Scarcity by design: how AI reset the memory chip industry

Data centres absorb most of the world's memory, and the few firms that make it have steered their wafers toward the chips accelerators need. The shortage looks built into the structure rather than passing through it.

Demand set the terms

AI infrastructure has become the memory industry's centre of gravity. Data centres consume around 70 percent of the memory produced worldwide, and by some estimates AI and server uses take about two-thirds of DRAM capacity in 2026, on IDC and brokerage research. The spending behind it is large and rising, with the biggest cloud operators lifting capital budgets from around 217 billion dollars two years earlier toward an estimated 650 billion. High-bandwidth memory, the stacked DRAM that sits beside AI processors, is sold out for the year, and each bit of it displaces two to three bits of ordinary DRAM on the line. HBM has grown to roughly a quarter of DRAM wafer output, up from under a fifth a year earlier. That displacement is a change in kind rather than a swing in a familiar cycle, and it is the hinge the rest of the mechanism turns on.

A shortage by design

The supply side tightened by choice. Three manufacturers, Samsung, SK Hynix and Micron, account for more than 95 percent of DRAM, and they moved capacity toward HBM and data-centre DDR5, leaving commodity DRAM and consumer NAND short. Supply growth is running near 16 percent for DRAM and 17 percent for NAND, below the 20 to 30 percent that was normal after 2018, and suppliers have been filling less than half of the wafer volumes customers request. The makers went further, halting legacy DDR4 orders in near lockstep, winding down older NAND lines, and stepping back from consumer products. Prices show the result: conventional DRAM up 58 to 63 percent, NAND flash up 70 to 75 percent, and server DRAM up more than 60 percent in a single quarter, on TrendForce data. Past shortages eased once new fabs opened. This gap reflects a deliberate choice of product mix, which fab expansions in 2027 and 2028 will not quickly undo.

The order at the top

Concentration intensified around HBM. SK Hynix moved ahead of Samsung as the largest memory supplier for the first time in 2025, on Counterpoint data, built on an early HBM bet and its position as the main supplier to the dominant AI-accelerator maker. It holds most of that customer's HBM business and a majority of the orders for the next generation. Samsung, after qualification stumbles that cost it share, has been recovering ground, while Micron sits third and is expanding, projecting the HBM market will reach around 100 billion dollars by 2028 from roughly 35 billion in 2025. Advanced HBM remains the preserve of these three, and the profits have followed, with the leaders posting record memory earnings as the same scarce wafers command far higher prices.

Korea's tape reads the same cycle

The concentration is visible one layer up, in a national export line and the equity index that tracks it. Korea's exports passed 100 billion dollars in a single month for the first time in June 2026, and semiconductors did most of the lifting, with shipments nearly tripling year on year to 44.82 billion dollars, the first month above 40 billion. Chip exports rose 163 percent across the first half. One category supplied close to half of a record month. The chip figure is a price and volume story in memory, the same pull-through of high-bandwidth memory and the commodity chips around it. The equity market moves with the same driver: two large-cap memory makers anchor the Kospi, so the benchmark re-rates with the chip cycle rather than with the broad economy. A tape that rises with one export category also falls with it, and the memory cycle has historically turned faster than the economy beneath it. The record and the fragility are the same fact seen from two sides.

The Chinese tier

China enters on two levels. At the fab, its makers still sit largely at the lower end, estimated at 5 to 10 percent of the market and without leading-edge HBM, though the largest, a mainland DRAM maker that by shipment describes itself as China's first and the world's fourth, has filed to go public. Around those fabs, a listed supply chain of module, controller and interface makers turned the price surge into record results. One representative name, the Shenzhen memory company Longsys, guided first-half net profit to between 9.2 billion and 11 billion yuan, more than six hundred times its year-earlier figure, in a preliminary exchange filing, and its shares multiplied. Peers reported similar swings, with national self-sufficiency goals adding a second tailwind on top of the price cycle.

The two-sided cycle

Whether the run holds rests on two things staying in place: AI capital spending, and the manufacturers' discipline on supply. For the Chinese names in particular, the gains are cyclical rather than structural. Profits balloon as prices rise and as low-cost inventory is revalued, and they can reverse quickly when the cycle turns, with heavy inventories and negative operating cash flow the usual tells. That risk sits alongside the same discipline the majors are showing, since capacity that is slow to expand also limits how far prices fall in the next downturn. The question the industry's own leaders now pose is whether AI demand has changed memory's economics durably enough to hold both the shortage and the prices where they are. The cycle has two owners, the buyers of compute and the sellers of wafers, and it turns when either one blinks.