SK hynix (SKHY.US) reported its third-quarter 2026 results and held an earnings call. The company posted quarterly revenue of 104.6 trillion won, up 32% quarter on quarter and 328% year on year, surpassing 100 trillion won in a single quarter for the first time in its history, while operating profit reached 83.3 trillion won, up 38% sequentially and 632% year on year, lifting the operating margin 3 percentage points sequentially to 80% and setting fresh quarterly records for both operating profit and margin.
By product, DRAM bit shipments rose about 10% quarter on quarter, with HBM4 entering full mass production and shipment in the third quarter and server DDR5 shipments up more than 40% sequentially, while the DRAM average selling price climbed to the high-20% range sequentially; NAND bit shipments edged down sequentially on portfolio restructuring, but average prices rose to the mid-20% range quarter on quarter.
Notably, net profit fell 58% sequentially to 39.8 trillion won, which management explicitly attributed to non-operating items: a foreign-exchange-related net loss of 11.6 trillion won and a valuation loss on investment assets of 22.9 trillion won, for a combined non-operating net expense of 32.6 trillion won, unrelated to the operating side.
On the outlook, the company expects global CSP capital expenditure to grow more than 60% in 2027, with secured power capacity up more than 100% year on year, and the global AI data center infrastructure market to sustain roughly 40% annual growth through 2030; constrained by supply capability, however, demand growth achievable in 2027 is seen as similar to this year, at a mid-20% level for DRAM and a high-teens level for NAND. Fourth-quarter guidance calls for DRAM bit shipments to rise in the mid-20% range sequentially and NAND to grow in the mid-single-digit range.
On capital expenditure, the company said 2027 spending will increase significantly from 2026, directed at advanced-node equipment, the ramp of Fab 1 in Yongin and domestic and overseas infrastructure, though it gave no specific figure.
Key takeaways from the analyst Q&A: questions from eight analysts clustered around four main threads. First, the durability of AI demand: Park Joon-deok, head of DRAM marketing, offered two grounds, namely that the latest earnings of major hyperscale customers showed cloud revenue growing 30% to 80% year on year, cloud order backlogs growing twofold to fivefold year on year and steady growth in paying users, indicating that investment is producing real business results, and that the investor base has spread from a handful of big tech firms to non-tech companies building their own capacity, sovereign AI and on-premise deployments, so that any single company's investment adjustment can be offset by other demand. Second, long-term agreements (LTAs): the company has signed LTAs with 14 core customers including major global CSPs, locking in more than 50% of total capacity for 2027 and 2028; he added that locked volumes plus demand under negotiation mean DRAM and HBM capacity through 2029 is effectively nearly sold out, with the caveat that volumes under negotiation are based on customers' specific procurement plans and carry a high probability of materializing.
Third, the risk of oversupply: CFO Kim Woohyun responded that more investment and more fabs do not equal more supply, with supply discipline resting on engineering pacing, as new plants such as the Yongin cluster and M17 first accelerate shell construction while cleanrooms are brought online in phases and equipment move-in and mass-production ramps follow demand; demand assessment also weighs customer financial health, financing capability and the readiness of key supply chains. Fourth, the downgrade in HBM stacking layers: Kim Ki-tae, head of HBM sales, characterized it as evidence of overheating demand, saying customers currently request more than twice the volume the company can supply, and that when shifting from 12-layer HBM4 to 8-layer HBM4E, higher per-die capacity can offset most of the impact without affecting supply and profitability plans.
What the market is watching. The real incremental information lies not in the 80% margin but in two points. The first is the phrase sold out through 2029, which goes a step further than the official press release, since the release only said LTAs cover more than 50% of 2027-2028 capacity, while the call pushed the horizon to 2029, and the accurate formulation is that locked LTA volumes plus demand under discussion are sold out in aggregate, not contracted volumes alone. The second is that Solidigm withdrew its U.S. listing plan: CEO Kwak Noh-jung said at the outset that Solidigm had hoped to bring in outside capital through an IPO to ease the investment burden, but after assessing recent financial market volatility and the correction in semiconductor share prices, concluded that an IPO at this time could not fully realize expected returns, and therefore halted its review of pre-IPO and IPO options.
Paired with that is roughly 100 trillion won of shareholder returns for fiscal 2026: more than 30 trillion won in cash dividends and more than 60 trillion won in buybacks and cancellations, including the 40 trillion won announced in August, completed last week and soon to be canceled; the board also resolved that day on 14 trillion won in third-quarter cash dividends and 20 trillion won in buybacks, of which about 10 trillion won for executive and employee compensation is excluded from shareholder returns, putting actual third-quarter returns at about 24 trillion won. The per-share dividend is 20,000 won, made up of 10,000 won as a performance-improvement dividend and 10,000 won as a special dividend from Kioxia investment gains, with the remainder to be folded into the year-end dividend after fiscal-year settlement.
Prepared remarks
Moderator: Good morning, and thank you for joining SK hynix's earnings call. Today we will begin with prepared remarks and then move to the Q&A session. If you have a question, please press star one. Please note that the prepared remarks will be simultaneously interpreted, while the Q&A will be consecutively interpreted. We will now begin. Park Seong-hwan, Head of IR: Good morning, good afternoon and good evening. I am Park Seong-hwan, head of investor relations at SK hynix. Welcome to SK hynix's third-quarter 2026 earnings call. Let me introduce the executives joining me today: President and CEO Kwak Noh-jung, CFO Kim Woohyun, head of DRAM marketing Park Joon-deok, head of NAND marketing Song Chang-seok, and head of HBM sales and marketing Kim Ki-tae. I would like to make a disclaimer: the third-quarter results discussed on this call are preliminary consolidated figures and have not yet been reviewed by an external auditor, so the data may still change. In addition, forward-looking statements including market outlooks and company plans may change with macroeconomic and market conditions. We will now begin SK hynix's third-quarter 2026 earnings call. Before the results presentation, CEO Kwak Noh-jung will first address Solidigm's IPO and shareholder returns. CFO Kim Woohyun will then present the third-quarter results, the company's future plans and the market outlook.
CEO: Solidigm ends listing review, Kioxia gains returned as special dividend
Kwak Noh-jung, President and CEO: Hello, I am CEO Kwak Noh-jung. I would like to express my sincere gratitude to our shareholders for your unwavering interest and trust in SK hynix even amid heightened market volatility recently. Today I want to speak directly to two matters: our decision on Solidigm's IPO, which has drawn considerable market attention recently, and a special dividend to share the returns from our strategic investments with you. First, Solidigm has been exploring various financing options to support initiatives such as expanding its manufacturing footprint and meeting growing demand. Among the options considered, bringing in outside capital through an IPO had been expected to ease our investment burden and strengthen our capacity for shareholder returns. However, after carefully assessing recent financial market volatility and the correction in semiconductor share prices, Solidigm concluded that an IPO at this time could not fully realize expected returns, and has therefore decided to halt its review of potential pre-IPO and IPO options. Solidigm will continue to strengthen its competitiveness and pursue growth opportunities in the AI data center storage market. Investments and financing arrangements supporting these efforts will be pursued in a way that enhances both Solidigm's and SK hynix's corporate and shareholder value. We have consistently upheld the principle of sharing most of the returns from strategic investments with shareholders. By using those returns to supplement our shareholder return capacity, which may temporarily decline as we invest for the future, we aim to both protect shareholder interests and continue bold investment in sustainable growth. In line with this principle, we intend to return the gains realized this year from our Kioxia investment to shareholders through a special dividend. We will firmly honor this commitment to ensure shareholders benefit directly from the success of our strategic investments. Next, the CFO will present this quarter's results and outlook, along with the details of the shareholder return plan. Thank you.
CFO: Third-quarter results
Kim Woohyun, CFO: Hello, I am CFO Kim Woohyun. Let me begin with the company's third-quarter 2026 results. In the third quarter, tight supply and demand persisted on growing AI memory demand, and memory prices continued to rise. DRAM and NAND recorded high price increases for a second consecutive quarter, with gains concentrated in HBM, server DRAM and enterprise SSDs. As a result, third-quarter revenue rose 32% sequentially and 328% year on year to 104.6 trillion won. This was the first time quarterly revenue exceeded 100 trillion won, setting a historic high. In DRAM, we expanded sales centered on AI servers, with bit shipments rising about 10% sequentially in line with plan. Notably, HBM4 began full mass production and shipment in the third quarter, and server DDR5 shipments grew more than 40% from the second quarter, driving overall DRAM bit shipment growth. With general DRAM prices remaining firm and the share of HBM4, priced above the prior generation, expanding, the DRAM average selling price rose to the high-20% range sequentially, a similar increase to the previous quarter. In NAND, as we shifted sales toward higher-value products such as 321-layer TLC enterprise SSDs, bit shipments declined slightly from the second quarter, but with overall prices strengthening and the SSD mix expanding, average prices rose to the mid-20% range sequentially. Despite unfavorable foreign exchange effects from a stronger won, higher product prices and expanded sales of high-value products drove third-quarter operating profit to 83.3 trillion won, up 38% sequentially and 632% year on year. The operating margin also improved 3 percentage points sequentially to 80%, with both operating profit and margin hitting quarterly record highs. Third-quarter depreciation and amortization totaled 4.3 trillion won, while EBITDA reached 87.7 trillion won for an EBITDA margin of 84%. On major non-operating items, foreign exchange movements produced a related net loss of 11.6 trillion won and investment asset valuation losses of 22.9 trillion won, for a combined non-operating net expense of 32.6 trillion won. Accordingly, pre-tax profit was 50.7 trillion won and net profit was 39.8 trillion won, for a net margin of 38%. As of the end of the third quarter, cash and cash equivalents stood at 114.5 trillion won, up 26.5 trillion won from the end of the prior quarter, while interest-bearing debt fell 2.8 trillion won to 15.8 trillion won. Net cash therefore expanded to 98.7 trillion won, and the debt-to-equity ratio improved 2 percentage points from the end of the prior quarter to 5%.
Market outlook
Next, let me discuss the market outlook. The AI market is continuously evolving. As AI drives productivity gains and new revenue creation across software development, customer response, search and other areas, a virtuous cycle is forming in which rising AI usage leads to expanded infrastructure investment. This shift is reflected in the investment scale and secured power capacity of major global cloud service providers. Recent estimates by market research firms put global CSP capital expenditure growth in 2027 at more than 60%, with secured power capacity also expected to grow more than 100% year on year. More notably, this growth will not stop next year, as the global AI data center infrastructure market is expected to sustain roughly 40% high annual growth through 2030. This medium- to long-term growth trend is likely to continue, driven by the commercialization of AI services and the expansion of application areas. In particular, as AI begins to plan complex tasks and execute them over longer periods using multiple tools, demand for AI compute and the volume of data generated, transmitted and stored are rising sharply. As a result, demand growth across the entire memory sector is significantly exceeding initial expectations. AI infrastructure investment has now moved past the early stage of a few companies racing to invest first, with participants expanding to more enterprises and national governments. Therefore, even with some degree of investment adjustment, we expect growth in AI infrastructure investment to remain strong. In the near term, enterprise and consumer AI will lead demand growth, while over the medium to long term, physical AI such as autonomous driving and robotics is expected to broaden the demand base. Although PC and mobile device shipments are expected to decline amid cost pressure from higher memory prices and supply shortages, demand growth from AI infrastructure is expected to more than offset that decline and sustain overall memory market growth. On the other hand, memory supply cannot keep up with demand growth at this level, so we expect it will be difficult to fully satisfy customer volume demand in the near term. Given the difficulty of advanced nodes and the time required to build new production facilities, the situation of supply lagging demand is expected to persist for some time, with this constraint likely extending to 2030. Despite very high demand, given actual supply conditions, demand growth achievable in 2027 is expected to be similar to this year, at a mid-20% level for DRAM and a high-teens level for NAND. We expect this favorable market environment to serve as a solid foundation supporting the company's earnings growth next year. Based on expanded capacity and the competitiveness of high-value products, we plan to achieve significantly higher results in both absolute revenue and profit.
Fourth-quarter sales plan and operating direction
In the fourth quarter, we plan to further accelerate supply expansion to respond more proactively to market demand. For DRAM, increased output from the expanded 1c nanometer process and higher HBM4 output from the M15X ramp will occur simultaneously, so fourth-quarter shipments are expected to grow significantly by a mid-20% range sequentially. For general DRAM, we are rapidly raising the share of 1c nanometer products for server applications, and in the fourth quarter we will further strengthen a product mix centered on high-value server products by expanding sales of 1c-based DDR5 and AI server LPDDR5 SOCAMM. Shipments of our 12-layer HBM4 are progressing smoothly on the schedule agreed with major customers. To actively respond to surging customer demand, we are accelerating the M15X ramp and plan to reach full capacity by year-end. As a result, HBM4 is expected to exceed HBM3E in both share of total HBM shipments and share of revenue in the fourth quarter, becoming our flagship product. The company has completed negotiations with major customers for next year, including HBM volume and price, and those negotiations fully reflect the enhanced product value of HBM4. We therefore expect the profitability gap between HBM and general DRAM to narrow significantly. For NAND, we plan to raise the share of our flagship 321-layer process to more than 50% of domestic capacity in the fourth quarter, expanding supply of 321-layer enterprise SSDs. Fourth-quarter NAND bit shipments are expected to grow by a mid-single-digit percentage sequentially. On that supply expansion, we will further broaden NAND's role and business opportunities in the AI infrastructure market by expanding supply of high-performance, high-density enterprise SSDs suitable for KV cache tiering, and over the medium to long term by evolving NAND into a core solution that improves AI service throughput and cost efficiency through HBF and system optimization technologies. To improve long-term demand visibility and solidify the foundation for stable growth, we continue to expand long-term agreements with customers. To date, we have signed LTAs with 14 core customers, including major global CSPs. LTAs typically run for five years, and we are actively negotiating with some customers to extend existing contract terms. Customers are requesting longer-term cooperation based on our product roadmap and supply capability, indicating that LTAs are evolving into a tool that strengthens customer relationships and business continuity. Price structures vary by customer and product but are generally designed to reflect market conditions within a price range agreed with the customer. In addition, we have included safeguards such as deposits to ensure customer procurement plans actually translate into orders and shipments. This structure both supports stable profitability and captures upside opportunities by reflecting improvements in market prices and product value. Currently, volumes locked in through LTAs account for more than 50% of our total capacity for 2027 and 2028, laying the foundation for long-term earnings stability and demand visibility. At the same time, we continue to push forward with capacity expansion investment to address medium- to long-term customer demand and growth opportunities. Because we must invest in equipment that raises the share of advanced nodes, advance the ramp of Fab 1 in Yongin and accelerate domestic and overseas infrastructure construction, we expect capital expenditure in 2027 to increase significantly from 2026. The scale and timing of investment will be decided by comprehensively considering the medium- to long-term demand visibility gained from customer communication, product competitiveness, profitability and investment efficiency.
Shareholder return plan
Last August, we announced a shareholder return plan including 40 trillion won of treasury share buybacks and cancellations, and said we would explain the size and method of additional shareholder returns at the third-quarter earnings call. Today, I will explain the progress of that plan and additional shareholder returns reflecting improvements in our operating performance and investment results. The company has decided to implement about 100 trillion won of shareholder returns for fiscal 2026: more than 30 trillion won in cash dividends and more than 60 trillion won in treasury share buybacks and cancellations, including the 40 trillion won announced in August. Of that, the 40 trillion won buyback was completed last week and will soon be canceled. Through dividends, we will directly share operating results with shareholders, while continuing to buy back and cancel shares so that per-share value appropriately reflects our intrinsic value. In doing so, we will honor our commitment to return more than 50% of free cash flow to shareholders. Today, the board resolved to implement 14 trillion won in cash dividends and 20 trillion won in share buybacks. Of that, treasury shares for executive and employee compensation are expected to total about 10 trillion won; excluding that portion, third-quarter shareholder returns come to about 24 trillion won. Let me break this down. First, we are expanding cash dividends to reflect improved performance and investment results. We will determine dividend levels based on operating performance, and when large investment returns arise, we will share that outcome additionally with shareholders through special dividends. We treat not only growth from our main business but also gains realized from the value of invested assets as sources of shareholder returns. Accordingly, the third-quarter dividend is set at 20,000 won per share, for a total of 14 trillion won. Of that, 10,000 won reflects performance improvement and 10,000 won is a special dividend sharing the Kioxia investment outcome. We plan to return most of the gains from the Kioxia investment to shareholders through special dividends: part is included in this quarter's dividend, and the remainder will be folded into the year-end dividend once fiscal 2026 results are finalized. Second, we will continue buying back shares every quarter. The size of each quarterly buyback will be decided by the board, taking into account factors such as the share price, market conditions and cash flow. This quarter we plan to buy back a total of 20 trillion won of shares, of which 10 trillion won is for shareholder return purposes and will be fully canceled; as noted, the 10 trillion won for executive and employee compensation is not counted in total fiscal 2026 shareholder returns. Going forward, based on strong memory demand growth and the company's technological competitiveness, we will steadily generate results and continue sharing them with shareholders. Thank you. Park Seong-hwan, Head of IR: We are now ready to take questions.
Q&A session
Moderator: The Q&A session will now begin. If you have a question, please press star one. Questions will be taken in the order in which star one is pressed. To cancel, please press star two on your phone. The first question comes from Peter Lee of Citi. Please go ahead.
Question one: What supports the sustainability of AI investment momentum? (Peter Lee, Citi)
Peter Lee: Congratulations on these results. My question relates to AI investment. We see AI investment continuing to grow, but monetization of AI demand or AI infrastructure buildout seems to be materializing slowly, and concerns have emerged in the market about whether AI demand can be sustained. Against that backdrop, what grounds does the company have for believing AI investment momentum will continue? Park Joon-deok, Head of DRAM Marketing: Thank you for the question. We are indeed aware of market discussion about the pace of AI investment, but we view it as part of customers more carefully examining the feasibility and profitability of individual projects and defining the timing and pace of investment. We see two main reasons why AI investment momentum will continue. First, we see that customers' AI investments are genuinely driving higher adoption of AI services and producing business results. In their latest earnings, major hyperscale customers saw cloud revenue grow 30% to 80% year on year, and we believe a considerable portion came from the expansion of their AI businesses. We also see long-term contracts for enterprise AI solutions increasing, cloud order backlogs growing twofold to fivefold year on year, and paying users rising steadily, which corroborates growth in AI-related sales and financial performance at major customers from multiple angles. Given this trend, AI sales and cash flow at major customers will only improve, which should also help ease current market concerns about the investment burden. At the same time, customers are addressing current constraints from multiple angles, such as multi-site deployment, mid-sized and smaller data centers with easier access to power, and upgrades to on-site power generation and cooling technologies, all of which should help reduce the investment burden. The second reason is that the investor base is structurally becoming more diversified. Early AI investment was concentrated among big tech companies, so an investment adjustment by a single company could directly cause overall demand volatility. Today, the investor base is rapidly expanding to include non-tech companies building their own AI infrastructure, sovereign AI and on-premise demand. This is a structural shift in which investment adjustments by some participants are offset by increased demand from others. For example, even if some projects slow due to interest rate pressure, policy-backed sovereign AI investment or broader enterprise AI adoption can fill the gap. In a multi-layered structure of diverse participants with different motivations and resources, the influence of any single company's financial capacity or investment restraint on the overall investment trend is becoming limited. In sum, supported by real business results and a more diversified set of investors, AI investment and demand intensity will remain sufficiently strong. On the memory market side, we also expect the current structural imbalance of demand exceeding supply to persist for the time being.
Question two: What is the basis for the dividend and buyback split? (Sun-woo Kim, Meritz Securities)
Sun-woo Kim: Thank you, and congratulations on the strong results. Special thanks to CEO Kwak Noh-jung for his remarks. My question concerns shareholder returns: what is the basis for determining the ratio of cash dividends to share buybacks within total shareholder returns? Kim Woohyun, CFO: Thank you very much for the question. We seek a balance between cash dividends and buybacks with cancellation, that is, between directly sharing improved performance with shareholders and enhancing per-share value by reducing the number of shares outstanding. First, we set cash dividends at a sustainable level. To do so, we comprehensively reflect performance improvement, our historical dividend yield, the dividend yields of global peers and an appropriate level of cash holdings. At the same time, when meaningful investment returns arise, namely returns on our Kioxia investment, we also plan to share them with shareholders, with most paid as a special dividend so shareholders truly feel the benefit. Although we have not closed the books for this year, we believe it is meaningful to share at least part of the predictable gains with shareholders in advance. Part of the special dividend will be paid early in the third quarter, and the remainder will be folded into the year-end dividend once the investment return is determined. After cash dividends, the remainder of total shareholder returns is allocated to share buybacks. We believe the company's share price is currently undervalued and does not fully reflect our earnings and cash flow, so we judge buybacks with cancellation to be an effective way to enhance per-share value. Buybacks will continue quarterly within the annual scale, with the size of each quarter decided by the board considering factors such as the share price, valuation and cash flow. Going forward, we will faithfully execute our shareholder return policy based on solid cash generation and continue working to enhance shareholder value.
Question three: What do LTAs mean for results and medium- to long-term growth? (Giuni Lee, Goldman Sachs)
Giuni Lee: Congratulations on the strong results. My question is about LTAs. You previously said that volumes currently locked in through LTAs account for about 50% of total capacity for 2027 and 2028. What does this mean for the company's results and longer-term incremental growth potential? Park Joon-deok: Thank you for the question. The purpose of raising LTA coverage is to improve longer-term demand visibility and thereby deliver stable revenue and earnings growth. Adding the volumes already locked in through LTAs to the demand currently under negotiation with customers, our DRAM and HBM capacity for the next three years, that is, through 2029, is in effect already sold out. The volumes under negotiation are based on customers' specific procurement plans and carry a high probability of materializing. Overall, we believe medium- to long-term sales visibility has improved significantly. As noted, committed LTA volumes account for more than 50% of total capacity for 2027 and 2028, and we plan to maintain a similar ratio over the medium to long term. This arrangement will help us secure reliable profit margins, while allowing us to use non-LTA capacity flexibly to respond to market conditions and next-generation product demand, continuously strengthening future competitiveness.
Question four: How to handle overcapacity if demand falls short? (Simon Woo, Bank of America)
Simon Woo: My question concerns possible overproduction, since there have been difficult cycles in the past. The concern about overproduction arises because, with strong market demand persisting, hynix is still building new plants or expanding and upgrading existing fabs, including the Yongin project. So if demand falls short of expectations, what is the company's strategy or management's response plan? I also want to congratulate the company on the strong results and the very detailed shareholder return plan. Kim Woohyun: Thank you for the question. The company is indeed pursuing medium- to long-term investment to meet strong demand and secure sufficient space for equipment and capacity needed for technology migration. But more investment and more fabs do not necessarily mean more supply. As the share of high-performance products such as HBM rises and processes become more complex, production requires more resources, and raising bit output through finer process improvements is becoming increasingly constrained. Our investment today is not only to expand capacity but also to build the production foundation needed for technology migration and a broader lineup of high-value products. Our investment will be executed in phases. Infrastructure for medium- to long-term demand will be prepared in advance, but the key resources that actually create supply, namely equipment move-in and mass-production ramps, will be adjusted with demand. For example, at new plants including the Yongin cluster and M17, we plan to accelerate construction of the shell structure, but cleanrooms will be brought online in phases. This approach is precisely to enable flexible operation in the future. In assessing demand, we look not only at customer capital expenditure or demand forecasts but also at customers' financial health and financing capability and the readiness of key supply chains, and reflect these in our investment plans. Our own financial soundness is also an important parameter: while balancing medium- to long-term net cash targets with an appropriate cash level and maintaining sufficient financial buffer, we will manage the scale and pace of investment based on demand visibility and investment returns. In summary, we will prepare medium- to long-term supply infrastructure as planned based on improved demand visibility, but actual capacity will only ramp in line with demand and profitability. If conditions do deviate from our expectations, we will certainly respond to risks by flexibly adjusting investment spending, production plans and resource allocation.
Question five: HBM4 supply progress and plans for HBM4E and custom HBM (Jay Kwon, JPMorgan)
Jay Kwon: Congratulations on the good results, and thank you for taking my question. You mentioned HBM4 supply earlier, and I would like an update on the specifics of HBM4 supply. Also, could you share plans for the follow-up products HBM4E and custom HBM? Kim Ki-tae, Head of HBM Sales and Marketing: Thank you very much for the question. As noted earlier, thanks to a smooth mass-production ramp, HBM4 is expected to exceed HBM3E in both bit shipment and revenue share in the fourth quarter. Yields are also quite stable. The company maintains industry-best yields in both wafer fabrication and back-end processes. In particular, what we call back-end yield refers to the entire flow including bumping, stacking and KGSD testing, not individual steps. On that basis, HBM4 has exceeded planned yields and targets have been raised, and HBM3E has also surpassed planned mature yields. Currently, we are focused on maximizing production efficiency to meet strong customer demand. Our performance in the HBM business will once again demonstrate our unmatched competitiveness. Of course, we are also steadily preparing next-generation products. For example, HBM4E completed sample shipments in the second quarter, and development and qualification are proceeding on the customer's timeline. Longer term, beyond stable supply of standard HBM, we see growing importance for custom HBM optimized for each customer's system and workload, and we are working closely with major customers on related projects. As the first supplier to major customers, we will continue to grow the HBM business based on proven product competitiveness and reliable mass-production capability.
Question six: The possibility and impact of lowering HBM stacking layers (Dong-hee Han, SK Securities)
Dong-hee Han: Thank you for taking my question, which is also about HBM, specifically the possibility of reducing HBM stacking layers. How does the company view the possibility of lowering HBM stacking layers, and what would be the potential impact on the HBM business, including shipments and profitability? Kim Ki-tae: Thank you for the question. The recent moves by some customers to adjust HBM specifications actually show that demand overwhelms supply. With HBM supply constrained, customers need to build more AI accelerators and systems with a limited volume. We see reducing stacking layers as a strategy to obtain more available units at the same capacity. However, fewer layers do not necessarily mean a proportional decline in HBM capacity. Even shifting from 12-layer HBM4 to 8-layer HBM4E, higher per-die capacity can offset most of the impact. Therefore, even considering such adjustments, our medium- to long-term HBM demand outlook is unchanged. Demand is so strong that customers are requesting more than twice the volume we can currently supply, which also reflects continued strong preference for our products. The supply side will not be affected either. Supplying the same bit volume with fewer layers requires more cubes, which in turn increases demand for base dies and back-end processing. We have incorporated these factors into our production and supply plans and will execute HBM shipments and supply as planned. As for profitability, it is not determined solely by stacking layers but by multiple factors including the terms of trade for each product, product mix, yields, back-end production efficiency and costs. Next year, the share and price of HBM4 will rise, and mass-production yields and cost competitiveness will improve further, so HBM revenue and profit are likely to grow significantly year on year, with a substantial improvement in profitability. Looking ahead, profitability is expected to remain high, well above peaks of past cycles.
Question seven: 4F square and next-generation DRAM technology preparation (Young-gun Kim, Mirae Asset Securities)
Young-gun Kim: Congratulations on the record results. My question is about 4F square. We see other companies also developing new cell structures and next-generation DRAM technologies, drawing attention to long-term technology competition. Is there any update on preparations for 4F square? In which areas is the company focusing development to strengthen technological competitiveness? Park Joon-deok: Thank you for the question. 4F square is a next-generation DRAM technology that increases density by using a smaller cell area than 6F square, and its applicability and strategy depend on each company's scaling capability, process environment and cost structure. For our company, our development plan is to make the fullest possible use of the 6F square structure through process maturity and a stable technology platform, and then transition to 4F square or a 3D DRAM platform. Through EUV-based scaling and process innovation, we already have the technical capability to further extend the life of the current 6F square platform. At present, we believe the most efficient strategy for the company is to make full use of a platform whose mass-production reliability and cost competitiveness have both been proven, thereby strengthening product competitiveness. To apply a new cell structure to products, we must not only ensure performance but also secure yields, reliability and cost competitiveness; product design and customer qualification must also be completed. This is why the timing of technology development and the timing of commercialization differ. We are preparing a phased roadmap covering everything from extreme scaling of existing DRAM to 4F square and then 3D DRAM. Each technology will be applied to the right product at the right time, meaning when both technology and economics have matured. We will continue to leverage our process and design capabilities, mass-production experience and collaboration with customers to balance performance, cost efficiency and manufacturability in next-generation DRAM and maintain technological competitiveness.
Question eight: The impact of KV cache expansion on NAND demand (Young-ho Ryu, NH Investment & Securities)
Young-ho Ryu: Thank you for taking my question, which is about NAND. We see the scale of KV caches expanding rapidly recently, and the industry is working on storing and reusing this data across multiple memory tiers. How does the company expect this to affect future NAND demand? What preparations have been made? Song Chang-seok, Head of NAND Marketing: Thank you for the question about NAND. As agentic AI and long-context use cases become widespread, KV cache sizes are expanding rapidly, and given capacity and cost constraints, it is becoming difficult to store all of it in HBM or DRAM. As a result, tiered memory architectures that distribute data across HBM, DRAM and SSD according to access frequency are becoming increasingly critical. This will create entirely new demand opportunities for NAND, to store and reuse the context data needed for AI inference. SSDs are already used in some AI systems for KV cache offload, and we believe their role will further expand into a context storage tier supporting large-scale KV caches shared across multiple servers and GPUs. To capture this opportunity, the company is developing supporting tiered storage technologies, including SALT-KV, a semantic-aware lifetime tiering technology for KV caches, and CMM-Hybrid. SALT-KV is a software technology that optimizes the allocation of data across HBM, DRAM and SSD by assessing the reuse value and storage cost of each KV cache. CMM-Hybrid is a CXL-based hybrid memory technology that stores all KV caches on SSD and then predicts which data is likely to be accessed and preloads it into DRAM. Recently, through system demonstrations based on real AI workloads, we have clarified the applicability and performance requirements of enterprise SSDs in inference scenarios, so they can be smoothly integrated into the system architectures of customers and platform partners. Over the medium to long term, we plan to use HBF, or high-bandwidth flash, to fill a new storage tier between HBM and SSD, combining high bandwidth with the large capacity and cost efficiency of NAND. By positioning early for the new NAND demand opportunities created by the evolution of AI system architectures, we will continue to grow our NAND business.
Closing remarks
Park Seong-hwan, Head of IR: Thank you to the management team for their answers. Due to time constraints, the Q&A session ends here. CEO Kwak Noh-jung will now deliver closing remarks to conclude SK hynix's third-quarter 2026 earnings call. Kwak Noh-jung, President and CEO: We have now delivered record results for six consecutive quarters. But we believe what matters more is that the underlying business environment is changing. The customer trust we have built through technology leadership is translating into long-term agreements one after another, giving us greater medium- to long-term demand visibility and building a more solid foundation for sustained profitability. This means the potential for a new growth phase, different from past cycles, is emerging through tangible changes in our business structure. Our task now is to turn these changes into durable competitive advantage. We will extend our current leadership in memory, including HBM, into next-generation products and become the partner of choice as customers plan for the future. We will translate market opportunities into concrete operating results and further solidify our ability to sustain that performance. As the company grows and investment returns materialize, the returns we share with shareholders should increase accordingly. The shareholder return plan announced today reflects that principle. Through consistent and solid performance and by sharing the fruits of growth with shareholders, we will continue to strengthen your confidence in our future growth. Thank you very much for your long-standing interest and support. Moderator: That concludes the closing remarks and also concludes this earnings call. Thank you very much.
Comments