Samsung and SK Hynix: memory is gold, shortage is the new normal

Memory stopped being a commodity. It became the bottleneck of the entire AI industry — and the bill has already been issued.

SK Hynix reported its third quarter: revenue of 104.6 trillion won, operating profit of 83.3 trillion. Revenue grew 327.9% year on year, profit 632.1%. The analyst consensus stood at 98.2 and 76.9 trillion — the company walked past it without slowing down. The reason is simple: average DRAM selling price jumped 28% in a quarter.

Not a one-off

In the fourth quarter SK Hynix will likely break the hundred-trillion-won mark in operating profit for the first time, according to ZDNet Korea. The forecast rests on two things: shipments grow by double digits, DRAM prices too. The company isn't cutting costs, isn't optimising, isn't launching a new line. It sells the same silicon for more money.

Samsung crossed that line a quarter earlier. The two Korean giants are running neck and neck toward what the industry already calls the new normal.

Now the unpleasant part

DigiTimes reports that Samsung and SK Hynix expect the memory shortage to drag on for years despite rising investment in fabs. AI demand grows faster than capacity can be brought online. Manufacturing is harder, cycles are longer, every new line costs as much as a small city.

Put simply: money pours into memory, and the shortage doesn't close. The usual market logic breaks. Normally a high price pulls in capital, capital builds fabs, fabs crash the price. Here the last link is gone.

Who pays

Everyone buying AI hardware. Servers, accelerators, laptops, phones — memory sits inside each one. While Nvidia and the rest carve up compute orders, the Koreans collect a toll at the entrance: without their chips any AI cluster stays a nice diagram on a slide.

For device makers that's bad news. For Korea it's excellent: two companies profit from the hype around other people's models without shipping a single one of their own.

There is a second, less obvious effect. When memory gets more expensive for years, the winners are those who squeeze more out of less: efficient architectures, quantisation, context savings. Expensive memory is a hidden tax on inefficient software — and possibly a better incentive to optimise models than any paper on power consumption.

Prediction: The shortage holds at least until the next capacity cycle, and consumer devices with large memory configurations get more expensive before anyone in the industry admits it out loud.

Disclaimer: not officially confirmed — the company has not confirmed next-quarter figures; market and supply chain outlook comes from ZDNet Korea and DigiTimes

ZDNet Korea, DigiTimes