SK hynix (SKHY) Q3 2025 Earnings Call Transcript

Date: October 29, 2025 | Source: StockAnalysis / SK hynix IR

---

Park Seong-hwan (Head of IR, SK hynix): Good morning, afternoon, and evening. This is Park Seong-Hwan, Head of IR at SK hynix. Welcome to the SK hynix 2025 third quarter earnings release conference call. Allow me to introduce the executives present here with me today. We're joined by Chief Financial Officer Kim Woo-Hyun, Head of DRAM Marketing Kim Kyu-Hyun, Head of NAND Marketing Kim Seok, and Head of HBM Sales and Marketing Kim Ki-tae. Let me issue a disclaimer that all outlooks presented by the company are subject to change depending on the macroeconomic and market circumstances. With that, we will now begin the SK hynix earnings release conference call for the third quarter of 2025. CFO Mr. Kim Woo-Hyun will first present the earnings, followed by the company's future plans and market outlook, and a Q&A session with the attending executives.

Kim Woo-hyun (CFO, SK hynix): Good morning, everyone. Allow me to first introduce the company's performance for the third quarter of 2025. Earlier in the year, we expected more moderate demand conditions in Q3 due to external uncertainties and the impact of some preemptive purchases in the first half. However, we ended up witnessing a highly favorable market environment with a spike in demand for memory products for servers, including HBM, driven by surging AI infrastructure investments by Big Tech companies. Third quarter revenue again recorded record quarterly revenue of KRW 24.4 trillion, up 10% QoQ and 39% YoY. This was driven by stronger DRAM and NAND pricing, as well as an increase in DRAM shipments from rising demand.

DRAM's bit shipments exceeded guidance by increasing high single digits sequentially, driven by growing sales of HBM3E 12-high products and server DDR5 to support AI demand, as well as seasonal demand recovery for LPDDR5 products. In particular, shipments of high-density DDR5 modules of over 128 GB doubled QoQ for two quarters in a row, clearly demonstrating robust growth in HPC-related DRAM demand. ASP rose by mid single digit QoQ, with strong ASP growth for conventional DRAM products. For NAND, bit shipments decreased by mid single digit QoQ, given the high base from the previous quarter, but enterprise SSD shipments grew by double digits amid rising demand from AI servers. ASP increased by low 10% compared to the previous quarter, supported by NAND price recovery and a higher mix of enterprise SSDs with pricing premium.

Operating profit reached KRW 11.4 trillion, up 24% QoQ and 62% YoY, also marking an all-time high. Operating margin improved by five percentage points QoQ and seven percentage points YoY to 47%, driven by strong sales of leading-edge products such as HBM, high-performance DRAM, and enterprise SSD. This marks the first time in the company's history that quarterly operating profit has exceeded KRW 10 trillion. Depreciation and amortization expenses in Q3 were KRW 3.6 trillion, resulting in EBITDA of KRW 14.9 trillion and an EBITDA margin of 61%. Non-operating income net of expenses was KRW 3.4 trillion, including KRW 0.21 trillion of foreign-currency-related net gain due to the stronger U.S. dollar at the end of the quarter, and KRW 3.3 trillion of valuation gains on investment assets. Pre-tax income was KRW 14.8 trillion, net income was KRW 12.6 trillion, and net profit margin stood at 52%, again reaching a record-high level.

At the end of Q3, cash and cash equivalents stood at KRW 27.9 trillion, up KRW 10.9 trillion from last quarter. Interest-bearing debt increased by KRW 2.2 trillion to KRW 24.1 trillion, resulting in a net cash position of KRW 3.8 trillion. Accordingly, the debt-to-equity ratio improved by one percentage point QoQ to 24%.

Now, let me share our market outlook. In 2025, despite ongoing geopolitical and macroeconomic uncertainties such as tariffs, the memory market saw mixed expectations — optimism regarding explosive AI growth alongside concerns about monetization. Recently, however, global investments in AI infrastructure have become the top priority for AI market expansion, driving significant demand growth not only for HBM but for broader memory demand such as DRAM for general-purpose servers and enterprise SSDs. The AI market is now shifting rapidly from the training phase of large models to the inference phase where users actively utilize AI services. The evolution toward inference requires handling vast numbers of tokens with no latency. KV cache, or Key-Value cache, which is intermediate computation results generated during inference, grows proportionally with context length. When HBM alone cannot store all of this data, it is offloaded sequentially to conventional DRAM and SSDs. As a result, expansion of the AI inference market is driving demand not only for HBM and high-performance DDR5, but also for enterprise SSDs, signaling a structural shift in both DRAM and NAND demand.

Leading AI companies are now accelerating investments backed by monetization and forming strategic partnerships to support this growth. This trend will lead to further expansion of AI data centers, creating robust demand across a wide range of memory products, from HBM to conventional DRAM and NAND. Meanwhile, the smartphone and PC markets are expected to show moderate growth, reflecting ongoing inflationary and macroeconomic uncertainty. However, as users increasingly experience on-device AI, AI functionality is spreading even to low- and mid-end smartphones, while AI PCs are expected to account for over half of the total PC market. Reflecting such demand environment, DRAM demand growth is expected to rise from high 10% this year to over 20% next year, while NAND demand growth is projected to improve from mid 10% this year to high 10% in 2026.

Next, I will discuss the company's plans. In the fourth quarter, we plan to continue to expand sales of HBM, server DRAM, and enterprise SSD. However, considering our normalized levels of inventories, we expect DRAM and NAND bit shipments to increase by low-single-digit QoQ for both products. For HBM, we have completed discussions with key customers for next year's HBM supply. Our HBM4, which we have completed development and mass-production preparation in September, not only fully meets customer performance requirements but also supports the highest speed in the industry. We will start HBM4 shipments in Q4 this year with further expansion planned for 2026, reinforcing our leadership position in the HBM market.

For conventional DRAM, we plan to meet increasing customer demand by securing a full lineup of the most advanced 1c-nanometer-based products across server, mobile, and graphic segments. Mass production of 1c-nanometer is already ongoing smoothly, and we plan to accelerate migration in 2026 to maintain our technology and cost leadership. For NAND, where demand recovery has been slower, we are deploying the world's highest 321-layer technology on various solution products to be ready when market conditions improve. We will also focus on supporting the growing enterprise SSD demand, all the while continuing to operate with a profitability-focused approach. Furthermore, in line with growing demand opportunities in AI servers, we are investing in tech migration to expand supply of both TLC and QLC products based on the 321-layer platform next year.

Meanwhile, we have secured customer demand across all DRAM and NAND products, including HBM, through next year. AI memory demand is significantly exceeding expectations, and this trend is expected to continue for the foreseeable future. To respond swiftly, we have recently opened the cleanroom ahead of schedule at M15X and begun equipment installation to rapidly secure new capacity. For conventional DRAM and NAND, we will accelerate the transition of existing capacity to advanced nodes to ensure robust responsiveness to rising demand. As a result, our CapEx in 2026 is expected to increase from this year's level. While continuing to maintain CapEx discipline, we will plan our investments in an optimal manner to support market demand.

Questions & Answers

Sunwoo Kim (Meritz Securities): It was mentioned that the HBM supply negotiations for 2026 have been completed. Could you share more details about the contract?

Management (SK hynix): We understand that there has been broad and deep interest in the HBM contract for next year. This year in particular has been challenging in fixing not only the supply volume but also the product mix due to various external factors. Not only that, there have been changes in the performance requirements for HBM products, which necessitated a longer time in discussing the supply contract than expected. That said, our discussions over major issues with our clients have been completed, and the HBM supply plan for next year for major customers has been finalized. Given the explosive growth in demand for HBM to keep building AI infrastructure and the company's product competitiveness, the company's HBM has been selling out since 2023, and the pricing has also been formed at a level that can sustain the current profitability. As the HBM demand continues to accelerate, driven by longer-term growth trends in the AI market, the company believes that it will be unlikely for supply to catch up with demand in a short period of time. The pace of HBM's growth will be determined by supply capacity, and the company's HBM is positioned for much higher growth than conventional DRAM products. The company's HBM supply will remain tight compared to demand into 2027, but we will continue to do our best to supply products that meet customers' needs in a timely and secure manner.

Jong-wook Lee (Samsung Securities): Thank you for taking my question, and congratulations on the performance. My question is on HBM. It was mentioned that there have been higher performance requirements for HBM4, and my understanding is that it was to be higher than the JEDEC specifications. Have there been any difficulties for the company in meeting such higher performance requirements? And for HBM4E, does the company expect the performance requirements to be similarly high, higher than the JEDEC specifications?

Management (SK hynix): With the AI inference market growing, memory bandwidth is increasingly seen as the key factor that can upgrade AI performance. As for HBM4, the number of I/Os is already fixed at 2048, double the number of HBM3E. So customers are now looking at higher speed as a way to increase HBM bandwidth. Based on our number-one technological competitiveness in HBM, the company is already fulfilling top-level specifications required by our customers. Moreover, we have already sampled products that meet customers' upgraded requirements faster than anyone in the industry and already started production for mass supply. With intensifying competition for AI chip performance, the memory wall phenomenon becomes more pronounced, where memory performance becomes the bottleneck for technological development. As such, performance requirements for next-generation memory products, including HBM, will continue to be upgraded. With the industry-leading design capability and know-how as the primary supplier, the company will respond to customers' requirements in a timely manner for our next-generation product line as well and maintain our number-one supplier position.

Simon Woo (Bank of America): Congratulations on the recent performance — KRW 10 trillion in operating profit in a quarter. Now, my question is about the memory cycle. In the past, a boom in the memory cycle was usually followed by a downturn. Does the company see any similarities between the recent memory boom and historic cycles? It appears the recent memory boom is also driving up demand for conventional memory. And in terms of inventory level, there was also a cloud-driven boom some time ago — how does the company see the inventory level from then and today?

Management (SK hynix): It is true that the memory market this year has entered into what can be called a super boom cycle with a surge in demand across all products, unlike earlier expectations. Such changes have only recently appeared, but the company sees this cycle to be a bit different from the super cycles we witnessed in 2017 and 2018. The biggest difference is that the current demand is driven by a much broader range of applications coming from the shift to the AI paradigm. AI creates upside to overall demand as it is being added on top of existing applications. At the same time, for the longer term, it is also creating new applications like autonomous driving and robotics AI. So what we are seeing is a fundamental shift in the demand for memory driven by AI. In particular, computing recently has expanded to inference, promoting demand for not just AI servers but general-purpose servers as well. The company believes that total server set shipments next year will grow at a high 10% level, and server DRAM will meet the overall demand for conventional DRAM. At the same time, looking at the supply side, production can only grow so much even if we use more cleanroom space and capacity because of the growing share of HBM. These circumstances create a structural constraint against supply increase in the DRAM industry and are likely to serve as the driver for a long-drawn-out memory super cycle.

Young Ho Ryu (NH Investment & Securities): Congratulations on the performance, and thank you for taking my question. My question is on NAND. The recently strong demand for eSSD was said to be a structural change following the advent of the AI era. Could you elaborate more on the rationale behind such an assessment?

Management (SK hynix): Let me explain in more detail the background to the higher NAND demand that we are seeing recently. First, there is stronger build demand for both AI servers and general-purpose servers, with our server customers expanding their investment in AI, which in turn is driving demand for TLC products. At the same time, demand for storage is also accelerating as a result of growth in AI-generated data like images and videos, leading to an HDD supply shortage. So for hyperscaler customers with high dependence on HDD, the recent developments have prompted them to turn instead to eSSD based on high-capacity TLC. Having said that, the company sees the recent change in demand as something that goes beyond current short-term supply and demand issues — we see this as potentially structurally increasing eSSD demand. First, with ever-advancing AI inference, the importance of RAG, or the Retrieval-Augmented Generation structure, is becoming even greater as a way to overcome the limitations of existing LLMs. RAG moves beyond the current LLM approach, which generates responses based solely on data it was trained on; it searches related documents in external databases and generates the final response based on such search, allowing it to refer to the latest data as well as user-specific data. To apply RAG on LLM, we need to additionally build outside databases that express and store data as vectors — in other words, vector databases. This is where eSSD becomes a must to enable speedy search of data. To support the scaling up of vector databases and performance upgrades in RAG, demand for storage based on high-performance TLC and high-density QLC eSSD is expected to rise. In addition, there has been a spike in data processing needed for the inference process, which led to the need to offload part of the key-value cache generated at the GPU level to the lower-layer memory. By offloading key-value cache processed at the GPU all the way to the SSD, depending on the frequency of data usage, they can increase throughput per unit of power when providing inference to many users and reduce response time per user. This is one of the reasons why use of high-performance TLC eSSD is expected to grow. As AI utilization keeps spreading, so will the role of the eSSD, resulting in higher content growth. Essentially, what we are seeing now is the benefits of AI infrastructure spreading from DRAM to NAND as well.

Min-sook Chae (Korea Investment & Securities): The memory market appears to be changing into a specialty market with an order-first, sell-later approach thanks to AI. Does the company see any differences from the past in your discussions or interactions with customers?

Management (SK hynix): It is true that in the memory market, some businesses have shifted to an order-first, produce-later approach with the emergence of HBM, marked by massive investment and long lead time. In addition, with strong HBM demand coming from our customers, the company was able to secure visibility into customer demand from the contracting stage with long-term agreements and respond with consistency. For both the memory industry and the company, this has led to greater market predictability and business stability than in the past, when it was much more volatile. And custom HBM will gradually increase from HBM4E, so products will be developed in close collaboration with customers from the early stage of design of the customer's GPU or ASIC products, unlike in the existing standardized HBM. This will lead to much longer-term and strategic transactions between customers and a small number of suppliers, contributing even more to business stability and profitability improvement on the part of memory suppliers. If I may add, memory companies are allocating capacity to ramp up HBM supply, and this has led to supply constraints in conventional memory, resulting in a supply shortage of conventional memory for which demand is actually growing. As a result, we are seeing an increase in customers who want to sign long-term agreements for conventional memory products as well. Some customers are very actively responding to the current supply shortage by issuing pre-purchase POs for 2026. Now, given customer demand and the company's capacity, for next year, not only HBM but DRAM and NAND capacity has essentially been sold out. The company will try to respond to customer demand with an optimum production and sales strategy, and we will keep discussing the implications of the HBM-driven changes with our customers.

Ricky Seo (HSBC): Congratulations on the performance, and my question is on CapEx. Recently, investments by global AI companies point to a very high investment level needed for the next few years to fulfill the demand for memory. It was mentioned that the company's CapEx in 2026 will increase over this year. What will be the extent of the increase? And for the longer term, it appears as if the investment or CapEx into new campuses like Yongin and others will also have to be far higher than one year ago.

Management (SK hynix): As global AI companies competitively expand investment with conviction in the growth and monetization of the AI market, there has been accelerated growth in demand for a wide range of memory products, including HBM, DDR5, and enterprise SSD. To respond to such surging demand, CapEx growth across the memory industry appears to be inevitable. For the company, CapEx next year will far outpace the level of this year. For M15X, equipment installation has begun in earnest to ramp up supply of HBM. For conventional DRAM and NAND, we will accelerate tech migration in the existing capacity as a way of responding to demand. At the same time, considering the Fab 1 construction in Yongin and preparation for construction of an advanced packaging plant in Indiana, U.S., investment in infrastructure is set to keep growing next year. But even with growing CapEx, the company will stick to its CapEx discipline and maintain a stable financial structure.

Bo-young Choi (Kyobo Securities): Thank you for taking my question. My questions are on the product and technology. First, it was mentioned that there will be conversion to 1c-nanometer next year, for which there is a lineup for all products, and also an increase in the portion of 321-layer NAND products. What is going to be the timeline for the ramp-up for each product? And what is the expected portion of the respective products by the end of next year?

Management (SK hynix): Under the principle that we respond with priority to demand with high visibility and profitability, our new capacity next year will center on HBM, for which supply contracts have already been completed. For DRAM and NAND, we plan to respond to demand through tech migration in existing capacity. For DRAM 1c-nanometer, development was completed, with mass production beginning this year. Ramp-up will begin in full swing next year, and 1c-nanometer is planned to take up over half of the conventional DRAM capacity inside Korea by the end of next year. Based on the 1c-nanometer process, with the best performance and cost competitiveness, we will build up lineups for all products, including DDR5, LPDDR5, and Graphics DRAM, to respond to customer demand in time and ensure profitability. In the case of NAND, our focus remains on improving profitability, and the plan is to keep improving profitability through tech migration rather than ramping up capacity. That has been the case this year with tech migration from 176-layer to 238-layer, then to 321. Next year, we will grow our supply not only in TLC but also in QLC, which will require ramp-up of 321-layer products. This means that we are making preparation for 321-layer products to take up more than half of our NAND bit production by the end of next year.

Dong-Hee Han (SK Securities): Thank you for taking my question, and my question is on the inventory level. There has been considerable inventory sell-down in the second quarter, and inventory appears to be much lower in the third quarter as well due to very strong demand. What is the company's inventory level now, and what about among customers?

Management (SK hynix): With customers' demand outpacing expectations in the previous quarter, there have been concerns over excessive inventory buildup in the memory supply chain, as well as a demand slowdown consequently, but customers' inventory level has become lower overall with accelerated set build. Added to that, investment in AI infrastructure has continued to grow, resulting in noticeably lower memory inventory among server customers. And for the company, the inventory level has also fallen QoQ in both DRAM and NAND as a result of the recent strength in memory demand. This is particularly true for DRAM inventory, which remains extremely low, so much so that in the case of DDR5, products must be shipped to customers straight out of production to ensure timely response. The company will continue to try to maintain a healthy inventory level for both DRAM and NAND to seamlessly respond to customers' demand.

Nicolas Gaudois (Managing Director and Head of Research, ASEAN and APAC Technology, UBS): Regarding M15X that you mentioned you're opening earlier — are you able to address the faster ramp-up, pulling in your equipment delivery schedule? And in that regard, could you more or less complete full equipment installation for the total vertical capacity for M15X by the end of 2026? And is it possible for you to pull in the schedule as well for Yongin Fab 1 Cleanroom Readiness, which I think initially you had pinned down for May 2027?

Management (SK hynix): Let me respond to the question about the company's plan for the fab. The company decided at the end of 2023 to make new investment in M15X to preempt the fast-rising demand for HBM, which requires relatively bigger wafer capacity. After around two years of construction, the fab finally opened early — M15X opened a while ago, with equipment installation starting. We are now making preparation for M15X to contribute to HBM production ramp-up starting next year. As the memory demand growth continues to accelerate much faster than expectation, we are also speedily moving ahead with the capacity ramp-up at M15X. As for Fab 1 in Yongin, which just started construction this year, we are working to pull up the schedule in light of the pace of demand growth and the earlier ramp-up at M15X. The company will keep trying to preempt capacity and fab space by building state-of-the-art production infrastructure from M15X to Yongin Fab to enable flexible response to the ever-growing AI memory demand.

Management (SK hynix, closing): Given the nature of HBM products, it is important to agree on the long-term volume with customers to make sure there is seamless supply. When we discuss long-term volume with customers, we also consider various factors like customer relations, long-term growth potential, as well as profitability. Having said that, there have also been discussions over stronger binding contracts for conventional memory products as well, with customers issuing pre-purchase POs or asking for multi-year LTAs, and our decisions on capacity mix will be made in a way that can ensure optimum productivity. The company also sees the current trend to potentially prompt changes in the nature of the memory business for the future. As the leading supplier of AI memory, the company was able to improve our fundamentals in the memory business based on the high and stable profitability from HBM and has achieved differentiated performance. Looking ahead, we will keep responding to customers' demand with a long-term view and achieve sustained growth along with the AI market.

Management (SK hynix, shareholder returns): Because of the good performance in the market as well as the growth in the AI market, good performance for the company has been achieved and is expected to continue for some time. As a result, the company has turned around to a net cash position this quarter, and its FCF is expected to continue to improve on the back of much stronger performance. Yes, it is true that the company's financial soundness is fast improving thanks to the stronger-than-expected performance in 2025, and we have achieved a net cash position in Q3 following the higher recovery of receivables with sales growth in Q2. It was explained as part of the current shareholder return policy that the company's aim in financial soundness is to maintain an appropriate level of cash that would allow us to keep stable business management through differing industry cycles and to execute CapEx that is necessary to maintain our competitiveness. The recent upturn in the memory market has fueled demand growth, which in turn is driving up CapEx necessary to fulfill the demand. So what we see as the appropriate level of cash also has to reflect this change. Furthermore, considering the huge growth potential in the AI memory market and the company's high return on investment, I believe that shareholders will also agree that for now, the best use of cash is to reinvest it into our business while maintaining CapEx discipline. As such, being in the first year of the new shareholder return policy announced at a three-year interval, we are not looking into additional shareholder return at this time. But we will keep looking into how we can maximize shareholder return by taking a comprehensive look into the changes in the environment, both inside and outside, such as market outlook and investment needs.

Operator: Thank you very much, and that concludes the SK hynix 2025 third quarter earnings release conference.