SK hynix shrugs off memory peak fears, cites strong AI demand and long-term supply deals

SK hynix shrugged off concerns about a so-called memory peak-out, saying big tech companies' investment in AI infrastructure will remain solid beyond 2027, driving broader memory demand higher.

The chipmaker, which has announced major investment plans in Yongin, Gyeonggi Province, and the South Jeolla-Gwangju integrated special city, said capacity expansion would be carried out flexibly based on confirmed demand. It added that medium- to long-term investment increases are unlikely to lead directly to oversupply.

SK hynix also said future production facility investments — whether domestic or overseas — will be evaluated based on factors including power, water, labor, supply chain and customer accessibility, while making clear that no additional plans have been decided at this point.

The company forecast DRAM demand to grow in the mid-20 percent range this year and NAND demand to rise in the high-10 percent range. It also projected that once supply constraints ease, latent demand will revive and market growth will accelerate further.

Song Hyun-jong, president and head of SK hynix's Corporate Center, said at the company's second-quarter earnings conference call Wednesday that major customers are still requesting more memory supply.

Having concluded long-term supply agreements, or LTAs, with about 10 customers, Song added that the contracts include financial mechanisms such as deposits to reinforce reliability and commitment.

Cash and cash equivalents at the end of the second quarter rose 33.6 trillion won ($22.9 billion) from the first quarter to reach 88 trillion won. The company said it will prioritize investment in future growth opportunities given its significantly strengthened cash generation, while also reviewing various options for additional shareholder returns. SK hynix's total investment this year is expected to reach the high-40 trillion won range.

SK hynix plans to ramp up production of its sixth-generation HBM, HBM4, which began mass-production shipments in the second quarter, along with 1c (10-nanometer-class sixth-generation) general-purpose DRAM, in the second half of the year. The company said this will drive stronger earnings growth in the second half. It also expressed confidence that the development schedule for products beyond HBM4E is proceeding smoothly.

SK hynix dismissed concerns about a pullback in big tech AI infrastructure spending, pointing to medium- and long-term demand it is currently discussing with major customers.

Park Jun-deok, executive vice president in charge of DRAM marketing, said at the conference call Wednesday that AI investment by major cloud service providers will continue over the medium to long term, based on memory demand discussions under way with customers.

Concerns about reduced AI infrastructure investment had emerged following the appearance of highly efficient AI models and reports that Meta is considering leasing out its AI data center capacity. Park said the trend reflects a process of maximizing utilization and monetizing large-scale AI infrastructure already built — not a pullback in investment.

Park acknowledged that physical constraints such as securing power and building data centers could cause some timing shifts in individual projects. He said AI infrastructure investment will remain solid beyond next year as competition among cloud service providers and service expansion continue.

SK hynix projected that demand will grow across the full memory spectrum — not only high-performance AI memory chips such as HBM used directly in AI computing, but also server DRAM for agentic AI and high-capacity NAND.

SK hynix said it is working to shorten construction timelines for new production facilities and is making every effort to expand capacity, but acknowledged that the current supply-demand tightness will be difficult to resolve in the short term. Instead, the company said it is focused on securing both supply stability and earnings stability over the medium to long term through LTA negotiations with customers.

Park said LTA contract terms typically default to five years, though specific conditions can vary depending on the customer and product.

He added that while it is difficult to disclose the exact share of LTAs in total sales, the company plans to manage them at an appropriate level in light of market conditions and customer demand.

SK hynix said the LTAs are designed to minimize future earnings volatility and to strengthen strategic partnerships for developing next-generation memory chips in line with customers' technology roadmaps.

Park said the agreements will help customers build medium- to long-term demand plans and will also allow SK hynix to manage its own investment and production planning more efficiently based on demand visibility.

With HBM competition intensifying from HBM4 onward following Samsung Electronics' early mass-production shipments, SK hynix said HBM4's mass-production yield and quality are on par with HBM3E, emphasizing that the product has quickly reached full stride.

SK hynix is currently in talks with major customers on HBM supply volumes and prices for next year, and said negotiations are proceeding smoothly given still-solid customer demand. HBM4E, for which sample shipments began last month, is on track for full-scale mass production starting next year.

SK hynix is also developing "iHBM" technology to address heat dissipation issues in next-generation products such as HBM5. iHBM is a next-generation memory technology that integrates a cooling element directly into the HBM package, reducing heat generation by more than 30 percent compared with conventional designs.

Park said iHBM is expected to contribute to improved system stability and operational efficiency in high-performance, high-density AI environments.

He also expressed confidence in the face of competition from rivals, saying that as the AI market expands and AI accelerator performance and packaging structures grow more sophisticated, customers will increasingly prefer partners with proven mass-production capabilities, quality and stable supply.

Meanwhile, Kim Woo-hyun, chief financial officer and executive vice president, said converting Wonju shares into American depositary receipts may require the issuer to complete a regulatory filing process, which typically takes at least several weeks. He added that any future expansion of the ADR share will be reviewed taking into account the regulatory environment and other factors.