The Zhitong Finance App learned that SK Hynix (SKHY.US) announced the results for the third quarter of 2026 and held a performance conference call.
SK Hynix's third-quarter revenue was 104.6 trillion won, up 32% month-on-month and 328% year-on-year, surpassing 100 trillion won for the first time in history; operating profit of 83.3 trillion won, up 38% month-on-month and 632% year-on-year; operating profit margin increased 3 percentage points month-on-month to 80%, and both operating profit and profit margins set new quarterly records. By product, DRAM shipments increased by about 10% month-on-month. Among them, HBM4 began full mass production and shipment in the third quarter, server DDR5 shipments increased by more than 40% month-on-month, and the average DRAM price rose to a high 20% month-on-month; NAND shipments declined slightly month-on-month due to structural adjustments, but the average price rose to the middle of 20% month-on-month. It is worth noting that net profit fell 58% month-on-month to 39.8 trillion won — management clearly attributed to non-operating projects: net loss related to exchange of 11.6 trillion yuan, loss of valuation of investment assets of 22.9 trillion yuan, total net non-operating expenses of 32.6 trillion won, unrelated to the operating level. Looking ahead, the company expects global CSP capital expenditure to increase by more than 60% in 2027, the locked power capacity will increase by more than 100% year over year, and the global AI data center infrastructure market will maintain an annual growth rate of about 40% until 2030; however, due to supply capacity constraints, the growth in demand that can be achieved in 2027 is similar to this year, with DRAM in the middle of 20% and NAND in the high ranking of ten percentage points. The fourth quarter guidance is a 20% month-on-month increase in DRAM shipments, and a single-digit increase for NAND in the middle. In terms of capital expenditure, the company said that 2027 will increase significantly compared to 2026, investing in advanced process equipment, Yongin Plant 1 climbing, and domestic and foreign infrastructure, but did not give specific figures.
Key content of the analyst Q&A: The eight analysts' questions focused on four main lines. First, the sustainability of AI demand: Park Joon-duk, the head of DRAM marketing, gave two grounds — the latest earnings report of major hyperscale customers increased 30% to 80% year over year, cloud order backlog increased by 2 to 5 times year on year, and the steady increase in paying users, indicating that investments are producing actual business results; at the same time, investors have spread from a few large technology companies to non-tech companies to self-built, sovereign AI and local deployment, and individual companies' investment adjustments can be hedged against other needs. Second, long-term supply agreements (LTAs): The company has signed LTAs with 14 core customers, including major global CSPs, with a target volume of more than 50% of the total production capacity in 2027 and 2028; he further stated that the amount already locked plus the demand being negotiated means that DRAM and HBM production capacity by 2029 is actually almost sold out, but with additional restrictions — the amount under negotiation “is highly likely to be realized based on the customer's specific procurement plan.” Third, risk of overcapacity: CFO Kim Woo-hyeon responded that “more investment, more plants do not equal more supply,” and supply discipline falls on the pace of engineering — new plants such as Yongin Cluster and M17 first speed up plant shell construction, clean rooms are being opened in stages, and equipment moving in and mass production climbing follows demand; when evaluating demand, it also depends on the customer's financial health, financing capacity, and key supply chain preparations. Fourth, the reduction in the number of layers in the HBM stack: HBM sales leader Kim Ki-tae characterized it as evidence of overheating demand, saying that the amount currently required by customers exceeds more than twice the company's supply. When switching from a 12-tier HBM4 to an 8-tier HBM4E, the increase in single die capacity can offset most of the impact, and supply and profit plans are not affected.
The focus of market attention. The real incremental information in the market is not 80% profit margin, but at two points. The first is that the phrase “sold out by 2029” goes further than the official press release — the press release only wrote that LTA covered more than 50% of production capacity from 2027 to 2028, and only pushed the caliber to 2029 through a conference call, and accurately stated that “LTA has already been locked out+demand under discussion” sold out in total, not that it has already been contracted out. Second, Solidigm has withdrawn its US listing plan: CEO Guo Luzheng explained at the beginning that Solidigm originally wanted to introduce external capital and reduce the investment burden through the IPO, but after evaluating recent financial market fluctuations and semiconductor stock price revisions, he believed that the current IPO could not fully achieve the expected benefits, so the review of pre-IPO and IPO options was suspended. This is accompanied by shareholder returns of approximately 100 trillion won for the 2026 fiscal year: cash dividends of 30 trillion dollars or more, repurchases and cancellations of more than 60 trillion dollars (including 40 trillion dollars announced in August, completed and cancelled last week); on the same day, the board of directors also decided on a cash dividend of 14 trillion yuan and a repurchase of 20 trillion dollars for the third quarter, of which about 10 trillion dollars was used for executive and employee remuneration, not included in shareholder returns, so the actual return for the third quarter was about 24 trillion yuan. The dividend of 20,000 won per share consists of a performance improvement dividend of 10,000 won and a special dividend on Kioxia investment income of 10,000 won. The rest will be incorporated into the year-end dividend after the fiscal year is settled.
Presentation session
Moderator: Good morning, thank you all for attending the SK Hynix results conference call. Today, we will first present our performance and then move on to the Q&A session.
If you have problems, press the star button (that is, an asterisk and 1). Please note that the presentation session is interpreted simultaneously, and the question and answer session is interpreted alternately.
So, let's get started.
Park Sung-hwan (IR Director): Good morning, good noon, good evening. I'm Park Sung-hwan, Head of Investor Relations at SK Hynix.
Welcome to SK Hynix's Q3 2026 results conference call. Let me introduce the executives who are attending with me today.
In attendance were President and CEO Kwak Lu-jeong, CFO Kim Woo-hyun, DRAM Marketing Director Park Joon-duk, NAND Marketing Director Song Chang-seok, and HBM Sales and Marketing Director Kim Ki-tae. I would like to disclaimer: The third quarter results included in this conference call are provisional data on a consolidated basis, and the external auditor's review has not yet been completed.
As a result, the relevant data is still subject to change. Furthermore, forward-looking statements, including market prospects and company plans, may change due to changes in the macroeconomic and market environment.
So, let's start the SK Hynix results conference call for the third quarter of 2026. Before starting the performance statement, CEO Guo Luzheng will first explain Solidigm's IPO and shareholder returns.
CFO Kim Woo-hyeon will then present the third quarter results, the company's future plans, and market outlook.
CEO: Solidigm terminates listing evaluation, Kioxia earnings returned with a special dividend
Guo Luzheng (President and CEO): Hello everyone, I'm Guo Luzheng, CEO.
I would like to express my sincere thanks to all shareholders for their continued concern and trust in SK Hynix even as market fluctuations have increased recently. Today, I would like to directly explain two things to you: first, our decision on the Solidigm IPO — which has recently attracted considerable attention in the market; and a special dividend to share the return on strategic investments with all shareholders.
First, Solidigm has been exploring various financing options to support initiatives such as expanding manufacturing layout and meeting growing demand. Among the options considered, bringing in external capital through an IPO was expected to ease our investment burden and enhance our ability to deliver shareholder returns.
However, after carefully evaluating recent financial market fluctuations and semiconductor stock price revisions, Solidigm concluded that the IPO was unable to fully achieve the expected returns at this time, so it decided to discontinue the review of potential pre-IPO and IPO options. Solidigm will continue to strengthen its competitiveness and seek growth opportunities in the AI data center storage market. Investment and financing arrangements to support these efforts will proceed in a way that simultaneously enhances the corporate value and shareholder value of Solidigm and SK Hynix.
We have always adhered to one principle: share most of the return on strategic investments with shareholders. Using these returns to supplement our shareholder return capacity — which may temporarily decline as we invest in the future — our goal is to protect shareholder interests while continuing to boldly invest in sustainable growth.
In accordance with this principle, we intend to return the profits achieved from Kioxia (Kioxia) investments this year to shareholders through a special dividend. We are committed to fulfilling this commitment to ensure that our shareholders directly benefit from the success of our strategic investments.
Next, the CFO will present the quarter's results and outlook, as well as details of the shareholder return plan. Thank you.
CFO: Third Quarter Results
Kim Woo-hyeon (CFO): Hi everyone, I'm CFO Kim Woo-hyun.
Let me first introduce the company's results for the third quarter of 2026. In the third quarter, driven by increased demand for AI storage, supply and demand continued to be tight, and storage prices continued to rise. DRAM and NAND recorded high gains for the second consecutive quarter, with price increases focused on HBM, server DRAM, and enterprise-grade SSDs.
As a result, third-quarter revenue increased 32% month-on-month and 328% year-on-year, reaching 104.6 trillion won. This was the first time that quarterly revenue surpassed 100 trillion won, a record high.
In terms of DRAM, we are expanding sales with AI servers as the center, and our shipments have increased by about 10% month-on-month, which is in line with the plan. In particular, it should be pointed out that HBM4 began full mass production and shipment in the third quarter; in addition, server DDR5 shipments increased by more than 40% compared to the second quarter, leading to an overall increase in DRAM shipments. Due to the combination of the continued strength of GM DRAM prices and the increase in sales share of HBM4, which is higher than the previous generation, the average price of DRAM rose to a high level of 20% (high -20%) month-on-month (high -20%), an increase similar to that of the previous quarter.
On the NAND side, since we adjusted our sales focus to high value-added products such as 321-tier TLC enterprise-grade SSDs, the average price dropped slightly compared to the second quarter, but driven by stronger overall prices and the expansion of the SSD product structure, the average price rose to the middle of 20% (mid -20%) month-on-month (mid -20%).
Despite the negative impact of the exchange rate due to the strengthening of the won, the increase in product prices and the expansion of sales of high value-added products led to operating profit of 83.3 trillion won in the third quarter, up 38% month-on-month and 632% year-on-year. The operating margin also increased by 3 percentage points to 80% month-on-month, and both operating profit and operating margin reached record quarterly highs.
The total depreciation and amortization expenses for the third quarter were 4.3 trillion won; EBITDA reached 87.7 trillion won, and the EBITDA profit margin was 84%.
In terms of major non-operating projects, net exchange-related losses of 11.6 trillion won due to exchange rate changes, investment asset valuation losses of 22.9 trillion won, and total net non-operating expenses of 32.6 trillion won. As a result, profit before tax was 50.7 trillion won, net profit 39.8 trillion won, and a net interest rate of 38%.
By the end of the third quarter, cash and cash equivalents were 114.5 trillion won, up 26.5 trillion won from the end of the previous quarter; interest-bearing liabilities decreased 2.8 trillion won to 15.8 trillion won. As a result, net cash increased to 98.7 trillion won, and liabilities improved by 2 percentage points to 5% compared to the end of the previous quarter.
Market outlook
Next, I'll talk about the market outlook. The AI market is constantly evolving.
As AI brings productivity improvements and new revenue generation in various fields such as software development, customer response, and search, a virtuous cycle is being formed: AI usage is increasing, which in turn is driving the expansion of infrastructure investment. This change is reflected in the scale of investment of the world's major cloud service providers and their locked power capacity.
According to recent estimates by market research institutes, global CSP capital expenditure is expected to increase by more than 60% in 2027, and locked electricity capacity is also expected to increase by more than 100% year over year. More importantly, this growth won't stop next year — the global AI data center infrastructure market is expected to maintain a high annual growth rate of around 40% by 2030. Driven by the commercialization of AI services and the expansion of application fields, this medium- to long-term growth trend is expected to continue.
In particular, as AI begins planning complex tasks and uses various tools to execute them over a longer period of time, the demand for AI computing power and the amount of data generated, transmitted, and stored is increasing dramatically. As a result, demand across the storage sector grew significantly beyond initial expectations.
AI infrastructure investment has now gone beyond the early stages of relying on a few companies to invest first, and participation is expanding to more companies and national governments. As a result, we expect AI infrastructure investment growth to remain strong even with some degree of investment adjustment.
In the short term, enterprise-level and personal AI will lead to demand growth; in the medium to long term, physical AI such as autonomous driving and robots is expected to broaden the scope of demand. Although PC and mobile device shipments are expected to decline due to cost pressure from rising storage prices and supply shortages, the increase in demand for AI infrastructure is expected to more than offset this decline, thereby maintaining the overall growth of the storage market.
On the other hand, the storage supply cannot keep up with this level of demand growth, so we expect it will be difficult to fully meet customer volume requirements in the short term. Given the difficulty of implementing advanced processes and the cycle required to build new production facilities, the situation where supply cannot keep up with demand is expected to continue for some time, and this constraint is expected to continue until 2030.
Despite the high level of demand, considering actual supply conditions, the growth in demand that can be achieved in 2027 is expected to be similar to this year: DRAM is 20% mid-range (mid -20%), and NAND is ten percentage points higher (high-teens%). We expect this favorable market environment to be a solid foundation to support the company's performance growth next year. Based on expanded production capacity and high value-added product competitiveness, we plan to achieve significantly higher results in absolute terms of revenue and profit.
Fourth quarter sales plan and business direction
In the fourth quarter, we plan to further accelerate supply expansion to respond more proactively to market demand.
DRAM: The increase in production brought about by the expansion of the 1c nano process will occur simultaneously with the increase in HBM4 production brought about by the M15X slope, so shipments in the fourth quarter are expected to increase significantly by 20% mid-quarter (mid -20%) compared to the previous quarter. In terms of general purpose DRAM, we are rapidly increasing the proportion of 1c nm products for server applications. In the fourth quarter, we will expand sales of the 1c nanomic-based DDR5 and AI server LPDDR5 SOCAMM to further strengthen our product portfolio centered on high-value-added server products.
Shipments of our 12-tier HBM4 are progressing smoothly according to the schedule agreed with our major customers. To respond positively to surging customer demand, we are speeding up the M15X climb and plan to reach full production before the end of the year. As a result, HBM4 is expected to surpass HBM3E in terms of total HBM shipments and revenue in the fourth quarter, becoming our flagship product. The company has completed next year's negotiations with major customers, including the volume and price of HBM. The relevant negotiations faithfully reflect the increased product value of HBM4. As a result, we expect the profitability gap between HBM and General Purpose DRAM to close significantly.
NAND: We plan to increase the share of the flagship 321 layer process to more than 50% of domestic production capacity in Korea in the fourth quarter, thereby expanding the supply of 321 tier enterprise SSDs. In the fourth quarter, NAND shipments are expected to increase by mid-single digits (mid-single%) month-on-month. Based on this supply expansion, we will further expand NAND's role and business opportunities in the AI infrastructure market — expand the supply of high-performance, high-density enterprise-grade SSDs suitable for KV cache tiering; in the medium to long term, evolve NAND into a core solution for improving AI service throughput and cost efficiency through HBF and system optimization technology.
To increase the visibility of medium- to long-term demand and solidify the foundation for steady growth, we are continuing to expand our long-term agreements with our customers. To date, we have signed LTAs with 14 core customers, including major CSPs around the world. LTAs usually last for five years, and we are actively negotiating with some of our clients to extend the term of existing contracts. Customers require longer-term partnerships based on our product roadmap and supply capacity, which shows that LTA is evolving into a tool to strengthen customer relationships and business continuity.
The price structure varies from customer to customer and product, but is generally designed to reflect market conditions within the price range agreed with the customer. Furthermore, in order to ensure that the customer's purchase plan is actually converted into an order and shipment, we have set up guarantee terms such as a deposit. This structure not only supports our stable profitability, but also captures opportunities for market improvement by reflecting improvements in market prices and product values. Currently, LTA has locked in more than 50% of our total production capacity in 2027 and 2028, laying the foundation for medium- to long-term profit stability and demand visibility.
At the same time, in order to meet medium- to long-term customer needs and growth opportunities, we are continuing to invest in production expansion. We expect capital expenditure in 2027 to increase significantly compared to 2026 due to investment in equipment to increase the proportion of advanced manufacturing processes, promote the climbing of Yongin Plant 1, and accelerate infrastructure construction at home and abroad. The investment scale and execution time are determined by comprehensive customer communication with medium- to long-term demand visibility, product competitiveness, profitability and investment efficiency.
Shareholder return plan
In August of last year (note: August 2026), we announced a shareholder return plan including the repurchase and cancellation of 40 trillion won of treasury shares, and stated that we will explain the scale and method of additional shareholder returns during the third quarter results conference call. Today, I will explain the progress of implementation of this plan, as well as additional shareholder returns reflecting improvements in our operating performance and investment results.
The company decided to implement shareholder returns of approximately 100 trillion won in fiscal year 2026: cash dividends of 30 trillion won or more, and treasury stock repurchases and cancellation of 60 trillion won or more (including 40 trillion won announced in August). Of these, the 40 trillion won treasury share repurchase was completed last week and will soon be cancelled.
Through dividends, we will directly share operating results with shareholders; at the same time, we will continue to repurchase and cancel shares, so that the value of each share can properly reflect our intrinsic value. With this, we will fulfill our promise to return more than 50% of our free cash flow to shareholders.
Today, the Board of Directors resolved to implement cash dividends of 14 trillion won and share repurchases of 20 trillion won. Among them, treasury stock used for executive and employee compensation is estimated to be approximately 10 trillion won; excluding this portion, shareholder returns for the third quarter are approximately 24 trillion won.
I'll explain them in detail below.
First, we are expanding our cash dividends to reflect improved performance and investment results. We will determine the level of dividends based on business performance; when there is a large return on investment, we will additionally share this part of the results with shareholders through a special dividend. We use not only the growth results of our main business, but also the results of realizing the value of our invested products as a source of shareholder return.
According to this, the third quarter dividend was set at 20,000 won per share, for a total amount of 14 trillion won. Among them, 10,000 won reflects improved performance, and 10,000 won is a special dividend for sharing the results of Kioxia's investment. We plan to return most of the proceeds from Kioxia Investments to shareholders through a special dividend: part of this quarter's dividend will be included in the current quarter's dividend, and the rest will be incorporated into the year-end dividend after the 2026 fiscal year results are determined.
Second, we will continue to repurchase shares every quarter. The size of a quarterly repurchase will be decided by the board of directors, taking into account factors such as stock prices, market conditions, and cash flow. We plan to repurchase a total of 20 trillion won of shares this quarter, of which 10 trillion won will be used for shareholder return purposes and will be fully cancelled; as mentioned above, 10 trillion won for executive and employee compensation will not be included in the total shareholder return for the 2026 fiscal year.
In the future, based on strong growth in storage demand and the company's technical competitiveness, we will steadily create results and continue to share results with shareholders. Thank you.
Park Sung-hwan (Head of IR): Well, now we're ready to take questions.
Q & A session
Moderator: The question and answer session will now begin. If you have problems, press the star button (that is, an asterisk and 1).
The questions will be asked in the order of each person with an asterisk 1. To cancel, press the asterisk two (that is, an asterisk and 2) on the phone.
The first question comes from Citigroup's Peter Lee, please ask.
Question 1: What is the basis for continued AI investment momentum? (Peter Lee, Citigroup)
Peter Lee: Congratulations on this achievement. My question relates to investing in AI. We have seen investment in AI continue to grow, but the demand for AI or monetization after AI infrastructure construction seems to be slow to materialize, and the market is worried about whether AI demand will continue as a result. In this context, what is the basis for the company's belief that AI investment momentum will continue?
Park Joon-duk (Head of DRAM Marketing): Thank you for asking. We do notice discussions in the market about the pace of AI investments, but we see this as part of a process where customers take a closer look at the viability and profitability of individual projects and define when and at what pace to invest.
We believe AI investment momentum will continue for two main reasons.
First, we're seeing that customers' AI investments are actually leading to increased adoption of AI services and generating business results. In the latest financial report, cloud revenue for major hyperscale customers increased 30% to 80% year over year, and we think a significant portion of this is due to the expansion of their AI business. We are also seeing an increase in long-term contracts for AI solutions for large enterprises, while the cloud order backlog is growing 2 to 5 times year over year, and paying users are also growing steadily — this confirms the growth in AI-related sales and financial performance of major customers in many ways.
Given this trend, AI sales and cash flow for key customers will only improve, which will help ease current market concerns about investment burdens. At the same time, customers are addressing current constraints from multiple perspectives, such as multi-site deployments, small to medium data centers with easier access to power, and upgrading on-site power generation and cooling technology, all of which will help reduce the investment burden.
The second reason is that investors are becoming more structurally diversified. Early AI investments were concentrated in big tech companies, and the adjustment of investments by a single company could directly cause fluctuations in overall demand. Today, investors are rapidly expanding to include non-tech companies investing in self-built AI infrastructure, sovereign AI, and local deployment requirements. This is a structural shift: investment adjustments by some players will be offset by increased demand from others.
For example, even if some projects slow down due to interest rate pressure, policy-endorsed sovereign AI investments or broader enterprise AI adoption can fill this gap. In a multi-tiered structure comprised of diverse players with diverse motivations and resources, the financial capacity or investment restraint of any single company is becoming limited in its impact on overall investment trends.
Overall, with real business results and more diversified investment support, AI investment and demand intensity will remain strong enough. In terms of the storage market, we also expect that the current structural imbalance where demand exceeds supply will continue for the time being.
Question 2: What is the basis for the ratio of dividends and repurchases? (Sun-woo Kim, Merritz Securities)
Sun-woo Kim: Thank you and congratulations to the company for its excellent results. Special thanks to CEO Guo Luzheng for his introduction. My question relates to shareholder returns: What is the basis for the company to determine the ratio of cash dividends to share repurchases in total shareholder returns?
Kim Woo-hyeon (CFO): Thank you so much for the question. We seek to balance cash dividends with share buybacks and cancellations, that is, between “directly sharing improved performance with shareholders” and “increasing the value per share by reducing the number of tradable shares.”
First, we set the cash dividend at a sustainable level. To this end, we are comprehensively reflecting improved performance, our historical dividend rates, global peer dividend rates, and appropriate levels of cash holding. At the same time, when there is a meaningful return on investment, that is, our return on investment in Kioxia, we also plan to share it with shareholders — most of which will be distributed in the form of special dividends, so that shareholders can truly feel it. Although we haven't settled our accounts this year, we think it makes sense to share at least part of the expected earnings with shareholders in advance. Part of the special dividend will be paid early in the third quarter, and the rest will be incorporated into the year-end dividend after the return on investment is determined.
After deducting cash dividends, the remainder of the total shareholder return is allocated to share repurchases. We believe that the company's stock price is currently undervalued and does not fully reflect our profits and cash flow, so we determine that buybacks and cancellations are an effective way to increase the value per share. Buybacks will continue to be carried out every quarter on an annual scale. The size of a single quarter will be determined by the board of directors taking into account factors such as stock price, valuation, and cash flow.
In the future, we will faithfully implement shareholder return policies based on our sound ability to generate cash, and continue to work to enhance shareholder value.
Question 3: What does LTA mean for performance and medium- to long-term growth? (Giuni Lee, Goldman Sachs)
Giuni Lee: Congratulations on the excellent results. My question is about LTA. As previously stated, the volume currently locked in through LTA accounts for about 50% of the total production capacity in 2027 and 2028. What does this mean for the company's performance and the potential for incremental growth in the longer term?
Park Joon-duk: Thank you for asking. The purpose of increasing LTA coverage is to increase longer-term visibility of demand, thereby bringing about stable revenue and profit growth.
If you add up the amount already locked in through LTA, plus the demand currently being negotiated with customers, we have actually sold out of DRAM and HBM production capacity for the next three years — that is, until 2029. The volume under negotiation is based on the customer's specific procurement plan and is likely to be realized. Overall, we think medium- to long-term sales visibility has improved significantly.
As mentioned before, LTA has promised more than 50% of total production capacity in 2027 and 2028, and we plan to maintain a similar ratio in the medium to long term. This arrangement will help us obtain reliable profit margins; at the same time, we can use non-LTA production capacity to respond flexibly to market conditions and next-generation product needs, and continue to strengthen our future competitiveness.
Question 4: If demand falls short of expectations, how to deal with overcapacity? (Simon Woo, Bank of America)
Simon Woo: My question relates to possible overproduction because there have been some tough cycles in the past. The reason why there are concerns about overproduction is that while strong market demand continues, Hynix is still building new plants or expanding and upgrading existing plants, including Yongin projects. So if demand falls short of expectations, what is the company's strategy or management's response plan? We would also like to congratulate the company for its excellent results and a very detailed shareholder return plan.
Kim Woo-hyun: Thank you for asking. The company is actually currently promoting medium- to long-term investments to meet strong demand and allocate sufficient space for the equipment and production capacity required for technology migration. But more investment and more factories don't necessarily mean more supply.
As the proportion of high-performance products such as HBM increases, the process becomes more complex, and more resources are required for production, increasing bit output through finer processes is becoming more and more limited. Our investment today is not only to expand production capacity, but also to build the production base needed for technology migration and expansion of high-value-added product lineups.
Our investment will be executed in stages. Infrastructure for medium- to long-term needs will be prepared in advance, but the key resources required to actually form supply — equipment moving in and mass production climbing — will adjust according to demand. For example, for the new plant, which includes the Yongin Cluster and M17, we plan to speed up the construction of the plant shell structure, but the clean room will be opened in stages. This approach is intended to be able to operate flexibly in the future.
When evaluating demand, we not only look at the customer's capital expenditure or demand forecast, but also examine the customer's financial health and financing ability, as well as the degree of preparation of key supply chains, and then reflect this in investment plans. Our financial soundness is also an important parameter: while balancing medium- to long-term net cash targets with appropriate cash levels and maintaining sufficient financial balances, we will manage the scale and pace of investment based on demand visibility and return on investment.
In summary, we will prepare medium- to long-term supply infrastructure according to the plan based on improved demand visibility, but actual production capacity will only climb on the premise that it meets demand and profitability. If the situation does deviate from our expectations, we will definitely respond to risks by flexibly adjusting investment expenses, production plans, and resource allocation.
Question 5: HBM4 supply progress and HBM4E, custom HBM plans (Jay Kwon, J.P. Morgan)
Jay Kwon: Congratulations on the great results and thank you for taking my questions. HBM4 supply was mentioned earlier, and I would like to ask the company to update the details of HBM4 supply. Also, can you share plans for the successor product HBM4E and custom HBM?
Kim Ki-tae (Head of Sales and Marketing at HBM): Thank you so much for asking. As mentioned earlier, thanks to the smooth progress of mass production, HBM4 is expected to surpass HBM3E in terms of both bit shipments and revenue share in the fourth quarter.
The yield is also fairly stable. The company maintains the best level of yield in the industry at both ends of wafer manufacturing and post-processing. In particular, what we mean by post-stage yield refers to the entire process, including bump, stacking, and KGSD testing, rather than individual steps. On this basis, HBM4 has exceeded the planned yield, and the target has been raised; HBM3E has also exceeded the planned mature yield.
Currently, we focus on maximizing production efficiency to meet strong customer demands. Our performance in the HBM business will once again prove our unrivaled competitiveness.
Of course, we are also steadily preparing the next generation of products. For example, the HBM4E completed sample shipments in the second quarter, and development and certification are progressing according to the customer's schedule. In the longer term, in addition to stabilizing the supply standard HBM, we are seeing a rise in the importance of customized HBM optimized for each customer's systems and workloads, and we are working closely with key customers on related projects. As the first supplier for major customers, we will continue to expand the HBM business by relying on proven product competitiveness and reliable mass production capacity.
Question 6: The possibility and impact of the reduction in the number of HBM stacks (Dong-hee Han, SK Securities)
Dong-hee Han: Thanks for taking the question. The question is also about HBM, and in particular the possibility of reducing the number of HBM stacks. How does the company view the possibility of reducing the number of HBM stacks and what is the potential impact on the HBM business (including shipping and profitability)?
Kim Ki-tae: Thank you for asking. The recent move by some customers to adjust HBM specifications just shows that demand is overwhelming supply. With the limited supply of HBM, customers need to use a limited amount to build more AI accelerators and systems. We believe that reducing the number of stacked layers is a strategy to get more units at the same capacity.
However, a reduction in the number of layers does not necessarily mean a proportional decrease in HBM capacity. Even with the shift from the 12-tier HBM4 to the 8-tier HBM4E, the increase in single die capacity can offset much of the impact. Therefore, even considering these adjustments, our medium- to long-term demand outlook for HBM remains unchanged. Strong demand has caused customers to demand more than double our current supply, which also reflects our continued strong preference for our products.
The supply side will also not be affected. Supplying the same amount of bits at a lower number of layers requires more cubes, which in turn increases the demand for basic die and post-processing. We have incorporated these factors into our production and supply plans, and will execute HBM shipment and supply as planned.
As for profitability, it is determined not only by the number of layers stacked, but by multiple factors such as trade conditions, product mix, yield, and back-end production efficiency and cost of each product. Next year, the share and price of HBM4 will increase, and mass production yield and cost competitiveness will be further improved. Therefore, HBM's revenue and profit are likely to increase significantly year-on-year, and profitability will also improve substantially. Looking ahead, profitability is expected to remain high, far above the peak of previous cycles.
Question 7:4F square and preparation for next-generation DRAM technology (Young-gun Kim, Future Asset Securities)
Young-gun Kim: Congratulations on the record results. My question is about 4F square. We have seen that other companies are also developing new unit structures and next-generation DRAM technology, drawing attention to long-term technology competition. Are there any updates on preparations for 4F Square? In order to strengthen technological competitiveness, what areas does the company focus on development?
Park Joon-duk: Thank you for asking. 4F square is a next-generation DRAM technology that increases density with a smaller unit area than 6F square. Its applicability and strategy depend on each company's miniaturization capabilities, process environment, and cost structure.
For our company, our development plan is to make the most of the 6F square structure through process maturity and a stable technology platform, and then transition to a 4F square or 3D DRAM platform. Through EUV-based miniaturization and process innovation, we have the technical ability to further extend the life cycle of the current 6F square platform. Currently, we believe that the most efficient strategy for the company is to make full use of a platform where mass production reliability and cost competitiveness have been verified to strengthen product competitiveness.
To apply the new unit structure to the product, it is necessary not only to ensure performance, but also to have yield, reliability and cost competitiveness; in addition, product design and customer certification must also be completed. This is the reason why there is a difference between when technology is developed and when it is commercialized.
We are preparing a phased roadmap covering everything from extreme miniaturization of existing DRAM, to 4F square, to 3D DRAM. Each technology will be applied to the right product at the right time — the so-called right time, that is, when both technology and economy are mature. We will continue to rely on process and design capabilities, mass production experience, and collaboration with customers to balance performance, cost efficiency, and mass producibility in next-generation DRAM to maintain technical competitiveness.
Question 8: The impact of KV cache expansion on NAND demand (Young-ho Ryu, NH Investment Securities)
Young-ho Ryu: Thanks for taking the question, the question is about NAND. We have seen the recent rapid expansion of KV caches, and the industry is also promoting the storage and reuse of this data across multiple storage tiers. How does the company anticipate this will impact future NAND demand? What preparations have been made?
Song Chang-seok (Head of Marketing for NAND): Thank you for asking about NAND. With the popularity of agent-based AI and long-context application scenarios, KV caches are rapidly expanding in size, and are constrained by capacity and cost, making it difficult to store them all in HBM or DRAM. As a result, hierarchical memory architectures that distribute data across HBM, DRAM, and SSD according to data access frequency are becoming increasingly critical.
This will open up new demand opportunities for NAND to store and reuse contextual data needed for AI inference. SSDs are already being used for KV cache offload in some AI systems; we think their role will be further extended to a contextual storage layer that supports multiple servers and multiple GPUs to share large-scale KV caches.
In response to this opportunity, the company is developing supporting tiered storage technologies, including SALT-KV, semantic-aware lifecycle tiering technology for KV caches, and CMM-Hybrid. SALT-KV is a suite of software technologies that optimizes the allocation of data between HBM, DRAM, and SSD by evaluating the reuse value and storage cost of each KV cache. CMM-Hybrid is a CXL-based hybrid memory technology that stores all of the KV cache on an SSD, then predicts what data might be accessed and preloads it to DRAM.
Recently, through system demonstrations based on real AI workloads, we have clarified the suitability and performance requirements of enterprise-grade SSDs in inference scenarios so that they can be smoothly integrated into the system architectures of customers and platform partners.
In the medium to long term, we plan to fill the new storage tier between HBM and SSD with HBF (High Bandwidth Flash) — it combines high bandwidth and NAND's high capacity and cost efficiency advantages. We will continue to maintain the growth of the NAND business by laying out the new demand opportunities for NAND brought about by the evolution of AI system architectures in advance.
Closing Remarks
Park Sung-hwan (Head of IR): Thank you to the management for your answers. Due to time limitations, this concludes the Q&A session. Next, CEO Guo Luzheng delivered a closing speech to conclude this SK Hynix results call for the third quarter of 2026.
Guo Luzheng (President and CEO): So far, we have delivered record results for six consecutive quarters. But what we think is more important is that the underlying business environment is changing.
The customer trust we have built through technology leadership is being transformed into a long-term agreement, bringing us greater visibility into medium- to long-term needs and building a more solid foundation for continued profit. This means that the potential for a new phase of growth that is different from previous cycles is being revealed through tangible changes in the business structure.
Our task now is to turn these changes into a lasting competitive advantage. We will continue our current leadership in the storage field, including HBM, to the next generation of products, and become the partner of choice for customers as they lay out the future. We will transform market opportunities into tangible business results and further strengthen our ability to continue to maintain this performance.
As the company's business grows and the return on investment becomes apparent, the return we share with shareholders should increase accordingly. The shareholder return plan announced today is an expression of this principle. We will continue to enhance everyone's confidence in our future growth through continuous and steady performance and sharing development results with shareholders.
Thank you very much for your continued attention and support.
Moderator: Closing remarks have come to an end, and the results conference call has come to an end. Thank you all so much.