SK hynix (Trans): LTAs include deposits; capacity adds won't cause immediate oversupply

Dolphin Research summarizes $SK Hynix(SKHY.US) FY26 Q2 earnings call Trans.

I. Key financial takeaways

1. Shareholder returns: Cash generation climbed to a record high, aided by the Kioxia stake sale and ADR issuance. The company is evaluating multiple incremental return options from various angles, but due to regulatory/process limits tied to the ADR, it cannot disclose form, size, or timing yet and plans to communicate within the year.

2. Outlook (Q3 2026): DRAM bit shipments are guided to rise ~10% QoQ, with a focus on servers to address demand. NAND bit shipments are guided to grow low single-digit % QoQ.Driven by HBM4 ramp and higher 1C nm conventional DRAM shipments, H2 bit growth should outpace H1. As the mix shifts to higher value products and HBM4 ramps, blended ASP should be positively supported.

3. Q2 earnings at a record high: Revenue was KRW 79.3tn (+51% QoQ, +257% YoY). OP reached KRW 60.5tn (+61% QoQ, approx. +57% YoY), with OPM up 5ppt QoQ to 76%, both OP and margin at all-time highs.D&A was KRW 4tn, EBITDA KRW 64.6tn with an 81% margin.

4. One-off investment gains boosted net income: Non-operating net profit was KRW 62.2tn, including KRW 1.1tn FX gains and KRW 63.3tn from asset disposal/revaluation (incl. Kioxia stake sale). As a result, PBT was KRW 122.7tn, net profit KRW 93.9tn, and net margin 118%.The net margin above revenue reflects the one-off investment gains.

5. Balance sheet sharply improved: Cash and equivalents (incl. short-term investments) at quarter-end were KRW 88tn, up KRW 33.6tn QoQ. Interest-bearing debt fell KRW 0.7tn to KRW 18.6tn, taking net cash to KRW 69.4tn.D/E fell 5ppt QoQ to 7%.

6. CapEx: With M15X mass production pulled forward and higher investment, 2026 CapEx is expected to be in the 'high KRW 40tn range'. New investments include PMD7 (advanced packaging) and M17 (new NAND site), and a new domestic semiconductor cluster is planned to meet long-term demand.

7. ADR listing: ADRs were listed on Nasdaq on Jul 10, the largest ADR issuance by a foreign company in US IPO history.

II. Earnings call details

2.1 Management highlights

1) Demand and market outlook

a. In Q2, amid expanding AI infrastructure investment and tight supply, pricing continued to rise. Both DRAM and NAND prices were up materially QoQ, led by AI-linked Server DRAM and enterprise SSDs.

b. AI is evolving into an agentic form that can execute complex tasks on behalf of users over long periods, with demand spreading from search and coding to productivity tools. Beyond HBM, demand is also expanding for Server DRAM that supports agent services and high-performance enterprise SSDs that process AI outputs, marking a structural shift where AI memory and conventional memory grow together.

c. While model upgrades and software optimization reduce compute per task and cost, they do not suppress overall infrastructure needs. Instead, they lower AI service prices and adoption thresholds, enlarging the user base and use cases.Major customers are still adding to infrastructure and increasing memory procurement.

d. Full-year DRAM and NAND demand are expected to grow in the mid-20% and high-10% range, respectively. Supply-demand is unlikely to ease materially near term due to higher process complexity for HBM and AI server memory and long lead times for new capacity.Tightness is expected to persist for an extended period.

2) HBM

a. HBM4: Through continuous product optimization, the company delivers the required data processing speed while achieving industry-leading power efficiency and cost competitiveness. Mass production shipments started in Q2 and will ramp fully in H2.Current yield and quality are approaching mature HBM3E levels.

b. HBM4E: Samples were shipped to a major customer in H1, using the optimal, proven, and mass-production-stable technology. Development is proceeding on schedule per the roadmap.

c. Next-gen technology: In addition to hybrid bonding, IHBM (in-package integrated cooling) for HBM5 and beyond is under development. It is expected to cut thermal resistance by 30%+, enhancing system stability and efficiency in high-performance, high-density AI environments.

3) DRAM (conventional)

a. Q2 saw full-scale supply of SOCAMM2 on the 1C nm node. The company will optimize the lineup and sample per customer roadmaps to broaden the customer base.

b. Bit shipments grew by high single-digit % in Q2, in line with guidance, with ASP up ~30% QoQ. Server ADR products (incl. SOCAMM2) saw significant sales growth.

4) NAND

a. Q2 ASP rose by mid-50% QoQ across products. Bit shipments grew by mid-teens % QoQ off a low Q1 base; enterprise SSD revenue doubled QoQ, while Solidigm 30TB+ enterprise SSD revenue rose over 3x QoQ.

b. 321-layer products had the largest share of NAND output in Q2. The company plans to lift their share to ~50% of domestic capacity by year-end, accelerating node migration and focusing on high-capacity, high-performance products.

5) Long-term agreements (LTA)

a. LTA negotiations have been concluded with about 10 customers, including key accounts, and discussions continue with major industry customers. LTAs are strategic partnerships beyond volume, intended to secure mid- to long-term supply stability and co-develop next-gen memory aligned to customers' technology roadmaps.

b. Pricing structures differ by customer and product and are designed to address price volatility. Financial mechanisms such as deposits are included to support performance and improve visibility and reliability of customers' mid- to long-term demand plans.

6) Investment and capacity

a. Mid-/long-term investment will proceed in phases as needed, balancing business feasibility and efficiency. Domestically, Icheon and Yongin will be core hubs for next-gen DRAM and AI memory, while Cheongju will be strengthened for NAND and advanced packaging.

b. M15X mass production has been pulled forward. The Yongin Fab1 cleanroom is expected to start in early 2027 and ramp quickly thereafter.

2.2 Q&A

Q: Some hyperscalers are considering leasing data centers, and more efficient AI models are emerging, raising concerns that AI infrastructure investment may slow. What are you hearing from major CSPs about AI capex, and what does it mean for HBM/DRAM/NAND demand?

A: We recognize market concerns about a slowdown, but we see leasing and the use of more efficient models as improving utilization and monetization of large-scale AI infrastructure, not a cut in investment. For major CSPs, AI capability is tightly linked to their core competitiveness in search, ads, cloud, and software, so investment aimed at strengthening AI should remain solid.

We also do not think more efficient models will reduce infrastructure needs; they broaden accessibility and adoption. Higher efficiency allows the same infrastructure to support more users and services.The recent surge in demand sparked by efficient models supports this view and is consistent with our discussions with key customers on their mid- to long-term outlooks.

We expect CSP AI-related investment to continue over the mid-/long term, and customers' memory needs reflect this trend. Timing of individual projects may vary due to power availability and data center construction constraints, but with CSP competition and AI service expansion, AI capex should stay firm into next year and beyond.As a result, overall memory demand will keep expanding, not only for HBM used in AI compute, but also for Server DRAM that enables agentic AI and high-performance, high-capacity NAND that supports AI service growth and data expansion.

Q: You plan significant capacity additions mid-/long term. What underpins your long-term memory demand outlook, and does it include LTA-locked demand? How do you address concerns about potential oversupply after expansion?

A: Our capacity strategy is based on structurally rising memory demand driven by AI expansion and ongoing dialogues with key customers about their long-term needs. Our relationships are evolving from transactional to more strategic partnerships, and stronger willingness to sign LTAs validates the sustainable demand created by the AI ecosystem.

The planned expansion is grounded in demand visibility obtained through customer collaboration. Actual capex and ramps will be staged, considering demand visibility and investment efficiency.Given expansion will flexibly align with confirmed customer needs, we do not expect our mid-/long-term plans to immediately create oversupply.

Q: Peers have announced LTAs. Can you elaborate on SK hynix's LTA framework, such as contract length and pricing?

A: Our LTAs are tailored by customer and product. Contract terms are typically around five years, but specifics vary; pricing structures also differ, and we are exploring mechanisms that better handle price volatility to reduce uncertainty from short-term swings while enhancing long-term stability for both sides.

Given the impact of demand swings on memory cycles, securing effective purchase commitments is also vital. In addition to long-term volume commitments, mechanisms such as deposits are included to strengthen performance and improve demand visibility, with terms depending on each customer's needs and the contract structure.While we cannot disclose what share of total sales will be under LTAs, we will keep it at an appropriate level based on market conditions and customer needs.

This approach should bolster downside resilience while preserving flexibility to capture incremental demand as the market improves. With long-standing partnerships with major AI customers like NVIDIA and a solid HBM-centric profit base, we will use visibility from LTAs and operating flexibility to reinforce HBM leadership and balance stability with profitability.

Q: Q2 DRAM ASP growth seemed below market expectations. Why, and what is the outlook for H2? How do you manage the mix between HBM and conventional DRAM?

A: Some high-value products shipped later than planned into H2, and mix shifts weighed on blended ASP in Q2. These factors should gradually fade in H2.With HBM4 ramping and 1C nm DRAM shipments increasing, H2 bit growth should exceed H1. Considering demand and mix changes, higher HBM4 sales and growing high-value contributions should also lift blended ASP.

This points to higher shipments and a steadily improving mix, supporting H2 ASP and profitability. We do not manage to near-term price or profit; rather, we consider demand visibility, long-term customer relationships, and segment-specific supply-demand dynamics.This principle will continue to guide us to capture growth while delivering stable, sustainable profit growth.

Q: Some say competitors are catching up fast in HBM. What differentiates your HBM4, and how will you maintain leadership?

A: HBM4 competitiveness is not only about performance delivery but also the ability to supply at scale with stable yields and consistent quality. Since HBM2E, SK hynix has demonstrated strengths in time-to-market, performance, mass production, yield, quality, and customer trust—advantages that are hard to replicate quickly.

We began HBM4 mass production for key customers in Q2, with yields and quality near mature HBM3E levels, and are focused on steady capacity ramps. HBM4E samples have been delivered using proven, mature processes with mass-production stability, and development is on track per the roadmap.

We are also investing ahead in next-gen technologies: beyond hybrid bonding, we are developing IHBM for effective cooling in HBM5 and beyond, integrating cooling elements within the package to cut thermal resistance by 30%+ and improve system stability and efficiency in high-performance, high-density AI setups. As AI expands and accelerators grow more complex in performance and packaging, customers will value partners with reliable manufacturing, quality, and stable supply.Given HBM's high value, quality issues can impose significant costs and systemic impacts on customers.

Leveraging early co-development and long-term strategic partnerships, we will keep delivering the right products at the right time and lead migration to next-gen technologies. This underpins our HBM leadership.

Q: How are 2027 HBM price talks progressing? Can you share the overall status of contracts and pricing outlook (HBM4E/HBM4)?

A: We are negotiating 2027 HBM volumes and pricing with key customers, progressing well on solid demand, but cannot disclose customer-specific terms. Recent sharp gains in conventional DRAM prices may influence HBM pricing discussions, but HBM pricing is not determined solely by conventional DRAM.

HBM requires more resources—greater wafer input, advanced processes, TSV and packaging capacity—than conventional DRAM. With each generation, performance and quality requirements rise, and development and qualification grow more complex.Thus, our pricing discussions consider conventional DRAM prices and market balances, HBM-related resource and opportunity costs, technology complexity, and the value created for customers.

Our goal is reasonable profitability commensurate with differentiated value while fostering healthy, sustainable AI ecosystem growth. With accumulated technology leadership, cost competitiveness, robust manufacturing, and deep customer trust and collaboration, we aim to sustain solid HBM profitability through successful generational transitions and ongoing value creation.We will reinforce our position as a strategic partner growing with customers in the AI era, focusing on long-term, sustainable growth and returns.

Q: Beyond the large investments announced in Korea, there is talk of expanding in the US and Japan. What is your domestic and overseas investment strategy?

A: In the AI era, technology leadership alone is not enough; delivering the required volume at the right time has become a core competitive capability. Especially amid acute shortages, supplying the memory the ecosystem needs is our responsibility.Our mid-/long-term investments will be timed to AI memory demand and executed with capex discipline based on feasibility and returns.

Over the mid-/long term, we will add manufacturing capacity through an optimal mix of maximizing existing sites and building new infrastructure where necessary. In Korea, Icheon and Yongin will remain the core hubs for next-gen DRAM and AI memory, while Cheongju will be strengthened for NAND and advanced packaging.The recently announced large-scale investments are part of this strategy, laying a manufacturing foundation for future demand.

We do not rigidly distinguish between domestic and overseas when selecting future sites. Decisions will optimize across power, water, talent, supply chain, the semiconductor ecosystem, and customer proximity.Beyond the announced plans, no further decisions have been made; we will secure capacity at the right time to meet customer needs and enhance investment efficiency by leveraging existing assets and assessing new investments prudently.

Q: AI inference growth and KV cache offload are boosting the importance of enterprise SSDs. What is your strategy across segments (QLC SSDs replacing HDDs and high-performance SSDs using SLC mode), given intensifying competition?

A: As observed, the AI market is shifting from training-centric to inference-centric, and NAND is becoming a core layer in AI memory hierarchies. Hence, NAND demand is rising rapidly around enterprise SSDs (eSSD), and this trend should persist.

We do not believe a single technology will serve the AI storage market, as latency, throughput, power, capacity, and TCO needs vary by customer. Customers do not necessarily mandate a specific technology or medium; the key is reliably delivering the required performance and responsiveness for each workload.Our NAND strategy is not about choosing SLC, PLC, or QLC, but providing the optimal storage mix tailored to each customer's workloads.

For AI Data Lakes and HDD replacement where storage efficiency and cost matter, high-capacity QLC eSSDs may be most competitive, and we are strengthening that lineup. We are also focusing on new-tier AI storage for emerging use cases such as KV cache offload and near-GPU storage.

Rather than any single NAND technology, we will pair firmware with NAND strengths to deliver the most efficient performance per workload, spanning high-performance TLC eSSD, high-capacity QLC eSSD, and high-performance SSDs operated in SLC mode. No single SSD will carry all workloads in the AI era; systems will employ storage tiers optimized per workload.With a comprehensive portfolio across segments, SK hynix will proactively address evolving AI storage needs and unlock new long-term growth for NAND.

Q: On the much-watched ADRs, how is two-way fungibility being managed? Will you raise the share of ADRs outstanding?

A: Starting Jul 30 (the day after KRX listing was completed), ADRs can be freely converted into shares. Share-to-ADR conversion may be limited by process and the conversion cap.Based on other Korean companies with DR programs, converting shares to ADRs may require issuer regulatory filings that can take several weeks.

In addition, ADR float cannot exceed the ADR conversion cap, currently set at 17,790,000 shares, equal to the ADRs issued. Whether to raise the ADR share will depend on the regulatory environment and other factors.No decision has been made.

Q: After the Kioxia stake sale and ADR issuance, cash has risen sharply. How will you allocate capital, and is additional shareholder return planned this year?

A: Our capital allocation now balances three goals: timely investment to capture structural AI growth, maintaining a solid financial profile, and enhancing shareholder value through returns. We understand investor focus on shareholder returns and are evaluating multiple incremental options.

However, due to ADR-related regulatory requirements and processes, we cannot disclose any new material information not included in the offering documents. While we cannot specify the form, size, or timing of additional returns, we intend to communicate our plans within the year once determined.We will continue to invest appropriately to support growth, maintain a robust balance sheet, and execute a capital allocation policy that lifts shareholder value through sustainable cash generation.

<End of text>

Risk disclosure and statement: Dolphin Research Disclaimer and General Disclosure. Please refer to the link.