SK Hynix Chairman Chey Tae-won on $720 billion AI memory expansion

Organized & Compiled by Deep潮 TechFlow

Guest: Chey Tae-won, Chairman of SK Group

Host: Katie Tarasov, CNBC

Podcast source: CNBC Television

SK Group Chairman Chey Tae-won on SK Hynix’s Major Memory Expansion — Full Interview

Broadcast date: August 13, 2026

Disclosure: This episode's guest, Choi Tae-won, is the Chairman of SK Group and the highest decision-maker of SK Hynix's parent company. The interview content directly focuses on the group’s expansion strategy and company stock price, representing an internal perspective from an interested party, not an independent analysis.

Under the leadership of Choi Tae-won, SK Group has over 100 subsidiaries, with SK Hynix now being its largest. This company, which began in the textile industry, nearly went bankrupt during the 1997 Asian financial crisis and spent over a decade under creditor management, is now a core player in global AI memory supply. Its market value has more than quintupled in a year, surpassing $1 trillion. In 2012, he made the decision to bet the group’s future on semiconductors, calling it "the most successful risk of my life." Now, he’s placing another bet: a $720 billion expansion plan to triple production capacity by 2034.

The core conflict in the interview is the age-old question: In the decades-long boom-bust cycle of the storage industry, every surge in production has ended in oversupply and collapse—why would this time be different? Choi Tae-won’s answer is that AI has fundamentally altered the underlying demand logic. Previously, storage demand was capped by the ceiling of “one device per person”—an individual buys only one phone per year. In the AI era, one person will have ten agents, each consuming memory. The cycle won’t disappear, but it will lengthen. He also acknowledges the other side of reality: memory prices have already risen 40% to 50%, forcing Apple to raise prices on its flagship products; he calls this “chip inflation” and admits there is no short-term solution, with next year set to be the most severe year of shortage.

AI needs memory, just as humans need a brain.

• AI is like a four-year-old child. As it grows, it will need more features, experience, and knowledge, and you'll need to store that knowledge somewhere, just like a person.
• In the past, we produced products; now, AI creates intelligence. Intelligence is different.

The 2012 Gamble

• At the time, I couldn’t see AI, but I believed a digital society would inevitably require massive memory. You have to store knowledge and data somewhere—and that’s what memory chips are for.
• This was a very risky bet, and looking back now, it’s one of the most successful bets I’ve ever made in my life.

Chip inflation and shortage

• We call this chip inflation. Chips are too expensive, so even Apple has to raise prices, pushing up prices across the entire economy. This isn’t a good thing, but it’s happening—and in the short term, I don’t have a solution.
• Customer demand has doubled this year, and everyone is requesting nearly twice as many chips. Next year will be the worst year for memory shortages.

Why is this cycle different?

• In the past, the memory business depended on population size—each person bought only one phone per year. In the AI era, you’ll have ten different agents doing tasks for you, and each agent will require more memory.
• I’m not saying there are no cycles, but the shape of the cycle must change—it will become longer. This is not the memory business we were familiar with in the past; this is a turning point, a structural shift.

Relationship with NVIDIA and TSMC

• NVIDIA is now our largest customer, launching new products every year—the pace is something I’ve always admired. We’re also advancing alongside them with new chips and technologies.
• Without NVIDIA, there would be no AI ecosystem. Even if TSMC has abundant production capacity, GPUs would be useless without sufficient HBM—we must work together.

Establishing a factory and going public in the United States

• Behind a large-scale wafer storage facility, approximately 600 to 700 companies support materials, chemicals, and services—without them, the factory cannot operate. So the issue isn’t just about water, electricity, and land; it’s about whether the entire supply chain can be relocated together.
• The customer wants us to build a factory in the U.S., and I’m willing to do it, but I’m still looking for the right location. I haven’t spoken directly with the government yet.

Memory is the most important segment of the SK Group.

Katie Tarasov: You oversee far more than just this one company—SK has over 100 subsidiaries. But why is memory considered SK Hynix’s most critical business?

We are now in the AI era, and both people and AI itself require vast amounts of memory. I often think of AI as a four-year-old child—while small now, as it grows, it will need more capabilities, experiences, and knowledge, which must be stored somewhere, just like a human. A child’s brain is small and can hold only limited memories; as a person matures and accumulates life experiences, they develop memory and wisdom. AI is nearly identical in this regard, which is why we need so many storage chips.

It’s no surprise that SK Hynix is now the largest subsidiary in the group. But for AI, we also need energy and networks—and our portfolio already includes both an energy company and a telecommunications company, which will build AI data centers in the future. This is a natural fit within our portfolio. You’re right: SK Hynix is now the most important component of the entire portfolio.

The 2012 Gamble: From Textiles to Semiconductors

Katie Tarasov: You started in the textile industry—I visited the museum yesterday. Was it your decision to turn it into a semiconductor company? And when did that happen?

It was 2012. At the time, the group already had a portfolio in energy, chemicals, and telecommunications, but I felt we needed something truly global. The digital trend of that era was already clear; I didn’t see AI coming, but I believed a digital society would inevitably require massive memory. You had to store knowledge and data somewhere—and that naturally pointed to memory chips. I went after those companies, thinking this might be the future. It was a very risky bet at the time, but looking back now, it’s one of the most successful bets I’ve ever made.

$720 billion bet: The reality of demand and capacity

Katie Tarasov: You're now making another huge bet with a $720 billion expansion. How do you convince others that this long-term, sustainable business can support such an investment?

Look at the demand forecast. In the future, people will use many more AI agents; I believe that over the next five years, the number of AI agents will increase by an order of magnitude compared to 2025, spreading across the entire globe, requiring massive amounts of memory. My assessment is that total storage demand will grow to nearly five times current production capacity within a decade—demand is enormous. To meet the world’s needs, we must expand production extremely rapidly, and that’s why I’ve made the decision to scale up.

Katie Tarasov: Is the capacity you're building sufficient to meet customer demand?

Customers are constantly asking for more chips, and current demand is almost explosive. This demand only emerged last year, and expanding production requires long lead times—at least four to five years. Next year will likely be the hardest, as no memory companies are yet ready to scale up, yet demand continues to rise. This year, demand has actually doubled; my customers are requesting nearly twice as much as last year. Over the coming years, they will need even more chips, but supply cannot keep up. It’s somewhat like a war—everyone wants more memory chips, and without them, they cannot produce AI computing systems or AI chips.

Chip inflation: Even Apple can't withstand it

Katie Tarasov: Memory prices have increased by 40% to 50%. Apple, one of your customers, has already had to raise prices on its flagship products. Will this continue?

We call this chip inflation. Chips are too expensive, forcing even Apple to raise prices, which in turn pushes up prices across the entire economy. This isn’t good, and I don’t want it to happen—but it has, and I don’t have a short-term solution. That’s why we’ve made massive investments to expand production, having already announced plans to double capacity within five years—but even that isn’t enough for most customers; we need even more. Yet physically, some things simply can’t be done—and that’s our current challenge, which we refer to as the AI bottleneck. Right now, memory chips are the bottleneck for AI.

Katie Tarasov: HBM and DRAM account for about three-quarters of your revenue. Are there risks to your business from others attempting to develop AI using alternative chip architectures that don’t require DRAM?

Technology is always evolving, and changes happen at any time. DRAM is the foundation, on which we build HBM, and in the future, solutions like CXL will emerge—there are numerous technologies capable of supporting AI data centers. But I believe our goal is to scale production by every possible means, with the ultimate aim of reducing the cost of AI tokens. Current token prices are too high; we must bring them down, otherwise people will spend too much on tokens.

Relationship with NVIDIA and TSMC

Katie Tarasov: Do you have a signed wafer in your office—is it from Jensen Huang? Are you friends? Who else in the AI circle are your friends?

The major hyperscale cloud providers—Satya Nadella of Microsoft, Sundar Pichai of Google, and Mark Zuckerberg of Meta—are primarily my clients and mostly my friends; I occasionally meet with them. TSMC is certainly among them too—I had dinner last month in Taiwan with C.C. Wei, and we exchanged a lot of insights and forecasts about the future. He also believes AI is a real technology, and we are both scaling up production significantly. Moreover, our efforts are mutually dependent: even if TSMC has ample capacity, GPUs are useless without sufficient HBM, and vice versa—so we must collaborate closely.

Katie Tarasov: How important is NVIDIA as a client? Are you too concentrated on this single client?

NVIDIA is currently our largest customer, launching new products annually—a pace I have always admired—and we work alongside them to advance new chips and technologies. Without NVIDIA, there would be no AI ecosystem; they have created vast solutions for AI infrastructure, and I am committed to supporting them as much as possible. But they are not my only customer—I also serve other clients such as Google, Microsoft, and other hyperscale cloud providers, some of whom are developing their own chips. NVIDIA is simply the largest at this moment; I am not truly dependent on a single customer—I simply aim to solve problems for the world’s largest clients.

Why is this cycle different?

Katie Tarasov: 2023 was a tough year for you, but you made it through. In 1997, the company nearly went bankrupt. What’s different this time?

I don’t usually believe claims that “this time is different,” but there is one thing that truly is new: we must confront AI, this newborn technology. AI has already transformed our society, consuming vast amounts of resources—energy, electricity, materials, silicon, and even water. Yet what it produces is intelligence, which is different from before. In the past, we only manufactured products; now, AI generates intelligence. We can become smarter and happier, though there are risks—we must carefully monitor those dangers. Overall, I truly believe AI’s greatest benefit is that it can liberate humans from labor, reduce disparities between people, and create a more relaxed society.

Katie Tarasov: So why will this investment last this time, rather than turning into another boom-bust cycle?

Look at demand. In past cycles, the memory business depended on population size, because it was tied to the number of hardware devices. One person can only use one phone at a time—you won’t buy ten phones in a single day—so population set an upper limit for the entire market. But in the AI era, it’s different: you’ll have ten different AI agents working for you, and each agent requires more memory, expanding the market significantly. I’m not saying cycles are gone, but the nature of the cycle must change—it will become longer. I can’t definitively tell you what the future holds, but I am certain this is no longer the memory business we knew in the past. This is a turning point, a structural shift. Strong demand momentum on the demand side is certain in the near term.

Building a factory in the U.S.: The supply chain is more complicated than hydropower.

Katie Tarasov: You mentioned that electricity and water are more important than location when choosing a site. Does the U.S. have sufficient resources to support an HBM production facility?

Look at the ecosystem of memory chips. Behind every major memory wafer fab in the world, there are approximately 600 to 700 companies supporting it with materials, chemicals, and services—without them, the factory cannot operate. This is the biggest challenge. Every country wants to build memory fabs on its own soil, but if we can bring water, electricity, land, and the entire supply chain with us, I’m willing to build anywhere—including the United States or anywhere else in the world. Our research team is evaluating where real opportunities exist, how quickly we can build new fabs, and how to alleviate memory shortages.

Katie Tarasov: You have plans to build a factory in Indiana. In July, the U.S. Secretary of Commerce pressured you and Samsung to bring complete memory production lines to the United States. What are your plans?

Everything I’ve told you is what we’re currently researching—I can’t directly announce, “We’re building a factory in X location.” We need to determine where, when, how, and what’s required, which involves a lot of detailed planning and numerous contracts to finalize. We’re actively moving forward on this. It’s not just the U.S.—I’m evaluating opportunities worldwide. I’m willing to do it, but first I need to find the right place. Honestly, we’ve been researching for more than a month—both in the U.S. and elsewhere. U.S. customers want us to build a factory domestically, but I haven’t yet had direct discussions with government officials; this is primarily a business decision. If our customers want it, I’m ready to seize this opportunity together.

Katie Tarasov: You've launched in the U.S. market—how did your first day perform?

That’s a significant step—even though it’s just the first one—as it marks our entry into the global financial market. We now have new shareholders and new financial supporters, which opens up tremendous opportunities for us in both financial and market terms. I can even offer better incentives to attract American engineers, which is why we decided to establish a new R&D center in the U.S. and build a factory in Indiana. We’ve also launched an AI company in the U.S., investing approximately $1 billion to integrate American technologies: combining our own memory technology with systems, simulation, and other AI capabilities to deliver superior solutions for AI data centers and infrastructure. Software must be tightly integrated with hardware; the U.S. has outstanding software engineers and a robust software industry, while Korea, Taiwan, and Japan excel in manufacturing. Bringing both strengths together is precisely why we founded this company.