Everyone is scrambling for memory chips, just like a war.
"It's like a war," said Chey Tae-won, Chairman of SK Group, "everyone wants to buy memory chips."
In the past, memory chips used to be the most typical cyclical product in the semiconductor industry: when demand rose, the entire industry expanded production, and after supply exceeded demand, prices plummeted, forcing manufacturers to collectively "weather the winter".
Today, AI has rewritten the new cycle of the memory industry, and SK Hynix has become the most direct beneficiary of this change. In 2025, relying on its leading edge in HBM, SK Hynix captured nearly half of the global market share, and also became one of the most critical suppliers for AI chip giants such as NVIDIA.
SK Hynix holds nearly half of the global HBM market share. Source: CNBC
To bet on the next round of demand, SK Hynix is carrying out the world's largest memory fab expansion, with an investment of 72 billion US dollars, planning to triple its production capacity by 2034.
The reason Chey Tae-won gave is very simple: AI requires an increasing amount of memory.
He believes that the number of AI agents may increase by about 10 times in the next five years. Within a decade, the total demand for memory may reach nearly 5 times the current production capacity. At the same time, HBM is being co-designed with AI processors, and memory is no longer just a standardized commodity in the traditional sense.
However, the risks are equally clear: the memory industry has never lacked lessons of "overbuilding". Expanding production takes 4 to 5 years, while customer demand is already nearly doubling; long-term agreements, joint venture fabs, and "Memory as a Service" have all become new ways for the company to try to spread risks.
In Chey Tae-won's view, AI is pushing the memory industry into a new stage. In the past, memory demand mainly depended on the population and the number of hardware devices, and one person would not use dozens of mobile phones at the same time. AI is different: one person may have multiple AI agents in the future, and each of them requires memory.
He does not believe that the cycle of the memory industry will disappear, but argues that the cycle may become longer. This is also the main basis for SK Hynix to dare to carry out large-scale production expansion.
Chey Tae-won during the exclusive interview with CNBC. Source: CNBC
The following is the edited transcript of the exclusive interview:
01 "It's Like a War"
Q: The SK Group has more than 100 subsidiaries. How important is the memory business to the entire group?
Chey Tae-won: We are now in the AI era, and people need a huge amount of memory, especially for AI. I think AI is like a four-year-old child: it is still very small now, but as it grows up, it will need more functions, more experience and more knowledge, and all that knowledge has to be stored somewhere.
The same is true for humans: a young child does not have much memory in their brain, but when they become an adult, as experience accumulates, we gain more memory and wisdom. The same logic applies to AI, so AI requires a large number of memory chips. At present, SK Hynix is already the largest subsidiary under the SK Group.
At the same time, AI also requires energy and network support. The SK Group also has energy and telecommunications businesses, which are participating in the construction of future AI data centers. From this perspective, we have a relatively complete business portfolio. But as of now, SK Hynix is indeed the most important business segment of the group.
Q: When you decided to enter the semiconductor industry in 2012, did you already foresee today's AI opportunities?
Chey Tae-won: I did not foresee the AI boom back then. In 2012, we already had energy, chemical and telecommunications businesses, but I wanted to add more globalized businesses.
At that time, we were in the digital age, and I believed that digital society required a large amount of memory, because massive amounts of knowledge and data needed to be stored. So I thought memory chips might represent the future, and I made that decision. It was a very risky bet, but now it turns out that it is one of the biggest and most successful bets in my life.
Q: Now you are betting another 72 billion US dollars to expand production capacity. Why do you think long-term demand can support such a large-scale investment?
Chey Tae-won: If you look at future demand, you will find that people may use more and more AI agents. I think the future is the era of AI agents.
Taking 2025 as the baseline, the number of AI agents may increase by about 10 times in the next five years.
After they are spread all over the world, they will also require a large amount of memory. Therefore, what I see is that the total demand for memory in the next ten years may reach nearly 5 times the current production capacity. The demand is extremely huge, and we must rapidly expand our supply capacity to meet global demand, which is why I decided to expand production.
The first of the four fabs built by SK Hynix in the Yongin cluster in South Korea is taking shape. Source: CNBC
Q: Who are your peers in the AI industry?
Chey Tae-won: My contacts in the AI field are mainly the heads of those hyperscale enterprises, including Jensen Huang of NVIDIA, Satya Nadella of Microsoft, Sundar Pichai of Google, and Mark Zuckerberg of Meta. Almost all of them are customers of SK Hynix, and I meet them occasionally.
Q: What about TSMC?
Chey Tae-won: Of course, including C.C. Wei. About a month ago, I just met him in Taiwan, China, we had dinner together, and exchanged a lot of information and predictions about the future. He also believes that AI is a real transformative technology. No matter how much memory you have, if TSMC's production capacity is limited, there will not be enough GPUs; conversely, no matter how large TSMC's production capacity is, it will not work without sufficient HBM. So we must build partnerships and cooperate together. That is our relationship with TSMC.
Q: How do you secure sufficient production capacity on TSMC's manufacturing lines? Everyone is queuing up to get that capacity.
Chey Tae-won: He knows which parts are critical, because without them, there will be no AI production, especially for GPUs. So no matter how much production capacity they have, the demand is overwhelming. But I believe he will manufacture base chips for us.
Q: NVIDIA CEO Jensen Huang once wrote "Please produce more" on the wafer you showed. How important is NVIDIA to SK Hynix?
Chey Tae-won: NVIDIA is the most important customer at present. It develops new products and new technologies every year, and I really admire that speed. We are also working with NVIDIA to promote new chips and new technologies.
NVIDIA is one of the most outstanding companies in the world right now, and I believe it will still occupy a dominant position in the future. It does not only have GPU technology, Jensen Huang is also cooperating with other companies to build the AI ecosystem. He not only develops his own company, but also builds different types of partnerships to jointly build the entire ecosystem, which is very important.
Q: Concentrating so much business on a single customer like NVIDIA, won't that bring risks?
Chey Tae-won: Without NVIDIA, the AI ecosystem would not exist. NVIDIA provides a lot of solutions for AI infrastructure, so we must do our best to support it.
However, NVIDIA is not our only customer. Google, Microsoft and other hyperscale enterprises are also our customers, and some of these companies are also manufacturing chips on their own. NVIDIA is the largest one at present, so we hope to provide solutions for this largest customer, but that does not mean we only rely on one customer.
Q: Among these customers, is anyone helping you with this production capacity construction? What can they do to help you increase production capacity quickly?
Chey Tae-won: They always ask for more chips, and I am working hard to provide some solutions, but the demand is extremely strong right now. Expanding production capacity requires a very long lead time, at least 4 to 5 years. Therefore, next year may be the most difficult year, because all memory companies are not actually ready to expand production, but demand is still growing.
We believe that demand will double this year, and almost all customers are asking us to provide nearly twice the number of chips. It's a bit like a war, everyone wants more memory chips. Without them, they cannot produce AI computing devices and AI chips.
Another bad thing is that prices are rising too fast. PC manufacturers, smartphone manufacturers, consumer electronics and even automakers are facing difficulties in sourcing chips, and the price of some chips is too high. One of my jobs is to protect the entire memory market as much as possible, expand production capacity, and do it as soon as possible.
02 "Chip Inflation"
Q: Memory chip prices have risen by 40% to 50%. Do you expect this situation to continue?
Chey Tae-won: This is what we call "chip inflation". Chip prices are so high that even Apple has to raise the prices of its flagship products, which actually pushes up prices across the whole society. I don't want this to happen either, but it has already happened. I have no short-term solution, which is why I am investing huge sums of money to expand fabs.
Due to the shortage of memory chips, Apple has raised the prices of MacBook and iPad. Source: CNBC
We have announced that we will double our production capacity within five years, but that is not enough for most customers, they need more. The problem is that some things are physically impossible to achieve, which is what we now call the AI bottleneck.
Q: HBM and DRAM account for about three-quarters of SK Hynix's revenue. If AI chips start adopting other technologies in the future, will that pose a risk to you?
Chey Tae-won: Technology is evolving all the time. DRAM is just the foundation, we stack on top of DRAM to form HBM, and in the future there may also be CXL (Compute Express Link) and other technologies to support AI data centers.
There will be multiple options in the future, but our goal is very clear, which is to expand memory capacity in various ways. Ultimately, we hope to reduce the Token cost of AI. Right now I think the price of Token is still too high, it must be further reduced, otherwise people will spend too much money on Tokens.
Chey Tae-won shows the HBM4 chip to the interviewer at SK Group headquarters in Seoul, South Korea. Source: CNBC
Q: The memory industry used to suffer from overbuilding caused by large-scale capital expenditure very often. How do you avoid this situation?
Chey Tae-won: We have always been very cautious. Although we have just announced a very large expansion plan, this plan is based on the premise that demand exists. Before every expansion, we will check whether the demand is real. Customers have signed some long-term agreements with us, and we also have different ways to share risks together, such as establishing fab joint ventures, or launching "Memory as a Service". This is not just selling chips, we can share risks through other methods.
We don't want oversupply, and we don't want undersupply either, so we keep checking demand. Now another important issue is HBM. People need a large number of AI accelerators, and probably every country and every market wants to have these devices. To meet demand, we produce a large amount of HBM, but there is a problem with HBM expansion: it consumes more than four times the wafer space of ordinary DRAM, and will take up a lot of wafer production capacity.
Therefore, we must expand the overall wafer production capacity. I think there are indeed risks on the supply side, but we are aware of that, and we are carefully confirming whether future demand will materialize.
Q: What changes have taken place in the long-term contracts requested by customers now?
Chey Tae-won: This is not something we unilaterally proposed. In fact, customers want to sign long-term agreements, because they must ensure the supply of memory.
If customers only pursue short-term profits, it will not help ease the current supply shortage. These agreements are indeed long-term agreements, but we want to give customers more certainty. We will check demand year by year and renew long-term agreements. This allows both parties to work together, continuously confirm demand, avoid over-investment, and also avoid under-investment.
Q: It is said that hotels all over South Korea are packed with technology companies coming to sign long-term contracts, is that true?
Chey Tae-won: Yes. But as I said, next year may be the year with the most severe chip shortage. A lot of people want to buy chips now, but our supply capacity is limited. Every time I meet a customer, I feel sorry about that, because if everyone wants all the chips, I can't provide all of them, I can only do my best.
03 2026, Hynix Enters the US Market
Q: 2023 was a difficult period for you, and the company almost went bankrupt in 1997. What will be different this time?
Chey Tae-won: I don't think it will be completely different. But there is one difference, we have to deal with a new thing, the new "baby" (AI). AI has already changed our society. It consumes massive amounts of energy, electricity, materials, silicon, and even water. It also creates intelligence. In the past, we just manufactured products, but people can use them to gain intelligence. We can become smarter and happier. Of course there are also dangers, but we must stay alert.
Overall, AI can help humans reduce the labor burden, narrow the gaps between people, and greatly ease social tension. We don't have to work so hard, and there will be more opportunities to build a happier society.
Q: You mentioned that power and water are more important than geographical location when choosing a site. Does the United States have enough resources to support HBM production facilities?
Chey Tae-won: Building a large memory fab requires the support of nearly 600 to 700 different companies to supply materials, chemicals and services, otherwise