South Korea is becoming the centre of an unprecedented race to expand memory-chip production, through the explosive growth of artificial intelligence. The reason behind the push is SK Hynix, which is committing hundreds of billions of dollars to expand its manufacturing footprint and meet what it expects to be sustained demand for advanced memory.
The company says its planned network of factories will become the world’s largest memory manufacturing complex. The scale of the investment shows how dramatically the economics of the memory business have changed as AI companies compete for limited supplies of high-bandwidth memory, or HBM.
SK Hynix’s market value has risen more than fivefold over the past year, crossing $1 trillion at one point, as investors have bet on its position in the AI supply chain. But the stock has also shown how volatile that enthusiasm can be. Its US-listed shares have fallen about 21% from their July 14 peak of nearly $195, closing Wednesday at $154.41.
HBM has become the critical bottleneck
HBM is a specialised form of dynamic random access memory designed to move data rapidly between memory and processors. It has become an essential component in the systems used to train and operate AI models, particularly as companies deploy increasingly powerful processors.
The technology requires large quantities of DRAM to be stacked together, putting additional pressure on overall memory supplies. This has created an unusual situation for an industry historically known for sharp cycles of oversupply, falling prices and weak margins.
AI companies are now locking in memory supplies much further in advance, while chipmakers such as Nvidia are taking greater risks to ensure they have access to the components needed for their next generation of processors.
“It’s like a war,” said Chey Tae-won, chairman of SK Group, the controlling owner of SK Hynix told CNBC. “Everybody wants to buy the memory chips. Without that, they cannot produce their AI computing and AI chips.”
The HBM market remains concentrated among a handful of manufacturers. SK Hynix accounted for 58% of the market in the first quarter, according to Counterpoint Research. Samsung and US-based Micron each held 21%.
This concentration has turned South Korea into an important destination for the world’s biggest technology companies. Research director MS Hwang of Counterpoint Research said hotels around the areas where SK Hynix and Samsung are expanding their factories are fully booked. “If you name any company in Big Tech, they are all in Korea to sign a contract,” he told CNBC.
The rush is also changing the relationship between memory manufacturers and their customers. Instead of treating memory as a largely interchangeable commodity, AI companies are increasingly working directly with manufacturers on customised versions designed around specific processors.
SK Hynix is building for years of demand
SK Hynix raised $26.5 billion through a July share listing on the Nasdaq, the largest amount ever raised by a foreign company through a US listing. The proceeds come as the company embarks on an enormous expansion programme in South Korea.
The effort is also aligned with a major government push. South Korean President Lee Jae Myung has called for the country’s memory production to double over the next five years, alongside plans for several new Samsung facilities.
At Yongin, south of Seoul, SK Hynix is already constructing the first of four planned fabs. The site illustrates the unusual engineering demands created by South Korea’s limited and mountainous terrain.
Unlike many newer US semiconductor facilities that spread horizontally across a large footprint, the Yongin fab is being built vertically. SK Hynix says the structure will eventually rise to roughly the height of a 50-storey apartment building and contain six cleanrooms spread across multiple floors.
The company is also expanding its infrastructure at Cheongju, where CNBC recently toured its memory manufacturing facilities.
Nvidia and other AI giants are locking in supply
The expansion comes as SK Hynix deepens its relationships with the companies driving the AI boom. The company said in July that it had signed 10 long-term agreements with major customers.
Nvidia has also secured a stable supply of HBM from SK Hynix and agreed to work with SK Group on next-generation memory. That agreement forms part of a wider $500 billion deal involving SK Group and plans for new data centres with SK Telecom by 2027.
SK Group chairman Chey Tae-won said he also meets periodically with senior executives from Microsoft, Google and Meta. Nvidia CEO Jensen Huang and AMD CEO Lisa Su have both visited South Korea for discussions around memory supplies.
Tae-won acknowledged that the surge in demand has pushed prices sharply higher.
“prices went up too fast,” he said, adding that he hopes the industry’s expansion will eventually ease the pressure. “I’m trying to do my best,” he told CNBC.
China adds another layer of pressure
SK Hynix operates three memory fabs in China, but U.S. export restrictions prevent the company from selling its most advanced HBM products there.
China is nevertheless trying to develop its own HBM capabilities as it seeks to reduce its dependence on foreign semiconductor technology. CXMT, one of the country’s emerging memory manufacturers, recently made a major debut on the Shanghai market.
For the Korean memory industry, China’s progress adds another dimension to the investment race.
“It’s a race, and now the counterparty of the race is China,” Hwang told CNBC. “That’s going to be a lot more dangerous than anything else in the world.”
South Korea is not the only country trying to build more memory capacity. Micron is investing $50 billion in two new memory fabs in Boise, Idaho, with the first expected to begin wafer production in 2027.
The company is also developing a much larger $100 billion semiconductor campus in Clay, New York, which could eventually accommodate four fabs.
SK Hynix itself is expanding its American presence. Its first U.S. manufacturing facility is being built in West Lafayette, Indiana, with construction expected to finish in 2028. The $4 billion facility will focus on advanced packaging, where memory chips are stacked and connected, rather than front-end chip manufacturing.
The company already has a significant US footprint through Solidigm, the business created after SK Hynix acquired Intel’s NAND operation for $9 billion in 2020. It later reorganised the business and launched a $10 billion AI-focused company in the U.S.
Custom memory could change the equation
The biggest long-term shift for SK Hynix may not simply be the amount of memory being produced, but the way that memory is designed.
The company believes customised HBM could make the industry less vulnerable to the traditional boom-and-bust cycle. Instead of producing a largely standard product for a broad customer base, memory manufacturers are increasingly designing products alongside AI-chip companies.
That trend is expected to become more important with future generations such as HBM5. “Nvidia wants their own custom chips and Google wants their own customized HBM, so it’s not just a commodity,” Tae-won said. “It actually changes the memory chip’s status.”