Hynix, and Micron, have been steadily redirecting their factories toward High Bandwidth Memory, the specialized, stacked chips that feed AI accelerators inside data centers. HBM now consumes an estimated 23 percent of global DRAM wafer capacity, up from just 8 percent in 2024. The reason is simple math: a wafer turned into HBM earns three to five times the revenue of the same wafer turned into ordinary DDR5.
Sold out through 2026
When a market gets that lopsided, the shelves empty fast. SK Hynix told investors on its most recent earnings call that its HBM, DRAM, and NAND capacity is essentially sold out for all of 2026. Micron went a step further and walked away from the consumer memory business entirely, choosing to serve enterprise and AI customers instead. TrendForce, which tracks the industry closely, recorded conventional DRAM contract prices rising 90 to 95 percent quarter over quarter heading into 2026. That is not a gentle climb. That is a doubling.
The effects are already rippling out to the products people actually buy. RAM prices for gaming and desktop builds have surged nearly 90 percent, turning what used to be an afterthought in a build budget into a line item that can rival the graphics card. Apple quietly discontinued the largest memory configurations of its Mac Studio, and the maximum unified memory available on some models slipped downward as supply tightened. Nvidia’s compact DGX Spark AI desktop rose from $3,999 to $4,699. When even the companies at the front of the line are paying more, everyone behind them feels it.
What makes this crunch different from past memory cycles is how long the fix will take. Memory factories, called fabs, are among the most expensive and complex facilities humans build, and they do not appear overnight. New plants from Samsung and SK Hynix are not expected to ramp up production before the second half of 2027. Micron has committed to a staggering $200 billion long-term investment in U.S. manufacturing, but its first new fab does not start turning out DRAM until the middle of 2027. SK Hynix has gone so far as to warn that the imbalance could persist well past 2030.
There is a genuinely exciting story buried inside the pain, though. This shortage exists because AI has arrived faster and hungrier than the supply chain planned for. Every large language model, every image generator, every AI feature baked into a phone or laptop runs on memory, and the world simply wants more of it than the factories can currently make. The billions now pouring into new fabs are a bet that this demand is not a bubble but a lasting shift in how computers are built and used.
For buyers in the meantime, the practical advice is straightforward. If a memory upgrade or a new machine is on the horizon, sooner is cheaper than later, because prices are still climbing rather than cooling. Builders chasing large amounts of memory for local AI work are increasingly looking at unified-memory systems and clever clustering rather than stuffing a single machine with pricey modules. And businesses planning server refreshes should budget for memory to cost far more than it did a year ago, and lock in orders early where they can.
The memory squeeze is a reminder that the AI boom is not just software and clever algorithms. It rests on physical silicon, made in a handful of factories, and for now those factories are pointed squarely at the data center. The rest of the market is paying the price, and will keep paying it until the next wave of fabs comes online toward the end of the decade.
