Why Is SK Hynix Expanding Memory Capacity?

TL;DR

SK Hynix is expanding memory production because Chey Tae-won expects AI agents and computing systems to require far more storage and high-bandwidth memory. He says capacity takes four to five years to add, making near-term shortages and chipflation difficult to solve, while long-term agreements, joint ventures, and phased investment checks can reduce overbuilding risk.

Transcript

Chairman Chey. Thank you so much for having us at SK Hynix here in Korea. I wanted to start out by asking you, you know, you are in charge of a lot more than just this company, but how important is memory within SK Hynix even though you have more than 100 subsidiaries? Yes, this is AI era. So people need a lot of memory. So does the AI. So ri... Read More

Key Insights

• SK Hynix is SK Group’s most important subsidiary because AI systems require memory to retain expanding stores of data, knowledge, and experience. Chey compares present AI with a four-year-old child whose memory needs should increase as its capabilities mature.
• SK’s acquisition of Hynix in 2012 was a risky strategic bet intended to add a global business to a portfolio already spanning energy, chemicals, and telecommunications. Chey anticipated that a digital society would require extensive memory and data storage even before seeing the AI opportunity.
• Memory demand could approach five times current production capacity within ten years, according to Chey’s forecast. He expects AI-agent use to grow sharply within five years, creating the need for SK Hynix to expand supply capacity quickly despite the investment risks.
• Memory manufacturing capacity requires at least four to five years of lead time, limiting the industry’s ability to answer sudden demand growth. Chey says customers are requesting almost twice as many chips, while producers have not yet prepared enough additional capacity.
• Chipflation is the broader price pressure caused by expensive memory chips. Chey says memory prices have risen 40%–50%, creating difficulties for PC, smartphone, consumer-electronics, and automobile manufacturers, with some companies raising product prices as component costs increase.
• Memory is an AI infrastructure bottleneck because accelerators and computing chips cannot be produced or effectively deployed without sufficient HBM and other memory. Chey says SK Hynix plans to double capacity within five years, although customers indicate that even this expansion may be insufficient.
• Nvidia is SK Hynix’s most important customer because of its scale and central role in AI infrastructure. Chey nevertheless says the company is not dependent on only one buyer, citing Google, Microsoft, and other hyperscalers as customers that may also design chips.
• Expansion risk can be managed through conditional investment, repeated demand checks, long-term customer agreements, fab joint ventures, and programs such as memory as a service. Chey says SK Hynix wants to avoid both oversupply and undersupply while sharing some investment risks with customers.

Questions & Answers

Q: Why is SK Hynix expanding memory production?

SK Hynix is expanding because Chey Tae-won expects AI systems and AI agents to require much more memory as they gain functions, knowledge, and experience. He estimates that total memory needs within ten years could approach five times current production capacity. Since customers are already requesting substantially more chips, the company believes rapid capacity growth is necessary to meet demand.

Q: How long does it take to add memory manufacturing capacity?

Adding memory manufacturing capacity requires at least four to five years of lead time, according to Chey Tae-won. That delay prevents suppliers from responding immediately when demand rises unexpectedly. He says the next year may be the hardest period because memory companies were not prepared to expand quickly enough, even as customers began requesting almost twice their previous demand.

Q: What is chipflation, and why does it matter?

Chipflation is Chey Tae-won’s term for broader price increases caused by chips becoming too expensive. He says memory prices have risen 40%–50%, affecting PC makers, smartphone manufacturers, consumer-electronics companies, and automakers. Some businesses have difficulty finding chips, while others face component costs high enough to contribute to increased prices for their finished products.

Q: How does SK Hynix plan to avoid overbuilding factories?

SK Hynix plans to treat its announced expansion as conditional rather than automatic. Before each stage, it will check whether customer demand remains present. Chey Tae-won also cites long-term customer agreements and risk-sharing arrangements, including fab joint ventures and memory-as-a-service programs, as ways to reduce exposure to the memory industry’s historical boom-and-bust cycle.

Q: Why is Nvidia important to SK Hynix?

Nvidia is SK Hynix’s most important customer because it is currently the largest participant among the company’s AI-related buyers and plays a central role in AI infrastructure. Chey Tae-won admires its annual pace of product development and its ecosystem partnerships. SK Hynix develops new memory chips and technologies alongside Nvidia to support that continuing product cycle.

Q: Is SK Hynix too dependent on Nvidia?

Chey Tae-won rejects the idea that SK Hynix depends on only one customer, although he acknowledges that Nvidia is currently its most important buyer. He points to Google, Microsoft, and other hyperscalers as additional customers, some of which make their own chips. His position is that supporting the market’s largest customer reflects Nvidia’s present scale, not exclusive dependence.

Q: Why does SK Hynix need to cooperate with TSMC?

SK Hynix and TSMC must coordinate because AI accelerators require both processing capacity and sufficient high-bandwidth memory. Chey Tae-won says abundant memory is not useful if TSMC capacity limits GPU production, while abundant processing capacity cannot help if HBM is unavailable. TSMC also produces SK Hynix’s base die, making manufacturing cooperation and shared forecasting important.

Q: How could new memory technology reduce AI costs?

SK Hynix aims to support multiple memory approaches rather than relying on only one format. Chey Tae-won identifies DRAM as the foundation used to build stacked HBM and also mentions CXL-type solutions for AI data centers. The company’s broader objective is to expand effective memory capacity and lower AI token costs, which Chey considers too high for users.

Summary & Key Takeaways

• Chey Tae-won describes SK Hynix as SK Group’s most important subsidiary in the AI era. He argues that increasingly capable AI systems will accumulate more knowledge and experience, requiring substantially more memory. SK Group can also support AI infrastructure through its energy and telecommunications businesses, including future AI data centers.

• SK acquired Hynix in 2012 as a risky move into a more global, digitally oriented business. Chey says the investment became one of his most successful bets. He now expects AI-agent adoption to rise sharply and total memory requirements within ten years to approach five times current production capacity.

• Explosive demand and four-to-five-year construction lead times have made memory an AI bottleneck. Chey expects the next year to be especially difficult because producers were not prepared for such rapid growth. SK Hynix plans to expand conditionally, checking demand and using customer agreements and shared-risk programs to avoid oversupply.