SK Hynix Inc. (000660.KS) - Deep Dive Research Report

Prepared: April 30, 2026

SECTION 1: WHAT THE COMPANY DOES

SK Hynix makes memory chips. Not "makes" in the sense of assembling components - it designs, fabricates, and packages the actual silicon that stores data inside every GPU, smartphone, and data center server on earth. More specifically, it manufactures two types of memory: DRAM (Dynamic Random Access Memory), which holds data that a processor is actively using, and NAND flash, which stores data persistently when power is off. These two product families are the fundamental building blocks of all modern computing.

Memory chips sound simple - hold data, retrieve it fast. The reality is one of the most capital-intensive, technically demanding, and cyclically violent businesses in the world. The physics of shrinking transistors to single-nanometer dimensions while stacking billions of them on a wafer the size of a dinner plate, maintaining near-perfect yields across millions of dies, and doing this consistently enough to supply the world's GPU manufacturers is a task that fewer than five companies on earth can manage competently. SK Hynix is one of them, and in the current generation of AI hardware, it holds the single most important position in the entire semiconductor supply chain.

The founding story matters here. The company was created in 1983 as Hyundai Electronics, a venture by the Hyundai conglomerate's founder Chung Ju-yung, who recognized that the automobile and heavy industry businesses he was building would need chips and that Korea needed to develop the capability domestically. Hyundai Electronics began with technology licensing from U.S. partners and produced its first 256K DRAMs in the late 1980s from its Icheon fabrication plant, the same site that remains a core production hub today. By 1992 it was the world's ninth-largest DRAM maker. By 1995, a top-20 semiconductor company globally.

The next pivotal moment came in 1999 under government pressure during the Asian financial crisis: Hyundai Electronics absorbed LG Semiconductor in a forced consolidation that made it the world's second-largest DRAM maker overnight. The combined entity inherited the scale but also a mountain of debt that eventually forced it into a near-liquidation event in the early 2000s. It restructured, rebranded as Hynix Semiconductor in 2001 - a portmanteau of "high" and "electronics" - and slowly rebuilt under government and creditor oversight. SK Group, South Korea's third-largest conglomerate, acquired Hynix in 2012 for approximately $3 billion, renamed it SK Hynix, and gave it the financial backing and strategic direction it needed to accelerate from survivor to global contender.

The second transformation came through acquisition. In December 2021, SK Hynix completed the first phase of acquiring Intel's NAND and SSD business for $6.61 billion - receiving Intel's Dalian, China NAND fab and its SSD business, which was rebranded Solidigm. The final phase closed March 2025 with a further $1.9 billion payment to absorb Intel's NAND IP and R&D team. This made SK Hynix a credible enterprise SSD player overnight, adding Intel's decades of controller firmware expertise and deep relationships with hyperscaler data center customers.

The current value proposition is this: SK Hynix is the only company that can supply high-bandwidth memory (HBM) at scale, at the leading generation, to the AI GPU manufacturers who need it. HBM is not an accessory - it is the bandwidth engine baked into every NVIDIA H100, H200, and B-series GPU that runs the world's AI training and inference workloads. Without HBM, these GPUs cannot function. Without SK Hynix, there is no adequate supply of HBM. That is the position the company occupies in 2026.

To understand why this matters, consider what happens inside an AI training cluster. A GPU like NVIDIA's H200 contains eight stacks of HBM3E memory, each stack placing twelve DRAM dies vertically and connecting them with microscopic copper pillars called through-silicon vias (TSVs). This architecture delivers over 3.35 terabytes per second of memory bandwidth to the GPU - roughly 18 times what conventional DDR5 can deliver at the channel level. When researchers run a large language model like GPT-4 or its successors, the model weights - billions of parameters - sit in HBM. The speed at which those weights can be read into the compute cores determines how fast the model runs and how much it costs per token to operate. Memory bandwidth, in other words, is the bottleneck of AI. SK Hynix makes the fastest memory.

"The importance of memory has become greater than ever... as this supply-demand imbalance persists, customers are prioritizing procurement over price."

• SK Hynix CFO, Q1 2026 Earnings Call, April 23, 2026

SECTION 2: BUSINESS SEGMENTS

2.1 DRAM - The Core Engine

DRAM accounts for roughly 65-70% of SK Hynix's revenue and is where the company's most important competitive battles are fought. Within DRAM, the product types serve fundamentally different end markets with different competitive dynamics.

High Bandwidth Memory (HBM) is the company's crown jewel and fastest-growing product. HBM is technically still DRAM - stacked DRAM dies connected vertically - but the manufacturing complexity, customer relationship dynamics, and economics are completely different from conventional DRAM. HBM revenue more than doubled year-over-year in 2025 and reached an estimated 40% of total DRAM revenue. SK Hynix commands approximately 55-62% of the global HBM market by revenue, and its position is protected by a combination of first-mover manufacturing advantage, proprietary packaging technology (MR-MUF - Mass Reflow Molded Underfill), and deep co-development relationships with NVIDIA that span years.

The core capability in HBM is not just DRAM fabrication - it is the ability to stack 12 or 16 DRAM dies, drill 50,000+ TSV micro-holes through each die, align the stack to tolerances measured in nanometers, and encapsulate the whole assembly in molding compound that manages heat without warping. SK Hynix holds key patents on MR-MUF, the process by which molten compound flows uniformly through the micro-gaps between dies. Competitors Samsung and Micron use Thermal Compression with Non-Conductive Film (TC-NCF), a slower and less yield-efficient process. SK Hynix also sources a proprietary epoxy molding compound exclusively from Namics Corporation of Japan, giving it both a patent barrier and a supply chain lock that competitors cannot easily replicate.

SK Hynix began its HBM development journey in 2009, when it identified TSV technology as a potential breakthrough for breaking memory bandwidth limits. The company shipped the world's first HBM in 2013, worked collaboratively with AMD on HBM1 and HBM2 in the early GPU market, and then positioned itself for the NVIDIA wave that came with HBM3 and HBM3E when AI training became the dominant computing workload. In September 2025, SK Hynix became the first company to achieve mass production readiness for HBM4, the next generation.

Server DRAM (DDR5 and high-density modules) is the second-largest sub-segment within DRAM. As AI infrastructure drives hyperscalers to build larger server clusters, the need for dense DDR5 server memory modules has grown sharply. High-density DDR5 module shipments grew approximately 50% quarter-over-quarter in Q4 2025 alone. SK Hynix's 1cnm (sixth-generation 10nm-class) process, now in full-scale mass production, gives it cost advantages per bit for server DRAM.

Mobile DRAM (LPDDR) serves smartphone and laptop manufacturers. This is a larger-volume, lower-ASP market where SK Hynix competes with Samsung and Micron across LPDDR4, LPDDR5, and the emerging LPDDR6 standard. It provides baseline volume and cash flow, though margins are tighter than HBM or server DRAM. SK Hynix completed development of LPDDR6 on its 1cnm process in early 2026, targeting next-generation flagship smartphones.

Graphics DRAM (GDDR) serves gaming GPU manufacturers and is a smaller but stable sub-segment. GDDR7 represents the current frontier.

The DRAM segment exists as the heart of SK Hynix because memory is fundamentally an industry where scale, process technology leadership, and customer relationships reinforce each other. The company that achieves the most advanced process node first gets better yields, lower costs, and can quote more competitive prices or capture better margins - then reinvests those profits into the next generation. HBM has added a third dimension to this dynamic: the customer who qualifies you first builds their GPU architecture around your memory stack geometry, making switching technically and commercially painful.

2.2 NAND Flash - The Acquired Scale Play

NAND accounts for roughly 30-35% of SK Hynix's revenue. Unlike DRAM, where SK Hynix developed its capabilities organically over 40 years, a major portion of its NAND franchise was acquired through the Intel transaction. The segment has two distinct heritage streams: SK Hynix's own NAND business, which excels in mobile flash, and Solidigm's enterprise SSD portfolio, which carries Intel's deep hyperscaler relationships.

SK Hynix's native NAND capability is centered on its 3D NAND technology. The company has been racing up the layer count - the number of cell layers stacked vertically in each die, which directly determines storage density and cost per bit. In 2025, the company completed development of its 321-layer TLC and QLC (Triple and Quad Level Cell) products, the latter representing an industry record. A two-generation leap in productivity, 321-layer NAND was expected to exceed 50% of total bit production by end of 2026.

Solidigm is a separate legal entity, a wholly-owned subsidiary that operates under its own brand. It inherited Intel's enterprise SSD product lines, controller and firmware teams, and Optane-era relationships with data center buyers. The product portfolio spans twelve enterprise SSD families organized into three groups: performance-optimized SSDs for databases and caching, capacity-optimized SSDs for bulk storage, and cost-optimized SSDs for cold storage tiers. Solidigm is ranked second in enterprise SSD market share by segment, and its controller firmware expertise - which Intel spent decades developing - allows SK Hynix NAND dies to be deployed in solutions that can actually compete with Samsung's vertically integrated SSD stack.

The NAND segment's strategic rationale within the group is hedging and completeness. HyperscalerAI workloads require both DRAM (for active inference) and large-capacity NAND storage (for model weights, training datasets, and vector databases). Being able to supply both as a single vendor simplifies qualification for customers and gives SK Hynix a broader footprint in AI infrastructure spending. The segment also generates stable volume revenue that partially offsets DRAM's cyclicality, though NAND itself is cyclical and the first half of 2025 saw pricing weakness before AI-driven eSSD demand revived it in H2.

The management priority for NAND is clear: concentrate on high-value eSSD (enterprise SSD) where Solidigm can command premium pricing, minimize exposure to low-margin commodity NAND, and use the 321-layer QLC technology to compete on density economics. The company explicitly stated it responded to Q2 2025 demand by centering output on eSSD rather than spreading across all flash products.

Segment Comparison Summary

SECTION 3: PRODUCTS AND BUSINESS DETAIL

The DRAM Manufacturing Process

DRAM manufacturing begins with growing a silicon crystal ingot, slicing it into thin 300mm wafers, and then applying hundreds of sequential photolithography steps to etch the capacitor and transistor structures that form each memory cell. Modern DRAM at SK Hynix uses the 1cnm process node - the sixth generation of 10nm-class DRAM - which employs extreme ultraviolet (EUV) lithography for precision layer exposure that older 193nm immersion tools cannot achieve at required dimensions.

SK Hynix was notably the first DRAM manufacturer in the world to install ASML's High-NA EUV system, the EXE:5200B, at its M16 fab in Icheon in 2025. High-NA EUV doubles the numerical aperture of the lens system from 0.33 to 0.55, enabling finer pattern resolution that allows further cell miniaturization. This positions SK Hynix a generation ahead of competitors on the next-generation 1anm node (the 7th generation 10nm-class DRAM), which will underpin HBM4E and future server DRAM generations.

The 1cnm process uses more than five EUV exposure layers - up from two or three in prior generations. Each EUV step requires scanning the wafer with extreme ultraviolet light at 13.5nm wavelength in a near-vacuum environment to expose the photoresist pattern. The process window for each step is measured in angstroms. Yield - the percentage of dies on each wafer that function correctly - is what separates profitable from unprofitable DRAM production. SK Hynix has reportedly achieved stable mass production yields on 1cnm, which it communicated in Q3 2025.

HBM: The Stack Architecture

HBM manufacturing is a separate process that begins after the base DRAM dies are fabricated. The steps include:

1. TSV drilling - Etching microscopic via holes (approximately 5 microns wide, 50+ microns deep) through each DRAM die using deep reactive ion etching
2. TSV filling - Depositing tungsten or copper into the vias
3. Wafer thinning - Grinding the back of each die to expose the TSV tips for bonding
4. Die stacking - Aligning 8, 12, or 16 DRAM dies on top of each other with sub-micron precision using flip-chip bonding equipment
5. MR-MUF encapsulation - Flowing SK Hynix's proprietary molding compound through the inter-die gaps at controlled temperature and pressure, using the Namics epoxy compound
6. Base die integration - Attaching the die stack to a logic/interposer base die (for HBM4, this base die is manufactured by TSMC on advanced logic nodes)
7. Testing - Running electrical validation at temperature to screen defects

HBM4 introduced a significant change: the base die, which handles the memory controller and interface logic, is now manufactured by TSMC on an advanced logic process rather than using a conventional DRAM process. This co-development with TSMC is essential for HBM4's performance improvement - specifically the doubling of bandwidth versus HBM3E - but also means SK Hynix's supply chain now depends on TSMC's CoWoS packaging availability.

Product Catalogue

DRAM Products:

• HBM3E 8Hi and 12Hi - Current mass production, 24GB and 36GB per stack, powering NVIDIA H200 and Blackwell GPUs
• HBM4 - First mass production revenues achieved in Q4 2025; Rubin GPU supply for NVIDIA from 2026, 48GB capacity at 11.7Gbps
• HBM4E - Targeted for 2027 mass production using 1cnm core dies
• DDR5 server DRAM - 32GB and 64GB modules; high-density 128GB+ RDIMM modules growing fastest
• LPDDR5X - Mobile DRAM, 5500MT/s, used in Samsung Galaxy, Qualcomm Snapdragon platforms
• LPDDR6 - Development completed in early 2026 on 1cnm; for next-gen flagship smartphones
• GDDR7 - Graphics DRAM for gaming GPUs
• SOCAMM2 - 192GB server on-chip addressable module, mass production started April 2026; for AI inference servers

NAND Products:

• 321-layer TLC 3D NAND - Mass production 2025; highest layer count commercially available
• 321-layer QLC 3D NAND - Development complete 2025; highest density per die globally
• Solidigm D5-P5336 - Enterprise eSSD using 192-layer QLC NAND; 61.44TB capacity
• Solidigm P5530 - Enterprise eSSD combining SK Hynix NAND with Solidigm controller
• Client NAND - Used in smartphones, USB drives, and consumer SSDs

Manufacturing Footprint

SK Hynix operates fabrication plants across South Korea and China:

South Korea:

• M14 (Icheon) - Legacy DRAM fab, still operational for older nodes
• M16 (Icheon) - Flagship DRAM fab, site of High-NA EUV installation; primary HBM production base
• M10 (Cheongju) - NAND production
• M15 (Cheongju) - NAND production
• M15X (Cheongju) - New DRAM fab; began trial operations May 2026, four months ahead of original schedule; invested over KRW 20 trillion; targeted to reach approximately 50,000-60,000 wafers/month at full capacity
• Yongin Semiconductor Cluster - New mega-complex; Phase 1 (first fab) moved up to February 2027; full fab of 6 cleanrooms to complete through 2030; ultimate capacity addition of approximately 350,000 wafers/month; KRW 21.6 trillion (~$15 billion) committed for Phase 1 alone

China:

• Wuxi fab - Approximately 40% of total DRAM output; conventional DRAM (DDR4, LPDDR5); cannot receive EUV equipment under U.S. export controls but operates profitably on existing equipment
• Chongqing - NAND packaging
• Dalian (formerly Intel) - NAND wafer fabrication, acquired via Intel transaction

United States:

• West Lafayette, Indiana - Advanced HBM packaging plant under construction; $3.87 billion total investment; received $458 million CHIPS Act grant and up to $500 million in CHIPS Act loans; targeted to mass-produce next-generation HBM packaging domestically, expected to support approximately 1,000 facility jobs

SECTION 4: CUSTOMERS

Who Buys and Why

SK Hynix's customer base divides cleanly into three categories: AI GPU manufacturers (primarily NVIDIA), hyperscale cloud providers who buy server memory independently, and consumer electronics and mobile OEMs.

NVIDIA is the dominant customer and the relationship that defines the current era. NVIDIA reportedly accounted for approximately 27% of SK Hynix's total revenue in the first half of 2025, generating around KRW 11 trillion in H1 2025 revenue alone. In 2024, NVIDIA contributed approximately KRW 10.9 trillion - 16% of SK Hynix's annual revenue. That concentration shifted further toward NVIDIA in 2025 as HBM revenue more than doubled. The buying decision inside NVIDIA is made at the GPU architecture level: NVIDIA's chip architects specify HBM requirements years in advance of a GPU's launch, co-develop the interface specifications with SK Hynix, and then build the Hopper, Blackwell, or Rubin architecture around those parameters. Once an HBM stack's physical dimensions, electrical interface, and bandwidth specifications are locked into a GPU chip design, switching memory suppliers requires a re-tapeout of the chip - a process that takes 12-18 months and costs hundreds of millions of dollars. The switching cost is architectural, not contractual.

Hyperscale cloud providers - Amazon Web Services, Microsoft Azure, and Google Cloud - buy both through NVIDIA GPUs (which contain SK Hynix HBM) and directly as independent server DRAM buyers. They typically negotiate through long-term supply agreements (LTAs) that provide price stability in exchange for volume commitments. SK Hynix management has indicated in recent concalls a willingness to expand LTAs to reduce industry volatility and improve investment predictability. These customers specify their DRAM needs through procurement teams that evaluate technical qualification, yield data, supply reliability, and price - a process that takes months but once established is sticky.

OpenAI has been publicly identified as an SK Hynix supply agreement partner for its Stargate AI infrastructure project. This represents the emerging direct-to-AI-developer customer relationship that bypasses the GPU manufacturer layer.

Mobile OEMs - Samsung (as a device maker separate from Samsung Semiconductor), Apple (via its supply chain for iPhones and Mac chips that use LPDDR), Qualcomm, and MediaTek - buy mobile DRAM through qualification processes administered by the OEM's hardware team. Apple's LPDDR qualification, for instance, requires extensive reliability and power testing over 18-24 months before a new process node is approved for iPhone use. Once qualified, a supplier stays in the bill of materials for the product's lifecycle.

Contract structure in HBM is distinctive: SK Hynix confirmed in the Q3 2025 concall that "the HBM supply plan for next year has been finalized with major customers" and that pricing was "set to sustain current profitability." In practice this means HBM is sold through annual supply agreements with volume allocations and reference pricing - not spot market pricing. Conventional DRAM has a mix of contract pricing (large server customers on quarterly or semi-annual contracts) and spot pricing (small buyers and speculators). The spot market is thin - management explicitly dismissed spot price weakness in Q1 2026 as "not representative" of the true supply-demand balance.

The concentration dynamic around NVIDIA is a double-edged fact. It is a risk of customer dependence. It is also evidence that SK Hynix has won the most critical technical qualification in the industry. No other company has made HBM at the scale and quality that NVIDIA's production volumes require. The concentration did not arise from aggressive pricing - it arose from being the only company that could actually deliver. Management noted in Q1 2026 that "customer demand for the next three years already exceeds our supply capacity," which means the concentration reflects structural supply shortage, not pricing leverage.

SECTION 5: COMPETITIVE LANDSCAPE

The memory industry is a global oligopoly with three players controlling over 90% of DRAM output and roughly 85% of NAND output. The barriers to entry are not just high - they are generationally compounding. Every year a potential entrant delays entering, the incumbents sink another $10-20 billion into next-generation capacity that produces chips the entrant cannot make.

DRAM Competition

Samsung Electronics (Semiconductor Division) is SK Hynix's primary DRAM rival. Samsung holds approximately 37% of the DRAM market versus SK Hynix's 33%. In conventional DRAM, Samsung has historically led on process transitions and volume. In HBM specifically, Samsung has fallen meaningfully behind. Samsung's HBM3E products were initially rejected by NVIDIA for qualification due to heat and yield issues in 2024. Samsung's HBM3E market share was reported as low as 17% in mid-2025, compared to SK Hynix's 62%, before recovering to approximately 35% by late 2025. The gap reflects Samsung's continued reliance on TC-NCF packaging rather than MR-MUF. Samsung has committed to delivering HBM4 at scale in 2026 with TSMC's base die support, but it is running a generation behind in qualification and customer confidence.

Micron Technology is the U.S.-based third DRAM player with approximately 30% market share overall. In HBM, Micron has surprised the industry by reaching 21% HBM market share in Q2 2025, overtaking Samsung to become the second-largest HBM supplier. Micron's HBM3E was reportedly qualified by NVIDIA for Blackwell-era GPUs, giving it a meaningful foothold. Micron uses TC-NCF packaging and benefits from U.S. geography for CHIPS Act subsidies and domestic AI customer preference. It is a credible No. 2 in HBM but operates at smaller scale than SK Hynix and lacks the same depth of co-development history with NVIDIA.

Barriers to entry in DRAM are extreme: a single modern DRAM fab costs $10-20 billion to build, requires 2-3 years from groundbreaking to first wafer, and then 12-18 months to ramp yields to commercial levels. Even with unlimited capital, a new entrant would need 4-5 years and $30+ billion before producing competitive dies. HBM adds a further layer: TSV processing expertise takes years to accumulate, and NVIDIA's architectural co-development timeline means new HBM suppliers would need to be embedded in customer development programs three generations before they can be relevant. China's CXMT (ChangXin Memory Technologies) is attempting DRAM entry with government backing but is confined to older nodes (DDR4) due to export controls on EUV equipment.

NAND Competition

In NAND, the competitive landscape is broader. Samsung controls approximately 30% of the market. Kioxia (formerly Toshiba Memory, partnered with Western Digital) and Western Digital together represent another 30-35%. SK Hynix/Solidigm has approximately 20-21% combined. Micron holds approximately 13%. For enterprise SSD specifically, Samsung and Solidigm are the top two by market segment share, with Samsung's vertically integrated controller, NAND, and firmware stack a formidable advantage that Solidigm partially matches through Intel's legacy controller IP.

The competitive dynamic in NAND is structurally more challenged than DRAM. Layer count races compress per-bit margins as everyone rushes to 200+ and 300+ layer products. QLC enables more bits per die but requires more sophisticated error correction, which is where Solidigm's firmware lineage earns its keep. SK Hynix does not win in commodity NAND - it wins in enterprise SSDs where controller intelligence and reliability qualification matter.

Where SK Hynix is strong: HBM unequivocally. The MR-MUF advantage, Namics exclusive supply agreement, deep NVIDIA co-development, and 16-year TSV head start create a moat that is real and measurable in market share terms. In server DRAM, process leadership on 1cnm and High-NA EUV gives a cost and performance edge that may translate to share gains over 18-24 months.

Where SK Hynix is exposed: Customer concentration in HBM. Wuxi fab vulnerability to U.S. export control escalation. NAND margins in a commodity environment. Samsung's potential HBM recovery with HBM4. Any technology transition where a proprietary packaging approach (MR-MUF) proves inadequate for next-generation thermal or bandwidth requirements.

SECTION 6: INDUSTRY

What Drives Demand

Memory demand has two distinct engines in 2026. The AI infrastructure engine, which is supply-constrained and growing faster than capacity can be added, and the consumer/enterprise server engine, which is more cyclical and tied to traditional IT spending.

On the AI side: every AI GPU shipped contains HBM. Every AI server contains DDR5 server DRAM. Every AI storage system contains enterprise SSDs. As hyperscalers - the Amazons, Googles, and Microsofts - and AI-native companies - the OpenAIs and Anthropics - compete to deploy more inference capacity faster, their orders for AI infrastructure hardware (GPU clusters) create multiplicative demand for memory at each layer. TrendForce estimates that AI applications will consume approximately 20% of global DRAM wafer capacity in 2026. HBM alone is projected to consume more than 15% of total DRAM die output by bit volume, despite representing a far smaller fraction of total wafer starts, because each HBM stack requires multiple full dies stacked together.

The evolution from AI training to AI inference compounds demand further. Training runs happen once per model, but inference runs billions of times daily. As AI agents - software systems that repeatedly reason and act in real time - proliferate, the inference workload grows continuously. SK Hynix management described this dynamic explicitly in Q3 2025: "The AI market is now shifting rapidly from the training phase of large models to the inference phase," noting that inference creates sustained, continuous demand rather than burst demand tied to model training cycles.

On the conventional side: server DRAM demand is tied to global data center buildout, which continues regardless of AI because email, databases, video streaming, and enterprise applications run on the same infrastructure. PC and mobile DRAM demand tracks consumer replacement cycles and smartphone upgrade rates. These segments are slower-growing but provide volume base load.

Industry Size and Structure

The global semiconductor memory market was valued at approximately $171 billion in 2025, projected to reach $447 billion by 2034. DRAM represents roughly 65% of memory market revenue; NAND the remainder. The HBM sub-segment is projected to grow from approximately $38 billion in 2025 to $58 billion in 2026, with trajectories toward $100 billion by 2028 if AI infrastructure investment sustains its current pace.

Global DRAM output is concentrated in South Korea (Samsung, SK Hynix), the United States (Micron), and increasingly China (CXMT for older nodes). NAND output is more geographically distributed, spanning Korea, Japan (Kioxia), the U.S. (Micron), and China.

Cyclicality

Memory is the most cyclical major industry in semiconductors. The classic cycle: supply expands based on demand forecasts, demand disappoints or shifts, inventory builds, prices collapse, producers cut capex, supply tightens, prices recover, the cycle restarts. SK Hynix itself went through exactly this in 2022-2023 when a post-COVID inventory correction wiped out DRAM profitability and pushed the company to operating losses in late 2022 and early 2023.

The current cycle has structural features that are prolonging the upturn. AI demand is genuinely new - it was not predicted in prior capital allocation decisions - and HBM capacity requires 3-4 years to meaningfully expand because it involves both DRAM fab capacity and TSV packaging capacity, which are separate bottlenecks. Meanwhile, both Samsung and SK Hynix are explicitly "minimizing the risk of oversupply" by restraining conventional DRAM capex, a rational response to the memory industry's self-inflicted 2022 wound. Goldman Sachs forecast a 4.9% DRAM undersupply in 2026, the worst supply deficit in 15+ years, persisting into 2027.

The bear scenario from a cycle perspective is not 2026 but 2028-2029, when M15X, Yongin Phase 1, and Samsung's own expansion waves all begin producing simultaneously. The industry has been here before.

Regulatory and Export Control Environment

U.S. export controls on semiconductor equipment are the primary regulatory risk for the industry. EUV lithography machines cannot be exported to China, preventing Chinese DRAM manufacturers from advancing beyond approximately 20nm-class process nodes. For SK Hynix, this means its Wuxi fab cannot receive the equipment needed for next-generation DRAM, effectively capping the Wuxi fab at conventional DRAM on existing equipment. The U.S. has shifted from an open-ended waiver to an annual licensing regime for equipment shipments to Wuxi - Samsung and SK Hynix received 2026 licenses in December 2025. Whether equivalent licenses are issued for 2027 is a policy decision, not a technology decision.

U.S. tariffs on Korean semiconductor exports remain a secondary consideration - Korea has a different trade relationship with the U.S. than China, and memory semiconductors have historically been excluded from broad tariff actions. However, the current trade environment creates uncertainty that management has explicitly flagged as requiring monitoring.

SECTION 7: GROWTH TRIGGERS

Sourced exclusively from the four most recent concalls: Q2 2025 (July 24, 2025), Q3 2025 (October 28-29, 2025), Q4 2025 (January 29, 2026), and Q1 2026 (April 23, 2026).

• HBM4 full-scale sales expansion throughout 2026. HBM4 completed development and entered mass production in September 2025, with initial shipments beginning in Q4 2025. Management committed to full-scale expansion in 2026. Shipments began on schedule. (Q3 2025 concall, October 28-29, 2025; confirmed Q4 2025 concall, January 29, 2026)

"HBM4 completed development in September and entered mass production, fully meeting customer performance requirements and supporting industry-leading speeds. Shipments will begin in Q4 2025, with full-scale sales expansion planned for next year."

• Q3 2025 concall, October 28-29, 2025

• M15X fab production beginning May 2026, four months ahead of original schedule. First cleanroom commencing trial operations drives meaningful incremental DRAM wafer output. Equipment installation in second cleanroom underway for year-end 2025 completion. (Q3 2025 and Q4 2025 concalls)

• 1cnm DRAM exceeding 50% of Korean fab capacity by end-2026. Migration to the most advanced and cost-efficient DRAM process improves margins per bit across server, mobile, and HBM product lines. (Q3 2025 concall, October 28-29, 2025)

• 321-layer NAND exceeding 50% of total bit production by end-2026. Density improvement reduces cost per bit in NAND and positions Solidigm's enterprise SSDs more competitively. (Q3 2025 concall; reaffirmed Q1 2026 concall)

• Yongin Semiconductor Cluster Phase 1 first cleanroom completion moved up to February 2027 from the original May 2027 schedule. Full Phase 1 capacity of approximately 350,000 wafers/month adds transformational long-term DRAM output. (Q1 2026 concall, April 23, 2026)

• US ADR listing targeted for June-July 2026. Confidential SEC Form F-1 filing submitted March 24, 2026. Listing is expected to raise capital ($9-14 billion estimated by industry analysts) to fund domestic U.S. expansion and global capex programs. (Q1 2026 concall)

• Indiana advanced packaging plant ramping to produce HBM domestically. CHIPS Act grant of $458 million secured; plant will mass-produce next-generation HBM packaging for U.S.-based customers, reducing geopolitical supply chain risk and deepening NVIDIA/hyperscaler relationships. (Q4 2025 concall, January 29, 2026)

• HBM4E development targeting 2027 mass production using 1cnm core dies. Next generational leap in HBM performance extends the product roadmap visibility. (Q1 2026 concall, April 23, 2026)

• SOCAMM2 (192GB server module) mass production started April 2026. New high-density server memory form factor targets AI inference servers requiring large memory per slot. (Q1 2026 concall)

• DRAM demand exceeding 20% growth in 2026, NAND demand growing high-teen percent. Management's own demand projections suggest volume tailwinds independent of pricing. (Q3 2025 concall; Q4 2025 concall reaffirmed)

• Long-term supply agreements (LTAs) expansion. Management signaled willingness to lock in multi-year agreements with hyperscalers and AI companies to reduce volatility and improve investment planning. (Q1 2026 concall)

"Customer demand for the next three years already exceeds our supply capacity."

• Kim Ki-tae, Head of HBM Sales and Marketing, Q1 2026 concall, April 23, 2026

• U.S. AI company establishment to evolve SK Hynix from pure memory component supplier toward being a "full-stack creator" in the AI ecosystem. (Q4 2025 concall, January 29, 2026)

SECTION 8: KEY RISKS

Risk 1: NVIDIA Customer Concentration

NVIDIA reportedly drove 27% of SK Hynix's revenue in H1 2025 and the trajectory through 2026 is likely higher given HBM4 ramp timing. If NVIDIA's GPU shipments slow materially - from U.S. export restrictions on AI chip exports to China, from hyperscaler capex pullback, or from a competitive GPU entrant - SK Hynix's revenue would fall sharply and disproportionately. The mechanism is direct: NVIDIA buys SK Hynix HBM because it is the best available. If NVIDIA orders fall 20%, SK Hynix loses roughly 5-8 percentage points of total revenue before other customer offsets. This is not a theoretical risk - NVIDIA itself disclosed in its filings that U.S. export controls on H20 chip exports to China cost it billions in revenue in 2025. Each time the U.S. government restricts NVIDIA's China business, it indirectly restricts SK Hynix's HBM demand.

Management acknowledged in the Q2 2025 concall that they were "committed to providing uninterrupted supply within regulations" regarding U.S. export controls, signaling awareness that the NVIDIA-China nexus is a live constraint.

Risk 2: China Operations - Export Control Escalation

SK Hynix's Wuxi fab produces approximately 40% of the company's total DRAM output. This fab operates on existing equipment that cannot be upgraded with EUV tools under current U.S. export rules. If annual U.S. licensing for equipment maintenance and spare parts to Wuxi is denied or restricted in future years, SK Hynix would face meaningful output losses from a fab it cannot close overnight without destroying customer relationships and losing 40% of production. The shift from blanket waivers to annual licensing in late 2025 already made this risk concrete. An escalation in U.S.-China technology tensions could make licensing renewal a political rather than technical decision.

The mechanism: equipment in a DRAM fab requires ongoing maintenance and replacement of U.S.-origin parts. Without licenses to ship those parts, fab output degrades over 12-24 months. SK Hynix cannot quickly transfer Wuxi production to Korean fabs because the wafer capacity does not exist in Korea yet - that is what M15X and Yongin are being built to address over 2-3 years.

Risk 3: HBM Technology Transition Failure

SK Hynix's MR-MUF advantage is real but not permanent. Each HBM generation requires revalidating the packaging approach. HBM4 introduces the TSMC-produced base die, adding a new integration risk between SK Hynix's DRAM stacks and TSMC's logic process. If HBM4 yields disappoint, if the TSMC CoWoS packaging throughput cannot scale, or if Samsung or Micron's HBM4 implementation proves superior, the competitive gap narrows. The risk is not that SK Hynix falls behind in one quarter - it is that the MR-MUF advantage that justified 55-62% market share proves node-specific rather than generationally durable.

Risk 4: Macro-Driven Capex Pullback by Hyperscalers

The current memory super-cycle rests entirely on hyperscale AI capex. Amazon, Microsoft, and Google collectively committed hundreds of billions of dollars to AI infrastructure through 2026-2027. If AI investment returns disappoint - if AI services do not generate enough revenue to justify the GPU cluster build-out - hyperscalers could slow or defer orders. A 10-15% reduction in hyperscaler AI capex would hit HBM demand disproportionately and cascade into server DRAM weakness. This risk is high-probability-moderate over a multi-year horizon; it is essentially the question of whether the AI infrastructure build-out is correctly sized.

Risk 5: Memory Cycle Overshoot - Overcapacity in 2028-2030

SK Hynix is committing KRW 21.6 trillion ($15 billion) to Yongin Phase 1 alone, plus the M15X investment, plus Indiana. Samsung is expanding simultaneously. Micron is investing in Idaho and New York fabs under CHIPS Act support. If all this capacity arrives within 18 months of each other and demand growth decelerates - from an AI bubble burst, from geopolitical disruption, from technological substitution - the industry could reenter the 2022-type inventory correction but from a higher capex base. The mechanism is the same as every prior memory cycle: disciplined producers eventually build too much. The question is timing.

"No one in Seoul wants to sink billions into concrete, equipment, and cleanrooms in 2026, only to wage discount wars against their own inventory again in 2028."

• Industry commentary on DRAM expansion caution, late 2025

Risk 6: Samsung HBM Recovery

Samsung's HBM market share collapsed from approximately 30-35% in 2023 to around 17% in H1 2025. This is not a structural exit - Samsung is the world's largest memory company, has equivalent engineering talent, and has committed to winning back HBM share with HBM4. If Samsung's TC-NCF yields improve materially, if it secures NVIDIA qualification for Blackwell or Rubin HBM4 at scale, and if it prices aggressively to recapture share, SK Hynix's pricing power in HBM normalizes from supernormal toward competitive. This is a medium-probability, medium-impact risk - Samsung will not disappear from HBM, but catching SK Hynix's process maturity and customer depth takes years.

SECTION 9: WALK THE TALK

Concalls used for this analysis:

1. Q2 2025 - July 24, 2025
2. Q3 2025 - October 28-29, 2025
3. Q4 2025 - January 29, 2026
4. Q1 2026 - April 23, 2026

Note: Q1 2026 was reported April 22-23, 2026, within 8 days of this report. This is within the 90-day currency requirement.

Starting from Q2 2025

The Q2 2025 concall was SK Hynix communicating from a position of confidence but still proving the HBM thesis. Management made two headline commitments. First: "the company plans to double HBM sales year-over-year." Second, regarding export controls and China supply: management committed to "providing uninterrupted supply within regulations."

By the Q4 2025 earnings, both were delivered. HBM revenue more than doubled in full-year 2025. Supply to customers including NVIDIA continued through Wuxi without interruption, aided by U.S. licensing approvals received in December 2025. The HBM doubling commitment was not aspirational - management had visibility into contracted volumes when they made it, and they delivered.

In Q2 2025, management also committed to maintaining profitability levels with HBM4's higher cost structure. They stated they were "striving to reflect cost increases in pricing strategy for HBM4" and aiming to "establish optimal pricing with customers while maintaining current profitability levels." This was a specific forward commitment. In Q4 2025, operating margins hit 58% and in Q1 2026 reached 72% - not just maintained but substantially expanded. On this promise, management undercommitted and overdelivered.

Through Q3 2025

The Q3 2025 concall set an important specific guidance: DRAM and NAND bit shipments in Q4 2025 would grow only at "low single-digit" quarter-on-quarter rates. This was cautious guidance - the market expected stronger Q4 seasonality. What happened: DRAM shipments in Q4 grew by low-single-digit percent, consistent with guidance on bit volume, but ASP surged 20% QoQ as HBM pricing strengthened. NAND shipments grew approximately 10% QoQ - slightly better than the "low single-digit" guide. So the guidance was directionally correct for DRAM and marginally conservative on NAND.

More importantly, in Q3 2025 management stated: "The HBM supply plan for next year has been finalized with major customers. Pricing has been set to sustain current profitability." This was a major commitment - it told investors that Q4 2025 and FY2026 HBM revenue was essentially contracted. The Q4 2025 results confirmed this - revenue hit a record and management reiterated full 2026 demand visibility. No quiet dropping of this commitment; in fact, they doubled down by adding that supply would remain tight into 2027.

Q4 2025 Commitments

At the January 2026 call, management guided "2026 CapEx expected to increase considerably" while maintaining "mid-30% range of revenue" - capital discipline with expansion. They committed to server DRAM shipment growth of "high teens percent" in 2026 and overall DRAM demand exceeding 20%. They also promised treasury share retirement of 50 million shares.

In Q1 2026, capex discipline and the HBM demand trajectory were both tracking to plan. The server DRAM growth guidance was consistent with Q1 2026's observed strength. The treasury share retirement commitment was reiterated.

Q1 2026 - The Most Recent and The Most Striking

The Q1 2026 concall came with results that management could not have safely guided toward just one quarter earlier: revenue of KRW 52.5 trillion in a typically slow season, 72% operating margin. CFO Kim Woo-hyun described it plainly:

"The first quarter is typically a slow season, but AI infrastructure investment created a need for more supply, and the pricing environment strengthened significantly."

What makes this credibility-enhancing is that SK Hynix did not guide Q1 2026 to be exceptional. They guided Q1 DRAM shipments conservatively (typical seasonal expectations). The fact that results massively exceeded expectations was not a guidance miss - it was a demand surprise that management could not predict. But they acknowledged the structural shift explicitly rather than taking undue credit for it.

Management also introduced new forward commitments in Q1 2026: Yongin Phase 1 acceleration to February 2027, US ADR listing targeting June-July 2026, SOCAMM2 production commenced, and the statement that "customer demand for the next three years already exceeds our supply capacity." These are testable claims. The Yongin acceleration is an engineering milestone that will either happen or be delayed. The ADR listing has a target window that will either be met or missed. These commitments will define the next two concalls' credibility narrative.

Assessment

SK Hynix management is among the more credible in the semiconductor industry across these four quarters. They made large-scale commitments (HBM doubling, profitability maintenance, 2026 supply sold out) and delivered on all of them. Where they guided conservatively (Q4 2025 DRAM bit volumes), they undershot the delivery - margins blew past guidance. They have not made promises they subsequently walked back; they have made promises and then exceeded them.

The one potential credibility test ahead: the ADR listing (June-July 2026 target) and Yongin Phase 1 (February 2027 target) are specific, datable commitments. If either slips substantially, it will be the first material guidance miss in this four-quarter window.

SECTION 10: SHAREHOLDER FRIENDLINESS INDEX

Dividends

SK Hynix paid a quarterly dividend of KRW 300 per share throughout 2022, 2023, and 2024 - a flat KRW 1,200 per share annually across three consecutive years. During 2022 and 2023, this was maintained even as the company swung to operating losses in the DRAM downturn (2022 H2 through 2023), reflecting a commitment to the fixed dividend floor regardless of near-term earnings conditions.

The payout ratio in the loss period (H2 2022, 2023) was technically negative on a net income basis - the company was paying cash dividends out of reserves while losing money. This should be read as commitment to the fixed minimum, not generosity; it also reflects that operating cash flow remained modestly positive even during the down-cycle due to working capital release.

In 2025, the dividend structure changed materially. SK Hynix announced a new shareholder return program for 2025-2027, raising the fixed annual dividend 25% to KRW 1,500 per share. Total dividends paid in 2025 reached KRW 2,429 per share, including a large KRW 1,304 payment in February 2025 that appears to encompass catch-up distribution from 2024 free cash flow accumulation, plus KRW 375 quarterly thereafter. The company announced a total cash dividend commitment of KRW 3,000 per share for fiscal year 2025, with a total payout of approximately KRW 2.1 trillion.

Source: StockAnalysis KRX:000660 dividend history; SK Hynix newsroom announcement December 2024

Share Buybacks and Treasury Share Actions

In connection with FY2025 results (announced January 2026), SK Hynix committed to retiring all remaining 50 million treasury shares it holds, excluding shares reserved for employee stock compensation plans. This represents approximately 2.1% of total shares outstanding. The market value of the 50 million shares at announcement was approximately KRW 12.2 trillion - a material capital return event.

The mechanism matters: retiring treasury shares reduces the share count permanently, concentrating future earnings per share for remaining holders. Combined with the KRW 2.1 trillion cash dividend, total shareholder returns for the FY2025 cycle approach KRW 14+ trillion, the highest in the company's history.

For 2026, the company is in a tension between reinvestment and return. CFO Kim Woo-hyun stated explicitly in the Q1 2026 concall: "Given the high return on investment demonstrated by our first quarter results, we believe that reinvesting the cash we generate back into the business is currently the best use of capital." This signals that the 2026 shareholder return cadence will prioritize the new Yongin and M15X capex over incremental buybacks or special dividends, though the fixed dividend of KRW 1,500/year continues.

At the March 2026 shareholder meeting, shareholders reportedly demanded increased dividends, and management responded by setting a target of achieving KRW 100 trillion in net cash as a milestone that would trigger accelerated shareholder returns. This creates a transparent framework: build cash through the current cycle, then distribute when a financial safety net is established.

Sources: SK Hynix newsroom; StockAnalysis dividend history; SK Hynix FY2025 results press release

Three-Year Summary

The picture that emerges: SK Hynix is not a prodigious capital returner. The base dividend is modest relative to earnings. But the company held the dividend flat through a severe earnings downturn rather than cutting it, and in the first year of the super-cycle upturn, executed the largest single shareholder return action in its history. The balance between reinvestment (Yongin, M15X, Indiana) and return is weighted toward reinvestment, which is the right call if the capex generates the returns management projects - but it means shareholders depend heavily on the company's capital allocation wisdom.

SECTION 11: SCENARIOS

Bull Case

In the bull scenario, the AI infrastructure investment cycle proves more durable and broader than even SK Hynix's management projects. Hyperscale capex continues to grow at double-digit rates through 2027-2028, driven by the proliferation of AI agents - software systems running continuous inference - that require persistent, large-memory AI servers at far greater scale than training clusters ever did. SK Hynix's supply commitments, already booked three years out according to its HBM marketing head, do not loosen. HBM4 performs as projected and NVIDIA secures the majority of its Rubin architecture GPU HBM from SK Hynix, cementing another product generation of co-development advantage.

M15X commissions on schedule in 2026, ramps cleanly, and Yongin Phase 1 begins producing first wafers in early 2027 with comparable ramp efficiency to M16. The Indiana packaging facility adds domestically-produced HBM capacity that deepens relationships with U.S. hyperscalers who prefer domestic supply chains. The U.S. ADR listing in mid-2026 is completed successfully, raising significant capital and expanding SK Hynix's global investor base, improving its valuation multiple.

In this world, Samsung's HBM4 qualification efforts by major customers remain behind schedule, Micron continues to hold single-digit to low-teens HBM share, and SK Hynix extends its position toward 2027 without meaningful share erosion. The 1cnm process rollout across Korean fabs drives the cost structure down while ASPs remain elevated due to supply scarcity. NAND recovers from its price softness as AI storage demands for enterprise SSDs absorb output from 321-layer QLC dies. The company reaches KRW 100 trillion in net cash ahead of its own target and begins a more aggressive capital return program that re-rates the stock with a broader global shareholder base.

Base Case

In the most likely path, SK Hynix executes on the roadmap as articulated in Q1 2026. M15X begins contributing meaningful output in H2 2026. HBM4 ramps to the majority of HBM revenue by late 2026. The 1cnm migration crosses 50% of Korean DRAM capacity by end of 2026 as guided. NVIDIA Rubin GPU launches drive the next HBM4 procurement wave in H2 2026-early 2027.

The supply-demand balance remains tight through 2026 and into early 2027. Pricing holds favorable for both DRAM and NAND, with HBM ASPs sustaining profitability at levels similar to 2025 rather than dramatically expanding further. Samsung begins recovering HBM market share gradually with HBM4 qualification, bringing SK Hynix's HBM share from the current ~55-60% range toward 45-50% by end of 2026, without materially disrupting the overall market's profitability since total HBM demand is growing faster than share shifts.

The US ADR listing completes in mid-2026, raising capital that funds Yongin Phase 1 construction. The Indiana facility progresses toward a late-2027 production start. China operations continue under annual U.S. licensing with no escalation. NAND margins normalize as QLC eSSD demand absorbs output.

In this base case, SK Hynix has moved from a commodity cyclical to a structurally more valuable company - one with contractually visible revenue years out, a dominant position in the most critical AI hardware component, and the capital allocation discipline to avoid over-building into the next downturn.

Bear Case

The bear scenario begins not with a single catalyst but with a convergence. First: U.S. export controls tighten further and the annual license for Wuxi is not renewed or is dramatically restricted in scope. SK Hynix's Wuxi fab, responsible for 40% of DRAM output, begins degrading as equipment maintenance is denied. The company cannot immediately backfill that capacity from Korean fabs because M15X and Yongin are still ramping. Supply disruption creates a short-term crisis of customer relationships and revenue.

Simultaneously, hyperscaler AI capex growth decelerates sharply. Not a collapse, but a deceleration from 30%+ growth to single-digit growth as hyperscalers pause to assess the return on their AI investments. NVIDIA's H1 2026 GPU orders slow from the frenetic pace of 2025. HBM demand, which has been fully committed years out, begins showing cancellation requests or push-outs. SK Hynix's revenue visibility - which management explicitly cited as a strength - becomes its liability, as take-or-pay clauses may not protect against customer renegotiation at scale.

Samsung, meanwhile, executes a clean HBM4 qualification with NVIDIA, winning an unexpectedly large share of Rubin GPU memory orders. Samsung's TC-NCF yields improve with a process fix, and the narrative that MR-MUF is generationally superior breaks down. SK Hynix's HBM market share drops from 55-60% to 35-40%, and with it goes the pricing power that has sustained extraordinary margins.

Finally, the industry's collective capex - M15X, Yongin, Samsung's new fabs, Micron's U.S. expansion - all arrive within an 18-month window in 2028. Conventional DRAM supply floods the market, pricing collapses, and the company faces the same ugly choices it faced in 2022: cutting guidance, suspending capex, and watching margins evaporate while still carrying the fixed cost of its massive new fab footprint. In this scenario, SK Hynix has built too much at the wrong moment - a mistake the memory industry has made with near-clockwork regularity.

Sources:

• SK Hynix Q4 2025 Earnings Call Transcript - Alpha Spread
• SK Hynix Q1 2026 Earnings Call - BigGo Finance
• SK Hynix Q3 2025 Earnings Call Highlights - Yahoo Finance
• SK Hynix Q2 2025 Earnings Call Highlights - Yahoo Finance
• SK Hynix FY2025 Financial Results - SK Hynix Newsroom
• SK Hynix Q1 2026 Financial Results - SK Hynix Newsroom
• SK Hynix Q3 2025 Financial Results - SK Hynix Newsroom
• SK Hynix HBM4 at TSMC Symposium - SK Hynix Newsroom
• SK Hynix 2026 Market Outlook - SK Hynix Newsroom
• SK Hynix 25% Dividend Hike Announcement - SK Hynix Newsroom
• NVIDIA Drives 27% of SK Hynix Revenue H1 2025 - TrendForce
• SK Hynix HBM - The Memory Bottleneck Powering AI - LongYield
• SK Hynix Indiana CHIPS Act Award - NIST
• SK Hynix Confidential US ADR Filing - CNBC
• SK Hynix M15X Fab Production Start - Seoul Economic Daily
• Yongin Semiconductor Cluster Investment - SK Hynix Newsroom
• SK Hynix Solidigm Intel NAND Acquisition Complete - TechSpot
• SK Hynix Holds 62% HBM - Astute Group
• SK Hynix HBM Journey to Leadership - SK Hynix Newsroom
• Global Memory Market Outlook 2025-2026 - Fortune Business Insights
• SK Hynix High-NA EUV Installation - SK Hynix Newsroom
• US Licenses Samsung SK Hynix China Equipment - Tom's Hardware
• SK Hynix Dividend History - StockAnalysis
• Memory Price Rally May Run Past 2028 - TrendForce
• SK Hynix Wikipedia History