The hardware behind the future


Samsung and SK hynix Race to Stack More Memory as Nvidia Comes Calling

Korea’s two memory giants are planning a roughly 50 percent jump in Samsung’s HBM output and tens of trillions of won in new fabs, with Nvidia’s Rubin accelerators setting the pace.

Samsung and SK hynix Race to Stack More Memory as Nvidia Comes Calling

The two companies that make most of the world's advanced memory intend to make a great deal more of it. Samsung Electronics is preparing to lift its high bandwidth memory output by roughly half during 2026, the Korea JoongAng Daily reports, while SK hynix fills a freshly built plant with production tools about as fast as the movers can carry them. Both are chasing the same customer, and that customer is Nvidia.

First, a quick decoder ring for anyone new here. HBM, which the industry pronounces letter by letter, stands for high bandwidth memory. Instead of scattering memory chips across a circuit board like houses in a suburb, manufacturers stack them into a tower right beside the processor, so data takes a short elevator ride rather than a long commute. AI accelerators devour exactly this kind of memory, which is why everyone suddenly wants taller towers, and more of them.

The numbers involved are the kind you check twice. Samsung's monthly HBM wafer output is expected to reach 250,000 wafers by the end of 2026, up more than 47 percent from about 170,000 today, according to the JoongAng Daily. Its P4 fab in Pyeongtaek should add around 60,000 wafers of monthly DRAM capacity by the second quarter, and construction on the neighboring P5 plant began in November 2025, a 60 trillion won project, roughly $41.5 billion, aiming to start operating in 2028.

All that urgency has a name, and the name is Rubin. Nvidia's next processor family calls for sixth-generation HBM4 that moves at least 10 gigabits per second through each connection pin, and Samsung's HBM4 is reported to have earned top marks in Nvidia's internal testing at 11 gigabits per second per pin. Picture a pin as one lane on the bridge between memory and processor; a chip has thousands of lanes, so an extra gigabit on each adds up in a hurry. "We are internally reviewing the possibility of expanding HBM production," Kim Jae-june, vice president of memory business at Samsung Electronics, said during the company's October conference call, which is the sort of corporate understatement you learn to read like a weather forecast.

If KB Securities analyst Kim Dong-won has it right, the payoff could be substantial. "Samsung's HBM bit shipments in 2026 are expected to triple year on year to 11.2 billion gigabits, with HBM4 accounting for roughly half of total shipments," he said, a jump that could double Samsung's share of the HBM market to about 35 percent.

SK hynix is spending just as cheerfully. Its new M15X fab in Cheongju represents an investment of more than 20 trillion won, with the first cleanroom moving from trial runs to mass production by February and a second due for completion around the end of the year. At full utilization, projected for mid-2027, the plant should turn out about 50,000 wafers a month, and the company is weighing a 600 trillion won buildout of its Yongin complex, quadruple the 128 trillion won originally planned.

My favorite detail hides in the fine print. Demand is surging not only for HBM but for ordinary DDR memory and NAND flash drives too, per the report, which means the unglamorous chips in your laptop now compete for factory space with the celebrity silicon in Nvidia's accelerators. New capacity arrives between now and 2028, while the orders are arriving today. How that gap gets rationed is the thing to watch this year.

Leave a Reply