Tech — Noozify Original — October 19, 2026

The AI Memory Squeeze: Why Your Next Laptop Costs More

By Adam Cannizzaro

Watch the story in half a minute

By June, a 32-gigabyte kit of Corsair desktop memory that had sold for about $100 last October was selling for about $440. The same month, a survey by the research firm SigmaIntel found that the 12-gigabyte memory package that goes into phones had gone from $77.10 to $145.90 in a single quarter, an 89 percent jump on top of a first quarter in which prices had already roughly doubled. Dell was reported to have raised prices on its business computers in December, Asus followed on January 5th with increases that Taiwanese resellers put at 15 to 20 percent on mainstream laptops, and in May Valve raised the Steam Deck OLED from $549 to $789, citing "rising memory and storage costs." The holiday shopping season that opens next month will be the first in which nearly every device on the shelf carries the cost. There was no factory fire and no shipping jam. The three companies that make almost all of the world's memory decided to sell it to someone else.

That memory is DRAM, the working memory that every phone, laptop, and server uses to hold whatever it is doing right now. It is a commodity, sold by the gigabyte on contracts that reset every quarter, and it comes almost entirely from three companies. Samsung, SK hynix, and Micron account for more than 90 percent of global production, and in the second quarter of this year they held 39, 26, and 25 percent of the market, according to Counterpoint Research. For most of the past decade the business ran in a familiar cycle. The makers would overbuild, prices would collapse, they would cut investment, a shortage would follow, and prices would spike until the next round of factories opened. What makes this spike different is where the demand came from.

An AI accelerator, the chip at the center of every data center that Microsoft, Meta, Amazon, and Alphabet are building, needs a kind of memory called high-bandwidth memory, or HBM, made by stacking DRAM dies on top of one another and bonding them beside the processor. Those four companies are projected to spend about $650 billion on AI data centers this year, up from $217 billion in 2024, and by one widely cited projection data centers will consume roughly 70 percent of the world's memory output in 2026. SK hynix reportedly sold out its entire 2026 production and stopped taking orders. The AI investment boom has had many prices attached to it. This is the first one to show up on a laptop.

The reason it reaches that far is arithmetic. HBM is expensive to make in the only currency that matters inside a chip factory, which is wafer area. A wafer is a thin disc of silicon, about 12 inches (300 millimeters) across, on which hundreds of chips are printed at once and then cut apart. A factory can process only so many wafers a month, so the share of each disc a product takes up is what it really costs to make. Micron has said that making a given amount of HBM3E, the generation shipping in volume now, takes roughly three times the wafer input of making the same capacity in DDR5, the memory in a new laptop, and that the ratio gets worse with each generation. The stacks are harder to make, too. Only dies that pass testing are stacked, the finished stack is inspected again, and a single bad die can ruin the whole package. So every wafer a maker moves from ordinary memory to HBM takes about three wafers' worth of ordinary memory off the market, at prices that make the trade worth it. TrendForce, the Taiwanese research firm that tracks contract prices, said on September 30th that suppliers continue to put their most advanced production lines on server products and that the rest of the market remains undersupplied. A new memory plant costs $15 to $20 billion and takes several years to build, so there is no quick way to make more.

The price path shows what that did. TrendForce's survey found that contract prices for conventional DRAM rose 93 to 98 percent in the first quarter of 2026 alone, which lifted the industry's quarterly revenue to about $97 billion, up 81 percent from the quarter before. The firm expected a further 58 to 63 percent in the second quarter, then 13 to 18 percent in the third, and now forecasts 10 to 15 percent in the fourth. Those later numbers are described as a slowdown, and they are, in the sense that a car doing 60 is going slower than one doing 90. Prices are still rising, from a base that has already doubled. Samsung's head of global marketing, Wonjin Lee, told Bloomberg at CES in January that "prices are going up even as we speak" and that the company was reaching the point where it would have to consider repricing its products. Samsung's phone division, the report added, was having trouble getting enough memory for its Galaxy handsets from the memory division down the hall.

Who pays is already clear. IDC, which tracks device shipments, modeled the shortage in December with a pessimistic scenario, and by February it reported that conditions had become worse than that scenario, forecast an 11.3 percent drop in PC shipments and a 12.9 percent drop in smartphones for 2026, pushed any real recovery to 2028, and said memory prices would not return to 2025 levels within its forecast horizon at all. The firm has called this "the end of an era of cheap, abundant memory and storage." Some of the cost shows up as a higher price, and some of it shows up as less: brands are shipping models with 12 gigabytes where last year's had 16. Phone makers, TrendForce says, are cutting production plans and living off existing inventory for the rest of the year, while PC makers are buying aggressively because they expect 2027 to be tighter still, and the supply of memory for PCs could actually shrink next year as more capacity moves to servers. Valve's hardware engineer Yazan Aldehayyat put it plainly to Bloomberg in July. "Honestly, it's getting worse," he said, and retail prices lag the supply picture by at least three to six months, which means the shelf price in December reflects a market that was already short in the summer.

How long is the question with the least satisfying answer. David McAfee, who runs AMD's consumer chip business, said in June that memory pricing should return to balance in about two years, which puts normal somewhere in 2028, and added that getting back to the old lows could take longer. The relief, when it comes, will come from new capacity at Samsung and Micron and from China's CXMT, the industry's newcomer, and all of it is years of construction away. Whether the data centers that bought the memory will earn it back is a separate argument. The bill, for now, is being paid at the checkout.

For a buyer, the arithmetic is simpler than the industry's. The memory inside the box is now one of the most expensive parts in it, so the step up from 8 to 16 gigabytes that used to cost a trivial premium is where the increase hides, and choosing the configuration you will need for the life of the machine beats planning to add memory later, when it will cost more. The refurbished market is still priced off last year's components. And the memory business has never had a shortage it did not eventually overbuild its way out of. The trouble is that the last three times it overbuilt, it was making chips for phones and PCs. This time it is building for customers who have already bought everything it can make through next year.