Micron Technology (MU) -- Financial Picture

Last Updated: April 23, 2026

PART A: Structured Numbers

Current Market Data (April 23, 2026)

Price History Context: MU traded at ~$93.77B market cap at year-end 2024. By year-end 2025 it was $321B. As of April 2026 it's $543B -- a 480% increase in ~16 months. The stock has gone from ~$85 in late 2024 to $481 today.

Income Statement -- Quarterly Trend (Last 10 Quarters)

Micron's fiscal year ends in late August. The table below shows the dramatic cycle swing from trough (FY2024 Q1-Q2) through the current super-cycle peak.

The cycle in one chart: Gross margins swung from 18% (FQ1 2024) to 74.4% (FQ2 2026) in just 9 quarters. Revenue went from $4.7B to $23.9B -- a 5x increase. This is the most extreme upcycle in Micron's history.

Q3 FY2026 Guidance: Revenue ~$33.5B, gross margins ~81%, EPS ~$19.15. This would be another massive step-up -- the guidance implies $33.5B in a single quarter, which exceeds Micron's entire FY2022 annual revenue ($30.8B).

Annual Income Statement Summary

Key Observation: Before this cycle, Micron's peak annual EPS was $11.50 (FY2018). The FY2026 consensus is $58.96, and Q3 2026 guidance alone implies ~$19/share in a single quarter. This cycle is 5x the magnitude of any previous cycle. The difference: HBM.

Revenue by Segment

New Reporting Structure (effective Q4 FY2025)

Micron reorganized into four new segments in May 2025:

1. CMBU (Cloud Memory Business Unit): HBM + hyperscale cloud DRAM
2. CDBU (Core Data Center Business Unit): OEM data center DRAM + all data center storage/SSDs
3. MCBU (Mobile & Client Business Unit): smartphone DRAM/NAND + PC client DRAM/NAND
4. AEBU (Automotive & Embedded Business Unit): automotive + industrial + consumer embedded

Segment Revenue (Q1 FY2026 -- new structure, from earnings call)

Note: Q2 FY2026 segment breakdown not available from public sources at time of writing ($23.86B total). Given the 75% sequential revenue increase, CMBU likely grew substantially to an estimated $10-12B range (driven by HBM pricing surge).

Legacy Segment Revenue (old CNBU/MBU/SBU/EBU -- annual)

Revenue by Product Type

Balance Sheet (Quarterly)

Key Balance Sheet Observations:

• Cash surged from $9.7B to $13.9B in a single quarter (Q1 to Q2 FY2026) despite massive capex -- that's the power of 74% gross margins on $24B revenue.
• Long-term debt is being paid down: $14.0B -> $11.2B -> $9.6B over the last 3 quarters. Micron is deleveraging rapidly.
• PP&E is growing fast ($38B -> $52B in 2 years) reflecting aggressive fab buildout.
• Equity has nearly doubled from $44B (Q2 FY2024) to $72B (Q2 FY2026) -- retained earnings accumulating rapidly.
• Net debt position: $9.6B debt - $14.6B cash = net cash of ~$5B. Micron is now net cash positive.

Cash Flow Statement (Quarterly)

Annualized run rates (last 2 quarters x2): OCF ~$40B, Capex ~$24B, FCF ~$17B.

FY2026 Capex Guidance: ~$20B (revised up from $18B), prioritizing HBM capacity and 1-gamma DRAM expansion.

Capex Plans & Fab Investments

Active Construction Projects

CHIPS Act Funding

Micron received up to $6.165 billion in direct funding from the CHIPS and Science Act, plus a separate $275M for Virginia. This supports the Idaho and New York facilities. The goal: grow U.S. advanced memory manufacturing from <2% to ~10% of global capacity by 2035. Micron has committed ~$50B of its own investment before decade's end for the Idaho and New York projects.

Japan Government Subsidies

Japan's METI will provide up to JPY 500B (~$3.2B) in subsidies for the Hiroshima HBM fab. This is part of Japan's JPY 5.7T semiconductor investment strategy.

Total Capital Spending Trajectory

Dividend & Capital Return

Micron is allocating virtually all capital to fab construction and HBM capacity expansion. The dividend is token -- symbolic rather than meaningful. Share buybacks are negligible. This makes sense: when you can invest $1 in HBM capacity and generate $3-5 in revenue, buying back stock is a bad use of capital.

PART B: What the Numbers Really Mean

1. The ASP x Volume Reality

Memory companies' earnings are almost entirely determined by average selling prices (ASPs) multiplied by bit shipments. In Q2 FY2026, the numbers are extraordinary:

• DRAM ASPs rose ~65% sequentially while bit shipments grew only ~5%
• NAND ASPs rose ~78% sequentially while bit shipments grew only ~2%

This means virtually all of the revenue growth is from PRICING, not volume. The 18 percentage point sequential gross margin expansion (from 56% to 74.4%) is driven almost entirely by ASP increases. This is the defining feature of a memory upcycle -- and also the warning sign. ASPs can fall just as fast as they rise.

Current ASP Context: Server DRAM prices surged 60-70% in Q1 CY2026 vs Q4 CY2025. Cumulative increases may nearly double server DRAM costs by mid-2026. This is driven by the HBM wafer cannibalization effect: HBM production consumes ~3 standard DRAM wafers per HBM wafer (and this ratio is increasing toward 4:1 with HBM4). As more DRAM wafer capacity shifts to HBM, conventional DRAM supply tightens, and ALL DRAM prices rise.

2. HBM Revenue and Margin Impact

HBM is the single most important variable in Micron's story.

HBM Revenue Trajectory:

• FY2023: Negligible (<$500M)
• FY2024: ~$2B (started ramping in second half)
• Q4 FY2025 alone: ~$2B
• Q1 FY2026: Estimated $3-4B (within the $5.3B CMBU segment, which also includes non-HBM cloud DRAM)
• Q2 FY2026: Estimated $8-10B (CMBU segment likely ~$10-12B)
• Full FY2026E: $25-30B (our estimate, representing 25-30% of total revenue)

HBM Pricing vs Conventional DRAM: HBM commands 5-10x the price per gigabyte compared to standard DDR5 DRAM. An 8-high HBM3E stack requires ~3 DRAM wafers worth of silicon plus advanced packaging. The premium reflects both the higher silicon content and the extreme manufacturing complexity (TSV drilling, die stacking, underfill, micro-bump bonding, testing).

HBM Margin Profile: Management has not disclosed segment-level margins, but the evidence strongly suggests HBM margins are accretive to overall margins. The fact that gross margins exploded from 44.7% to 74.4% as HBM revenue scaled suggests HBM margins are well above 70% -- possibly 80%+. CFO acknowledged current margins are "not sustainable" at these levels, but even normalized HBM margins are likely 50-60%, well above conventional DRAM at 30-40%.

HBM Market Share: Micron targets 23-24% HBM market share, roughly in line with its overall DRAM share (~25%). SK Hynix leads with 53-62%, Samsung at 15-25%. Micron has confirmed shipping HBM3E 12-high to NVIDIA Blackwell and AMD MI350 platforms. Micron was a late entrant but has achieved competitive parity: its HBM3E has a claimed 20% power efficiency advantage over competitors' HBM3E 8-high products.

3. Inventory Analysis -- The Cycle Indicator

Inventory is the canary in the memory coal mine. Rising inventory days = oversupply building = downturn approaching.

Analysis: Days of inventory has been DECLINING from 171 days (FQ2 2024) to 123 days (FQ2 2026). This is extremely bullish -- it means Micron is selling through inventory faster than it can produce. At 123 days, inventory is LEAN by historical standards. Micron's 5-year median is around 123 days (per GuruFocus), meaning current inventory is at the median -- not overstocked.

The fact that absolute inventory ($8.3B) has remained flat while revenue has tripled means Micron is drawing down inventory relative to sales. In past cycles, the turn happens when inventory builds start exceeding revenue growth. We are not there yet.

However, HBM has a confounding effect: HBM inventory is pre-sold (customers have committed to buy it), so the "available" inventory is even leaner than the headline number suggests.

4. Capex/Depreciation Ratio -- Growth vs Maintenance

Analysis: Micron is spending 2x depreciation, meaning roughly half of capex is growth investment. The $20B FY2026 capex is focused on: (1) HBM capacity expansion (Singapore packaging, Japan fab), (2) 1-gamma DRAM node conversion, (3) Idaho and New York fab construction for future supply. Management has noted the Idaho fab won't contribute meaningful bits until FY2027, and New York not until FY2028+.

This heavy capex cycle means Micron is building capacity for the NEXT wave of demand, not the current one. This is forward-looking capital deployment, betting that AI-driven demand will continue to grow. The risk: if demand disappoints or pricing collapses, this capex becomes stranded investment and depresses returns.

5. Historical Earnings Power: Peak, Mid-Cycle, Trough

The current cycle is categorically different from all prior cycles. Even the previous peak (FY2018, $30.4B revenue) is less than one quarter's revenue at the current run rate. HBM has added an entirely new demand vector that didn't exist before. The question is whether HBM makes the CYCLE different (permanently higher mid-cycle earnings) or just makes the PEAK higher (bigger amplitude but still mean-reverting).

My assessment: it's a combination. Mid-cycle earnings power is structurally higher because HBM revenue will persist even in a conventional DRAM downturn. But peak margins of 74-81% are unsustainable. A reasonable estimate for "new mid-cycle" earnings:

• Revenue: $50-60B (HBM $15-20B + conventional $35-40B)
• Gross margin: 35-45%
• Net income: $10-15B
• EPS: $9-13
• This implies the stock at $481 trades at 37-53x "new mid-cycle" earnings. That is expensive.

6. What is the Floor?

Tangible Book Value: Stockholders' equity of $72.5B minus goodwill and intangibles (~$4-5B) = ~$68B tangible equity = ~$60/share.

But tangible book value is misleading for a company in a capex upsurge. The real question is: what is the replacement cost of Micron's assets?

Replacement Cost Analysis:

• PP&E (net): $52B, but this is depreciated. The gross PP&E is likely $80B+. To replicate Micron's fab network from scratch would cost well over $100B given current construction costs ($15-20B per leading-edge DRAM fab).
• IP and process knowledge: Priceless. Decades of 1-beta/1-gamma DRAM and NAND process development can't be replicated.
• Customer qualifications: Years of qualification testing with NVIDIA, AMD, hyperscalers. Can't be bought.

Practical Floor: In past cycle troughs, MU has bottomed at approximately 1x tangible book value:

• 2023 trough: ~$49/share, book value ~$39/share (1.3x)
• 2019 trough: ~$30/share, book value ~$28/share (1.1x)
• 2016 trough: ~$10/share, book value ~$15/share (0.7x)

Current book value: $64/share. If the stock were to fall to 1x book, that's $64. But book value is growing rapidly (was $54 last quarter, now $64). By the time a trough arrives (2027-2028?), book value could be $80-100/share.

The floor is probably $80-120/share in a downturn scenario, given higher tangible book value from accumulated earnings and the structural HBM premium. This is meaningfully higher than the v2 analysis estimate of $30-50, which did not account for the rapid book value buildup.

Summary Financial Snapshot

Micron Technology (MU) -- Business Understanding

Last Updated: April 23, 2026

What Micron Does

Micron Technology is one of only three companies in the world that manufactures DRAM and NAND flash memory semiconductors at scale. The other two are Samsung Electronics (South Korea) and SK Hynix (South Korea). This is a tight oligopoly with combined 95%+ market share in DRAM and ~50%+ in NAND (NAND has additional players: Western Digital, Kioxia, and several Chinese entrants).

Micron is headquartered in Boise, Idaho. It has ~48,000 employees and operates fabs in the US (Boise, Manassas VA), Japan (Hiroshima), Singapore, and Taiwan.

Business Segments (Weighted by Future Cash Flow)

Micron reorganized its reporting structure in May 2025, moving from the old CNBU/MBU/SBU/EBU segments to four new market-focused units. The reorg was explicitly designed to highlight AI/data center exposure.

1. CMBU -- Cloud Memory Business Unit [~45-50% of Future Cash Flow]

What it does: Sells memory solutions to large hyperscale cloud customers (Microsoft Azure, Google Cloud, AWS, Meta, etc.) and manufactures High Bandwidth Memory (HBM).

Revenue: Q1 FY2026: $5.3B (+100% YoY). Estimated Q2 FY2026: $10-12B. This is the fastest-growing segment by a wide margin.

Why it matters most: This segment contains HBM, which is the highest-margin, highest-growth product in the entire memory industry. Every NVIDIA GPU (H100, B200, GB200, and future Rubin chips) requires multiple HBM stacks. An NVIDIA B200 GPU uses 8 stacks of HBM3E 12-high, totaling 192GB. A single GB200 NVL72 rack contains 36 GPUs and ~6.9TB of HBM. At current HBM pricing (~$20-30/GB), the HBM content alone in an NVL72 rack is $140,000-$200,000.

HBM Deep Dive:

HBM is a 3D-stacked DRAM architecture using Through-Silicon Vias (TSVs) to vertically connect multiple DRAM dies. Current products:

HBM Manufacturing Complexity:

• The wafer trade ratio is 3:1 (one HBM wafer consumes the equivalent of 3 standard DRAM wafers), increasing to 4:1 with HBM4 16-high stacking. This means every HBM wafer produced REMOVES 3-4 conventional DRAM wafers from supply.
• Manufacturing requires: (1) advanced DRAM fab using 1-beta or 1-gamma node, (2) TSV drilling, (3) die thinning to <30um, (4) micro-bump bonding, (5) underfill, (6) advanced packaging to mount on a logic base die. The base die for Micron's HBM4 uses a 1-beta DRAM node (SK Hynix outsources its base die to TSMC).
• Yields on HBM are lower than conventional DRAM due to the stacking complexity and the fact that one bad die can destroy an entire stack.

Micron's HBM Competitive Position:

NVIDIA Relationship: This is critical. Micron is qualified for NVIDIA's Blackwell platform (B200/GB200). NVIDIA's HBM qualification process is notoriously rigorous -- it takes 6-12 months of testing. Being qualified means Micron has locked-in demand for the current generation. For Rubin (expected 2026-2027), Micron will need to requalify, but its existing relationship and demonstrated technical capability make this likely.

Micron has signed its first five-year Strategic Customer Agreement (SCA) in Q2 FY2026, committing to both volume and pricing terms. This is a major departure from the traditional spot-market memory business model and suggests Micron is locking in long-term HBM demand at favorable prices.

Moat Assessment (CMBU): STRONG. The moat is physical and technical:

1. Only 3 companies can make advanced DRAM at all (cost of entry: $15-20B per fab, 5-10 years to develop process technology)
2. HBM adds additional complexity (TSV, stacking, packaging) that narrows the field further
3. Customer qualification with NVIDIA/AMD takes 6-12 months
4. Multi-year supply agreements create switching costs
5. The wafer trade ratio creates a natural supply constraint -- you literally cannot make enough

2. CDBU -- Core Data Center Business Unit [~20-25% of Future Cash Flow]

What it does: Sells memory solutions to OEM data center customers (not hyperscalers -- think Dell, HPE, Lenovo servers) and ALL data center storage solutions (enterprise SSDs).

Revenue: Q1 FY2026: $2.4B (+4% YoY). This is the slowest-growing segment, which is telling.

Why it matters: Enterprise SSDs (especially PCIe Gen5/Gen6 NVMe drives using 232-layer NAND and the G9 NAND node) are a growing category. Data center NAND revenue exceeded $1B in Q1 FY2026. However, the OEM server DRAM business is more competitive and lower-margin than hyperscale.

Moat Assessment (CDBU): MODERATE. Enterprise SSDs benefit from Micron's vertically integrated NAND + controller stack, which gives performance and power advantages. But this segment is more competitive than CMBU -- Western Digital, Samsung, and Kioxia all compete here.

3. MCBU -- Mobile & Client Business Unit [~15-20% of Future Cash Flow]

What it does: Sells DRAM and NAND to smartphone OEMs (Apple, Samsung, Xiaomi, etc.) and PC OEMs (Dell, HP, Lenovo).

Revenue: Q1 FY2026: $4.3B (+63% YoY).

Key dynamics:

• Mobile DRAM is the second-largest DRAM category after server DRAM
• Content per phone is growing (8GB -> 12GB -> 16GB DRAM, 128GB -> 256GB -> 512GB NAND) driven by on-device AI features
• But smartphone unit volumes are mature/declining -- growth comes from content increase only
• PC market is benefiting from "AI PC" refresh cycle (PCs with dedicated NPUs require more DRAM)

Moat Assessment (MCBU): WEAK-MODERATE. This is more commodity-like. Apple, Samsung, and Qualcomm have significant bargaining power. Micron competes primarily on cost and technology node. LPDDR5X is the current product; the transition to LPDDR6 (with CAMM2 form factor for PCs) is upcoming.

4. AEBU -- Automotive & Embedded Business Unit [~10-15% of Future Cash Flow]

What it does: Sells memory solutions for automotive (ADAS, infotainment, autonomous driving), industrial IoT, and consumer electronics.

Revenue: Q1 FY2026: $1.7B (+49% YoY).

Key dynamics:

• Automotive is the highest-margin end market for conventional DRAM/NAND because automotive qualification requirements (AEC-Q100) create significant barriers and pricing power
• Content per vehicle is growing rapidly: from <4GB DRAM + 32GB NAND in 2020 to 16-32GB DRAM + 256GB-1TB NAND in premium EVs (2026)
• Autonomous driving is a long-term tailwind -- Level 4/5 autonomy could require 64-128GB DRAM per vehicle
• Growth is steady but the TAM is small relative to data center

Moat Assessment (AEBU): MODERATE-STRONG. Automotive qualification cycles are 2-3 years, creating significant switching costs. Safety-critical applications demand extreme reliability. But Samsung and SK Hynix also compete here.

Technology Position

DRAM Technology Roadmap

Micron's 1-gamma node uses EUV lithography (from ASML) and delivers significant die shrinkage, enabling more bits per wafer and lower cost per bit. Management claims yields matured 50% faster than any prior node transition, suggesting strong execution.

NAND Technology Roadmap

Micron's vertically integrated NAND controller + flash stack (Micron designs its own SSD controllers) gives it a system-level advantage in enterprise SSDs, particularly for AI training workloads that require high sequential write bandwidth.

Key Technology Advantages

1. Power Efficiency: Micron claims a 20% power advantage in HBM3E vs competitors. In data centers where power is the bottleneck, this is a real differentiator. For HBM4, Micron claims an additional 20% power reduction.

1. 1-gamma DRAM yield leadership: Achieving mature yields 50% faster than prior nodes translates directly to lower cost per bit and higher profitability.

1. Vertically integrated SSD stack: Micron designs its own NAND controllers, which most competitors (except Samsung) do not. This enables optimized firmware and better performance.

Manufacturing Footprint & Geopolitical Risk

Fab Locations

Key Geopolitical Risk: Micron's Taiwan fab represents a meaningful portion of its DRAM output. A China-Taiwan conflict would disrupt this capacity. However, Micron has been actively diversifying AWAY from Taiwan toward the US, Japan, and Singapore -- the fab construction spree is partly a de-risking strategy.

China Risk: Micron was banned from critical Chinese infrastructure in May 2023 by the Cyberspace Administration of China, citing "national security risks." This cost Micron roughly 10-15% of revenue at the time. The ban remains in place. Chinese memory makers (CXMT, Yangtze Memory/YMTC) are attempting to enter the market but remain 2-3 generations behind in DRAM and face export control restrictions on EUV equipment.

India Expansion

Micron is building an assembly and test facility in Gujarat, India -- not a fab, but a back-end packaging facility. Investment of ~$2.75B. This diversifies packaging away from East Asia.

Competitive Landscape

DRAM Market Share (Q1 CY2025)

SK Hynix is the HBM leader and has been NVIDIA's primary HBM supplier since HBM2. It overtook Samsung as the world's largest memory company by revenue in 2025, entirely because of HBM. SK Hynix's stock has outperformed even Micron.

Samsung fumbled the HBM transition. It was late to qualify HBM3E with NVIDIA and lost significant market share to both SK Hynix and Micron. Samsung is now aggressively investing in HBM4 to recover share.

Micron was the last to enter HBM but has executed well. It went from near-zero HBM share to ~24% in approximately 18 months. The five-year SCA signed in Q2 FY2026 suggests customers view Micron as a credible long-term HBM supplier.

NAND Market Share (CY2025)

Micron is the smallest major NAND player. NAND is more competitive than DRAM (5+ serious players vs 3 in DRAM). Micron's NAND strategy focuses on enterprise SSDs where its vertically integrated controller stack creates differentiation, rather than competing in commodity consumer NAND.

Cost Structure: Is Micron the Lowest-Cost Producer?

In conventional DRAM: Micron is competitive with SK Hynix and Samsung on a cost-per-bit basis. The transition to 1-gamma (smaller die = more bits per wafer) is a cost reduction lever. Micron claims its 1-gamma yields are ramping faster than any prior node. However, Samsung historically has had a cost advantage from its sheer scale (larger wafer starts), and SK Hynix has been the efficiency leader in recent years.

In HBM: Cost structure is harder to assess because HBM pricing is so high relative to costs. The key variable is YIELD. A low-yield HBM process wastes 3x the silicon of a conventional DRAM process failure. SK Hynix, having started earlier, likely has a yield advantage. But Micron claims its HBM3E yields are reaching competitive levels.

The honest answer: Micron is probably the third-lowest-cost DRAM producer behind SK Hynix (#1, benefiting from HBM first-mover yield advantages) and Samsung (#2, scale advantage). In NAND, Micron is also third or fourth. But in the current environment of extreme pricing power, cost structure matters far less than it does in a downturn. In a downturn, the highest-cost producer suffers most. Micron has historically been the most cyclically punished -- its stock falls more in downturns than Samsung or SK Hynix (partly because Samsung and SK Hynix have diversified revenue streams).

Capital Allocation Philosophy

Micron's capital allocation priorities, in order:

1. Invest in capacity and technology -- consuming the vast majority of capital ($20B/yr capex)
2. Maintain investment-grade balance sheet -- deleveraging debt (paid down $4.5B in FY2026 so far)
3. Token dividend -- $0.60/year, 0.12% yield, purely symbolic
4. Minimal buybacks -- net share count essentially flat

This is the right priority ordering for a company in Micron's position. When you can invest $1 in HBM capacity and generate $3-5 in revenue at 70%+ margins, every dollar of buybacks is a dollar of forgone investment. The time for buybacks is during the downturn when capex needs shrink and stock is cheap.

Historical Pattern: Micron has a poor track record on buybacks -- it tends to buy back stock near peaks (when cash is plentiful and it "feels" right) and not buy during troughs (when cash is scarce but the stock is cheap). This is a common corporate finance mistake. However, the company deserves credit for currently prioritizing investment over returns during the boom.

Key Risks Specific to the Business

1. Cyclicality: Memory is the most cyclical part of the semiconductor industry. Revenue has dropped 50% in a single year (FY2022 -> FY2023). This WILL happen again, though the timing is uncertain and HBM may reduce the amplitude.

1. Customer Concentration: Hyperscalers (AWS, Google, Microsoft, Meta) and GPU makers (NVIDIA, AMD) represent a large and growing share of revenue. These are sophisticated buyers with immense bargaining power.

1. Technology Execution Risk: Falling behind on technology nodes (1-gamma, HBM4, HBM4E) would be catastrophic. A single generation of yield problems could cede market share that takes years to recover.

1. China: The China ban costs Micron 10-15% of revenue. If the ban expands or if Chinese DRAM makers (CXMT) successfully ramp competitive products, this is a headwind.

1. Taiwan Exposure: A meaningful portion of production is in Taiwan, which faces geopolitical risk from China. Micron is diversifying but this takes years.

1. Capex Overcommitment: $20B/yr capex in a boom creates fixed costs that become painful in a downturn. Fab construction can't be easily stopped once started.

1. HBM Commoditization Risk: As all three memory makers ramp HBM capacity, HBM could gradually commoditize like DDR DRAM before it. The five-year SCAs push back against this, but pricing power will inevitably moderate.

Micron Technology (MU) -- Future Prediction

Last Updated: April 23, 2026

THE CRITICAL QUESTION: Where Are We in the Memory Cycle?

This is the most important question for any Micron investment thesis. Memory is the most cyclical business in all of technology. Getting the cycle timing right is worth more than any other analysis.

Evidence that we are AT OR NEAR the cycle peak:

1. Gross margins at 74.4% with guidance to 81%. Micron's 20-year average gross margin is ~30%. The previous peak was 59% (FY2018). We are 22 percentage points ABOVE the prior all-time peak. This is not normal.

1. ASPs surging on constrained supply. DRAM ASPs up 65% sequentially, NAND up 78%. This is driven by the HBM wafer cannibalization effect (HBM consuming 23%+ of DRAM wafers, up from near-zero). The question: does this supply constraint persist or does new capacity relieve it?

1. Revenue growth is almost entirely pricing, not volume. DRAM bit shipments grew only 5%, NAND only 2%. In Q2 FY2026, Micron earned $23.9B on similar volume to what produced $8B a year earlier. That's pure pricing. Pricing can reverse.

1. Management explicitly acknowledged margins are "not sustainable." When the CFO says this, listen.

1. Massive capex cycle underway. $20B/year from Micron alone, plus SK Hynix expanding from $85B to $410B investment in Yongin, plus Samsung building HBM capacity. Supply is being added aggressively.

1. Analyst euphoria. "Memory supercycle," "structural upcycle," "this time is different" -- these phrases appear in virtually every analyst report. The last time analysts were this bullish on memory was Q3 2018. The stock peaked at $64 and fell to $28.

Evidence that this cycle may have more runway than usual:

1. HBM is structurally different from past demand drivers. In previous cycles, the demand driver was PCs (1990s-2000s), smartphones (2010s), or servers (2017-2018). AI/HBM is a genuinely new demand vector with multi-year visibility -- NVIDIA's GPU roadmap extends through 2028 (Blackwell -> Rubin -> next), and each generation requires MORE HBM per GPU.

1. HBM supply is pre-sold. Micron's HBM capacity for CY2025 and CY2026 is fully booked. Five-year SCAs have been signed. This is NOT spot market exposure -- it's contracted revenue.

1. Hyperscaler capex is still accelerating. Combined hyperscaler capex is ~$660-690B in 2026, nearly doubling from ~$380B in 2025. There are no signs of a pullback.

1. The wafer trade ratio creates a self-reinforcing shortage. As more wafers go to HBM (3:1 to 4:1 ratio), conventional DRAM supply shrinks, which pushes up ALL DRAM prices. This feedback loop tightens supply across the entire memory market.

1. OpenAI's Stargate commitment: 900,000 DRAM wafers per month through 2029 -- representing ~40% of total global DRAM output. This is an unprecedented demand commitment.

My Assessment: We are in the late stages of the most extreme memory upcycle in history, but the peak may be further away than in a traditional cycle. I estimate 2-4 more quarters of very strong results (through end of CY2026) before any meaningful pricing correction. However, the magnitude of the eventual correction is highly uncertain.

2-Year Outlook (Pre-AGI, through 2028)

DRAM Pricing: 2026-2028

Near-term (Q3 FY2026 through end CY2026):

• Server DRAM prices have already surged 60-70% QoQ in early CY2026 and could nearly double by mid-2026.
• HBM pricing remains firm due to pre-sold supply agreements and capacity constraints.
• The HBM wafer cannibalization is tightening conventional DRAM supply, supporting DDR5 pricing.
• My estimate: DRAM ASPs increase another 10-20% through CY2026 H2, then plateau.

Medium-term (CY2027):

• New fab capacity begins coming online: Micron Idaho (late 2026/early 2027), Hiroshima new fab (2028), Samsung Pyeongtaek expansion, SK Hynix Yongin expansion.
• The critical question: does new supply outrun demand growth?
• HBM demand should continue growing (NVIDIA Rubin GPUs, custom ASIC chips from hyperscalers), but at some point supply catches up.
• Conventional DRAM may see some pricing weakness as HBM capacity expansion allows some wafers to return to DDR5/LPDDR5X production.
• My estimate: DRAM ASPs plateau in H1 CY2027, then begin declining 5-15% in H2 CY2027.

Tail-end (CY2028):

• Full impact of new fab capacity felt. Three major DRAM makers all ramping HBM4/HBM4E with more efficient processes.
• If AI capex slows even modestly, the impact on memory pricing would be amplified by the new supply.
• Risk of conventional DRAM oversupply is real -- the mobile and PC markets are mature, and server DRAM demand may moderate as hyperscaler capex growth decelerates.
• My estimate: DRAM ASPs decline 15-30% from peak through CY2028. This would still leave ASPs well above 2024 levels.

NAND Pricing: 2026-2028

NAND is a different story from DRAM. The NAND market has more competitors (5-6 players vs 3 in DRAM), and the QLC transition (4 bits per cell vs 3 for TLC) is dramatically increasing bits per wafer. NAND is structurally more prone to oversupply.

Current situation: NAND ASPs are surging (+78% QoQ in Q2 FY2026), partly because NAND production was cut aggressively during the 2023 downturn and partly because enterprise SSD demand for AI training is strong. But this is likely more cyclical than structural.

My estimate: NAND ASPs decline 20-40% from current peaks by CY2028 as QLC adoption increases supply per wafer and Chinese NAND makers (YMTC) regain some market access.

HBM Demand and Pricing: 2026-2028

HBM TAM Forecast:

• CY2025: ~$35B (Micron's estimate)
• CY2026: ~$55-65B
• CY2027: ~$75-85B
• CY2028: ~$100B (Micron's estimate)

This represents a ~40% CAGR through CY2028. The key drivers:

1. NVIDIA GPU roadmap: Blackwell (2024-2025) -> Rubin (2026-2027) -> Next-gen (2028+). Each generation increases HBM content per GPU.
2. Custom ASIC proliferation: Google TPU, Amazon Trainium, Microsoft Maia -- all require HBM. The custom chip market is growing faster than NVIDIA's.
3. Inference scaling: As AI models are deployed at scale, inference workloads grow, each requiring HBM-equipped GPUs or accelerators.

HBM Pricing Risk: HBM prices are currently ~$20-30 per GB, representing 5-10x conventional DRAM. As all three memory makers ramp HBM capacity, this premium will compress. My estimate: HBM pricing declines 15-25% by CY2028, but volume growth more than compensates, so total HBM revenue continues growing.

Micron EPS Trajectory: FY2026-FY2028

Consensus: FY2026 EPS $58.96, FY2027 EPS $99.94 (but with enormous range: $53.71 to $149.60).

My base case:

• FY2026: Revenue $115B, gross margin ~68%, EPS ~$60
• FY2027: Revenue $140B, gross margin ~55%, EPS ~$55 (margins compress even as revenue grows)
• FY2028: Revenue $120B, gross margin ~45%, EPS ~$35 (beginning of cyclical normalization)

My bull case:

• FY2027: Revenue $180B, gross margin 60%, EPS ~$80
• FY2028: Revenue $200B+, gross margin 55%, EPS ~$80+
• This requires: no memory downturn, continued hyperscaler capex acceleration, HBM pricing holds

My bear case:

• FY2027: Revenue $110B, gross margin 40%, EPS ~$25
• FY2028: Revenue $80B, gross margin 30%, EPS ~$10
• This requires: memory cycle turns, DRAM oversupply, hyperscaler capex slows, HBM pricing collapses

Risk of a Memory Downturn

The pattern is unmistakable. Every memory cycle peak in history has been followed by a severe downturn:

• FY2018 peak ($30.4B revenue) -> FY2023 trough ($15.5B) = -49%
• FY2022 peak ($30.8B) -> FY2023 trough ($15.5B) = -50%
• 2014 peak ($16.4B) -> 2016 trough ($12.4B) = -24%

The question is whether HBM breaks this pattern. My assessment: HBM reduces the amplitude of the downturn but does not eliminate cyclicality. Even if HBM revenue is stable ($20-30B/year sustained by long-term contracts), conventional DRAM and NAND will still cycle. If conventional memory revenue drops 30-40% from peak while HBM holds, total revenue might fall 15-25% rather than 50%. That is still a painful downturn, but less severe than historical patterns.

Timing the downturn: The earliest realistic timing for a downturn is late CY2027, when new fab capacity comes online and if hyperscaler capex growth decelerates. The latest realistic timing is CY2029. The consensus expectation of "no downturn ever because AI" is, in my view, wrong. Supply response is inevitable when margins are 80%.

Where in the Cycle Are We?

Late-cycle upcycle. The unprecedented margins (74-81%) and management's own admission that they're "not sustainable" confirm this. But "late cycle" can last several quarters when the demand driver (AI infrastructure buildout) has genuine multi-year duration. I estimate we're in inning 7 of a 9-inning game. The last 2 innings could still be very profitable, but the risk of a sudden end is growing.

5-10 Year Outlook (Post-AGI)

How Does AGI Affect Memory Demand?

AGI is coming within the next 2-3 years. By 2028, we will likely have automated AI researchers capable of recursive self-improvement. Here's what this means for memory:

The bullish case for memory in an AGI world is extremely strong:

1. Every inference requires memory. Unlike compute (where algorithmic efficiency can reduce FLOPS needed), memory bandwidth is a physical bottleneck. You cannot run a large language model without enough HBM to hold the model weights. A 1-trillion-parameter model requires ~2TB of HBM just for weights at FP16 precision.

1. Recursive self-improvement scales memory demand exponentially. As AI models get larger and more capable, they require more memory. GPT-4 class models use ~1.8T parameters. GPT-5 and beyond will be larger. Each generation of model requires proportionally more HBM.

1. The number of deployed AI models will explode. Post-AGI, every company, every device, every service will run AI inference. The total installed base of AI accelerators (GPUs, TPUs, custom ASICs) will grow 10-100x from current levels. Each accelerator needs HBM.

1. Edge AI requires memory. AI-capable smartphones, laptops, robots, autonomous vehicles, and IoT devices all need more DRAM and NAND. The content-per-device growth trend accelerates post-AGI.

1. Robotics requires memory. When robotics is solved (likely 2029-2031), every robot will need onboard memory for inference. Millions, then billions of robots = massive memory demand.

Does the Memory Oligopoly Survive AGI?

Yes, with high confidence. The memory oligopoly is one of the most durable physical moats in technology:

1. The capital barrier is $100B+. Building a competitive DRAM fab from scratch requires $15-20B in construction, $5-10B in equipment, and 5-10 years of process development. The total cost to become a competitive DRAM manufacturer is easily $100B over a decade. No new entrant has successfully entered the DRAM market in 30 years.

1. The knowledge barrier is decades of accumulated process IP. DRAM manufacturing requires precise control of atomic-scale processes (EUV lithography, atomic layer deposition, high-aspect-ratio etching). This knowledge is proprietary and can't be simply purchased or reverse-engineered, even by an AGI.

1. Equipment constraints. Leading-edge DRAM fabs require EUV lithography machines from ASML, which has a multi-year backlog. Even if an AGI designed a better memory technology, it would need physical manufacturing infrastructure that takes years to build.

1. AGI companies need memory manufacturers, not the reverse. OpenAI, Anthropic, Google, and Microsoft are the customers. They cannot vertically integrate into memory manufacturing -- it would cost $100B+ and take 10 years, and they have more valuable things to do with that capital and time. They NEED Samsung, SK Hynix, and Micron.

1. China's attempts prove how hard this is. China has been trying to build a competitive DRAM industry (CXMT) with billions in government subsidies since 2015. After 10+ years and $50B+ in investment, CXMT is still 2-3 generations behind. And they're now blocked from EUV equipment by export controls.

The risk: An AGI could potentially design a fundamentally new memory technology (neuromorphic memory, phase-change memory, RRAM, or something we haven't conceived of). However:

• New memory technologies have been "5 years away" for 20 years. None has displaced DRAM or NAND at scale.
• Even if a new technology is designed, manufacturing it at scale would require new fabs and process development -- a decade of work.
• DRAM and NAND will be the dominant memory technologies through at least 2035, probably 2040.
• The existing manufacturers would be best positioned to adopt any new technology anyway, given their fab infrastructure and process engineering talent.

Conclusion: The memory oligopoly is among the most AGI-resilient businesses in the world. It is a physical moat protected by atoms, not bits. Billions of geniuses in data centers cannot conjure a fab out of thin air.

Long-Term Structural Demand for HBM

HBM is not going away. The fundamental physics of AI inference require memory bandwidth close to the compute die. The Von Neumann bottleneck (data movement between memory and processor being the performance constraint) gets worse as models get larger and inference requirements grow. HBM solves this by putting memory physically on top of the processor.

Long-term HBM demand drivers:

• CY2025: ~$35B TAM
• CY2028: ~$100B TAM (Micron estimate)
• CY2030: ~$150-200B TAM (my estimate, driven by post-AGI infrastructure buildout)
• CY2035: Potentially $300B+ if robotics and edge AI adopt HBM-like architectures

Could AGI design a better memory technology?

Yes, eventually. An AGI could potentially design:

• More efficient 3D stacking technologies (reducing the wafer trade ratio)
• Hybrid memory/compute architectures (processing-in-memory)
• New memory materials (emerging non-volatile memories)

But even if designed in 2028, manufacturing at scale would take until 2033-2035. And the existing memory manufacturers would likely be the ones to manufacture it. The moat is the fab, not the specific DRAM cell design.

10x Framework

Bull Case 5-Year EPS

If the AI infrastructure buildout continues through 2031 (which it almost certainly will given the AGI timeline), and HBM demand grows to $150B+, and Micron maintains its ~24% share:

• FY2031 Revenue: $100-120B (HBM $35-45B + conventional $65-75B)
• Gross margin: 40-50% (normalized from current peak, but structurally higher than historical 30%)
• Net income: $30-45B
• EPS: $26-40

Historical Peak Multiple

Memory stocks trade at different multiples depending on cycle position:

• Trough: 1-2x book value, negative P/E (losses)
• Early recovery: 15-25x P/E
• Mid-cycle: 8-15x P/E
• Late-cycle/peak: 5-8x P/E (low P/E because earnings are peaking)

At the current $481 stock price and ~$60 FY2026 EPS, MU trades at ~8x current-year earnings. The forward P/E of 5.2x uses FY2027 consensus, which assumes continued earnings growth.

The market is telling you that it expects earnings to DECLINE from these levels -- that's why the multiple is so low. When the market gives a memory stock a 5x P/E, it's pricing in a downturn.

10x Entry Price Calculation

Working backwards from a bull case:

• 5-year target price: FY2031 EPS of $35 x 12x mid-cycle multiple = $420
• Alternative: FY2031 EPS of $40 x 15x multiple (reflecting structural HBM premium) = $600
• 10x entry price: $420/10 = $42. Or $600/10 = $60.

Where is the stock now vs that entry price? At $481, MU is 8-11x above the 10x entry price. There is no 10x upside from here under any realistic scenario.

What Would Make MU a Buy?

Working backwards from what constitutes a "good investment" (3-5x in 5 years, since 10x is unrealistic for a $543B company):

3x scenario (requires $1,440 in 5 years from current $481):

• FY2031 EPS of $80 x 18x = $1,440. This requires permanently elevated earnings AND a high multiple. Very aggressive.
• Probability: <10%

2x scenario (requires $960 in 5 years):

• FY2031 EPS of $48 x 20x = $960. Requires strong mid-cycle earnings and expansion multiple.
• Probability: ~15-20%

Base case ($481 -> $300-600 in 5 years):

• This is the most likely range. Micron earns enormous profits in FY2026-FY2027, then experiences some degree of normalization.
• At $300: Cycle turned harder than expected, stock at ~6x normalized earnings
• At $600: AI demand persists, HBM grows, cycle downturn is mild
• Probability: ~50-60%

Bear case ($481 -> $80-200):

• Memory cycle turns violently. DRAM/NAND pricing collapses. HBM demand slows. Stock falls to 1-2x book value.
• Probability: 20-30%

The Real 10x Opportunity

The 10x opportunity in MU is NOT buying at $481. It is:

1. Wait for the cycle downturn. History says this is coming. The question is when, not if.
2. Buy at the trough. The trough price is now likely $80-120 (book value will be $80-100 by the time a trough arrives in 2027-2028).
3. Hold through the next upcycle. If HBM demand continues (which AGI virtually guarantees), the next upcycle takes MU to $400-800 from a trough of $80-120 = 4-10x.

The v2 analysis estimated a $30-50 trough. I am revising this upward to $80-120 because:

• Tangible book value has grown from ~$40/share to ~$60/share and will be $80-100 by 2027-2028
• HBM provides a revenue floor that didn't exist in past troughs
• Multi-year SCAs provide contracted demand that reduces trough severity
• But this is still a 75-83% drawdown from current prices, consistent with historical memory cycle amplitude

Scenario Analysis Summary

Probability-weighted expected return from $481: Negative. The asymmetry is wrong at this price. You're risking 50-80% downside for 20-50% upside in most scenarios.

Key Uncertainties and Open Questions

1. When does the cycle turn? This is unknowable with precision. My best estimate: late CY2027 to CY2028. But it could be as early as H2 CY2026 if hyperscaler capex guidance disappoints, or as late as CY2029 if AI demand exceeds all forecasts.

1. How much does HBM change the cycle dynamics? HBM contracted revenue (via SCAs) should provide a floor. But how firm are these contracts? Can customers renegotiate pricing? What if NVIDIA GPU demand slows -- do the SCAs still hold?

1. Chinese DRAM competition. CXMT is 2-3 generations behind and blocked from EUV. But if export controls are relaxed (possible under Trump administration), Chinese entry could add meaningful supply and depress pricing.

1. The $100B New York megafab. Construction hasn't started. The project has been delayed. If Micron builds it, it adds enormous supply in ~2030. If they don't build it, the CHIPS Act funding may be at risk.

1. Does the DRAM oligopoly hold discipline? In past cycles, Samsung has deliberately overproduced to crush competitors' margins (especially Micron). If Samsung aggressively ramps HBM and conventional DRAM capacity simultaneously, it could trigger a downturn.

Bottom Line

MU at $481 is not a buy for a patient, value-oriented investor. The stock has run 5x in 16 months. It's trading at the peak of the most extreme memory cycle in history. Gross margins at 74-81% are unsustainable by management's own admission. The probability-weighted expected return is negative from here.

MU at $80-120 IS a buy. That's the 10x setup. Wait for the cycle downturn. It's coming -- it always comes. The difference this time is that HBM and AGI demand provide a structurally higher floor, which means the recovery from the trough will be more powerful and more certain than in past cycles.

Action items:

• Set price alerts at $200, $150, $120, $100, $80
• Monitor quarterly earnings for signs of cycle turn: rising inventory days, declining ASPs, softer guidance
• Track hyperscaler capex guidance -- any deceleration is the first domino
• Watch DRAM spot prices (DRAMeXchange) for early warning signals
• When the crash comes, buy aggressively. The AGI thesis means memory demand is structurally higher for the next decade. The oligopoly moat means Micron survives. The only question is the entry price.

SK Hynix (000660.KS) — Financial Picture

Last Updated: April 23, 2026

Stock Price: ₩1,216,000 (~$832 USD at ₩1,461/USD)

Market Cap: ₩859.6 trillion (~$588 billion USD)

Shares Outstanding: 701.69 million (728M total minus ~26M treasury/cancelled)

Enterprise Value: ₩824.6 trillion (~$564 billion USD)

Step 0: Investment Access for US Investors

Current Stock Price

• KRX Price (Apr 23, 2026): ₩1,216,000 per share
• USD Equivalent: ~$832 per share (at ₩1,461/USD mid-April 2026 rate)
• 52-Week Range: ₩176,700 – ₩1,267,000 (the all-time high of ₩1,173,000 was set April 15, 2026, but a new ATH of ₩1,267,000 was hit subsequently)
• 1-Year Return: +572%
• YTD 2026 Return: +66%

How a US Investor Can Buy

1. Direct Purchase on KRX (Korean Exchange)

The primary method today. SK Hynix trades as 000660 on the Korea Exchange. US brokerages that support Korean equities:

• Interactive Brokers — best option, direct KRX access, reasonable FX fees
• Charles Schwab International — available but higher fees
• Fidelity International — available with international account

You must convert USD → KRW through the broker, and settlement follows Korean market hours (KST, 13-14 hours ahead of EST).

2. ADR Listing (PENDING — Expected 2026)

SK Hynix filed a confidential Form F-1 with the SEC in March 2026, targeting a US ADR listing within 2026. Approximately 2.4% of outstanding shares (~₩10 trillion worth) may be offered as ADRs. This would be a game-changer for US investor access. No final exchange decision yet (likely NYSE). The SEC review is ongoing and the company said timing depends on "SEC review results, market conditions, and book-building demand." If approved, this eliminates the need for international brokerage accounts and Korean market access.

3. ETFs with Significant SK Hynix Exposure

• iShares MSCI South Korea ETF (EWY) — SK Hynix is the #2 holding at 22.2% weighting. Samsung is #1 at 22.8%. Together they are 45% of the fund. EWY has $19.6B in AUM. This gives significant but diluted exposure.
• Various Korean semiconductor ETFs — Korean-listed ETFs with 30-65% combined Samsung/SK Hynix weighting exist (e.g., Shinhan launched one at 65% combined weight in March 2026), but these trade on KRX and don't solve the US access problem.

4. Frankfurt-Listed GDR (HY9H.F / HY9H.DU)

SK Hynix has Global Depositary Receipts trading on Frankfurt Stock Exchange, but liquidity is very low. Not recommended for significant positions.

Tax Implications for US Investors

Dividend Withholding Tax:

Under the US-Korea Income Tax Treaty, Korean dividend withholding is capped at 15% for portfolio investors (under 10% ownership). This is claimable as a foreign tax credit on US tax returns, so effectively no double taxation if your US marginal rate exceeds 15%.

Capital Gains Tax:

Under the treaty, US residents are generally exempt from Korean capital gains tax on portfolio stock dispositions, provided they don't maintain 183+ days presence in Korea or have a permanent establishment there. Capital gains are taxed only in the US at normal US rates (long-term/short-term depending on holding period).

Practical Impact: The tax treatment is relatively clean for US investors. The 15% dividend withholding is recoverable via foreign tax credit, and capital gains are taxed only in the US.

Currency Risk: KRW/USD

2026 KRW/USD Trend:

• January 2026: ~₩1,433-1,479/USD (avg ₩1,455)
• February 2026: ~₩1,429-1,472/USD (avg ₩1,448)
• March 2026: ~₩1,454-1,517/USD (avg ₩1,490)
• April 2026: ~₩1,467-1,514/USD (avg ₩1,486)

The Korean won has weakened ~2.5% against the USD year-to-date in 2026. For a USD-denominated investor, this means:

• If you buy at ₩1,460/USD and the won strengthens to ₩1,350/USD, you gain ~8% on currency alone
• If it weakens to ₩1,550/USD, you lose ~6% on currency
• SK Hynix is a net exporter (sells chips globally in USD), so a weaker won actually boosts KRW-denominated earnings
• Bottom line: Currency risk is real but partially self-hedging because SK Hynix's revenues are predominantly USD-denominated while costs are KRW-denominated. A weaker won is actually good for the company's earnings, even if it reduces USD returns for the investor.

Part A — Structured Numbers

Income Statement (K-IFRS, Consolidated)

Annual Summary (Billion KRW)

Quarterly Trend — The Cycle in Full (Billion KRW)

The story in these numbers is astonishing. From an operating loss of ₩1.8 trillion in Q3 2023 to an operating profit of ₩37.6 trillion in Q1 2026 — a swing of ₩39.4 trillion in just 10 quarters. Operating margin went from negative 20% to positive 72%. This is the most dramatic earnings ramp in semiconductor history, driven almost entirely by HBM.

Revenue Segmentation (Estimated, FY2025)

SK Hynix does not publicly break out DRAM vs. NAND revenue in quarterly filings. Based on analyst estimates and partial disclosures:

FY2026 Estimates (Equity Research):

• Total Revenue: ~₩190 trillion (+95% YoY)
• Operating Profit: ~₩125 trillion
• Operating Margin: ~65%
• Net Profit: ~₩98 trillion

Balance Sheet (Billion KRW)

Balance sheet transformation: From ₩21.9 trillion net debt at end of 2023 to ₩35 trillion net cash at end of Q1 2026. This company went from financial distress to fortress balance sheet in under 3 years.

Cash Flow Statement (Billion KRW)

Capex Plans (2026-2028):

• FY2025 capex was ₩27.5 trillion (₩30.2T per CEO, discrepancy likely due to classification)
• FY2026 capex guidance: ₩35 trillion+ (significant increase)
• Major investments:
• Yongin Semiconductor Cluster: Total commitment expanded from ₩128 trillion to ₩600 trillion (multi-decade). First fab launches 2027 for HBM4/HBM4E and 321-layer NAND.
• M15X Facility (Cheongju): Active HBM4 production ramp, from ~10K wafers/month scaling to "several times" that by end-2026
• Advanced Packaging Facility: ₩19 trillion investment for HBM packaging
• EUV Equipment Procurement: Major spend in 2026

CEO Kwak has set a target of accumulating ₩100 trillion in net cash (matching Samsung's position) to fund this expansion cycle. At current run rates of ₩40+ trillion FCF/year, this is achievable within 2-3 years.

Valuation Metrics

Key valuation insight: SK Hynix trades at a forward P/E of 3.8x vs. Micron's 5.2x. This means the market is pricing SK Hynix at a ~27% discount to Micron on forward earnings, despite SK Hynix having 3x Micron's HBM market share and demonstrably higher margins. The "Korea discount" is real and significant.

Shareholder Returns

Dividends:

• FY2025 total: ₩3,000/share (₩1,500 fixed + ₩1,500 special/variable) = ₩2.1 trillion total payout
• Yield at current price: 0.25% — this is not a dividend story

Share Buyback:

• January 2026: Announced cancellation of ~50 million treasury shares (~2.1% of total), worth ~₩12.2 trillion
• This is more meaningful than the dividend — it's EPS-accretive

Capital Allocation Policy (2025-2027):

• Fixed dividend: ₩1,500/share/year (raised 25% in Nov 2024)
• Variable dividend: 50% of cumulative FCF after fixed dividends and investment
• Priority: Investment in capacity > share buyback/cancellation > variable dividend
• CEO's stated goal: Accumulate ₩100 trillion net cash first, then return more to shareholders

Part B — What the Numbers Really Mean

The HBM Revenue/Margin Revolution

Q1 2026's 72% operating margin is not a normal semiconductor margin. It is a reflection of HBM's extraordinary pricing power. HBM ASPs are 5-10x conventional DRAM, and HBM margins are substantially above corporate average. HBM likely accounts for well over 50% of Q1 2026's operating profit despite being perhaps 40-50% of revenue.

The revenue progression tells the story:

• Q3 2024: HBM was 30% of DRAM revenue
• Q4 2024: HBM was 40% of DRAM revenue
• Q3 2025: HBM revenue share was 57% (Counterpoint Research)
• Q1 2026: HBM likely 50-60% of total company revenue (extrapolated)

This means a company that was producing commodity memory chips 3 years ago is now primarily an HBM company. The entire margin structure has been transformed.

The Solidigm (Intel NAND) Acquisition

SK Hynix completed the $8.85 billion acquisition of Intel's NAND business (Solidigm) in March 2025, after a 5-year process that began in 2020. Key facts:

• Final payment of $1.9 billion to Intel closed the deal
• Solidigm brings enterprise SSD (eSSD) capabilities and the Dalian, China NAND fab
• Integration is progressing: NAND/eSSD revenue grew ~39% YoY in FY2025 to ~₩25 trillion
• The eSSD business is benefiting from AI server demand (AI training/inference requires massive storage)
• Solidigm is now profitable and contributing to growth, not a drag
• NAND remains the weaker segment (~26% of revenue) with structurally lower margins than DRAM

Inventory as Cycle Indicator

SK Hynix's current inventory levels are extremely healthy — the company has stated that its entire 2026 DRAM and NAND production is fully sold out (contracted). This is unprecedented. Normally, memory companies carry inventory risk; right now, SK Hynix is in a supply-constrained position where demand exceeds supply. This is the polar opposite of 2023 when bloated channel inventory crushed prices.

Historical Earnings: Peak, Mid-Cycle, Trough

The current earnings are not mid-cycle. They are near/at peak. The critical question is whether the HBM-driven supercycle represents a structural shift (new permanently higher baseline) or a cyclical peak that will correct. More on this in the future.md report.

What Is the Floor? (Downside Protection)

Tangible Book Value: ₩166,880/share (end FY2025), ~₩227,000/share (end Q1 2026). At the current price of ₩1,216,000, you're paying 5.4x tangible book. This is NOT a deep value situation. If earnings revert to mid-cycle, the stock could fall substantially before hitting book value support.

Fab Replacement Cost: SK Hynix's fabs cannot be replicated for less than ₩200-300 trillion. The Yongin cluster alone is budgeted at ₩600 trillion over its full buildout. But replacement cost is a theoretical floor — you don't get to access that value as a minority shareholder unless someone acquires the company, which is essentially impossible given the SK Group structure and Korean regulatory environment.

Realistic Downside in a Downturn:

• In the 2023 trough, SK Hynix's stock hit ₩75,500 (Jan 2023) — that was 0.6x book value at the time
• In a severe downturn, 1.0-1.5x book value would be the floor (~₩230,000-345,000 per share)
• This implies 70-80% downside from current levels in a trough scenario
• However, the HBM franchise did not exist at this scale during the last trough, so the floor may be structurally higher this time

Real Earnings Power: Normalized vs. Peak

If we strip out the HBM superprofits and assume a mid-cycle scenario:

• Conventional DRAM at mid-cycle margins (~20%): ~₩55T revenue × 20% = ₩11T operating income
• NAND at mid-cycle margins (~10%): ~₩30T revenue × 10% = ₩3T operating income
• HBM at reduced but still-premium margins (~35%): ~₩40T revenue × 35% = ₩14T operating income
• Normalized operating income: ~₩28 trillion (vs. current ₩77T TTM)
• Normalized EPS: ~₩30,000-35,000 (vs. current ₩106,606 TTM)
• At ₩1,216,000 per share, normalized P/E would be ~35-40x — expensive

This is the key risk. You are buying peak earnings at what looks like a low P/E (3.8x forward), but if earnings normalize, the stock is actually expensive. The bull case requires believing these earnings are sustainable and still growing.

Free Cash Flow Conversion

FCF conversion has been excellent since the recovery:

• FY2024: FCF ₩13.9T on ₩19.8T net income = 70% conversion
• FY2025: FCF ₩25.9T on ₩42.9T net income = 60% conversion
• TTM: FCF ₩41.8T on ₩75.2T net income = 56% conversion

The declining conversion rate is driven by rising capex (₩28-35T/year for HBM expansion). This is intentional — the company is investing heavily for growth. But it means SK Hynix is a capital-intensive growth story, not a cash cow. Don't expect large shareholder returns while the investment cycle is running.

Sources consulted: SK Hynix official earnings releases (Q1-Q4 2023, Q1-Q4 2024, Q1-Q4 2025, Q1 2026), StockAnalysis.com, Korea Herald, Joongang Daily, Seoul Economic Daily, equity research from Daishin Securities, LS Securities, Citi, KB Securities, Macquarie.

SK Hynix (000660.KS) — Business Understanding

Last Updated: April 23, 2026

Business Overview

SK Hynix is the world's #1 DRAM manufacturer (as of Q1 2025, overtaking Samsung for the first time since 1992) and the undisputed global leader in High Bandwidth Memory (HBM) with ~53-62% market share. The company designs, manufactures, and sells memory semiconductors — DRAM and NAND flash — from fabs in South Korea and China. It is a subsidiary of the SK Group chaebol, with SK Square holding a 20.5% controlling stake.

Revenue Mix (FY2025 estimated):

• DRAM: ~71% of revenue (~₩68-70 trillion)
• HBM: ~31% of total revenue (~₩30T+)
• Conventional Server/PC/Mobile DRAM: ~39% of total (~₩38T)
• NAND/eSSD (including Solidigm): ~26% of revenue (~₩25T)
• Other: ~3%

Weighting by future cash flow potential: HBM is ~31% of revenue but generates over 50% of operating profit due to its 5-10x ASP premium and higher margins. The correct way to analyze SK Hynix is as an HBM company with a DRAM/NAND commodity business attached. The HBM franchise is where virtually all the excess value creation occurs.

Segment 1: HBM (High Bandwidth Memory) — THE Franchise

What HBM Is and Why It Matters

HBM is a 3D-stacked DRAM product specifically designed for AI accelerators (GPUs and custom ASICs). It achieves bandwidth of 1,000+ GB/s by stacking 8-16 DRAM dies vertically using through-silicon vias (TSVs) and micro-bumps, then connecting them to the GPU via an interposer. This is orders of magnitude more bandwidth than conventional DRAM modules.

Every major AI GPU requires HBM:

• NVIDIA H100/H200: HBM3/HBM3E
• NVIDIA B100/B200: HBM3E
• NVIDIA Vera Rubin (next-gen): HBM4
• AMD MI300X/MI350: HBM3E
• Google TPU: HBM3E
• Custom ASICs from Broadcom, Marvell: HBM3E/HBM4

HBM is not optional — without it, AI accelerators cannot function. This makes HBM the most critical bottleneck component in the AI infrastructure stack.

How SK Hynix Became #1 in HBM

SK Hynix's HBM dominance is the result of a deliberate, decade-long bet that competitors dismissed:

1. Early Investment (2013-2018): SK Hynix invested heavily in 3D stacking and TSV technology when Samsung and Micron viewed it as niche. They were the first to mass-produce HBM2E (2018-2019).

1. NVIDIA Partnership: SK Hynix became NVIDIA's primary HBM supplier starting with the A100 GPU. This was not accidental — SK Hynix's engineering team worked closely with NVIDIA to co-optimize HBM specifications for NVIDIA's GPU architectures. This co-development created deep technical lock-in.

1. Consistent Generation Leadership:

• HBM2E: SK Hynix first (2018-2019)
• HBM3: SK Hynix first (Q4 2022), Micron followed 6-9 months later
• HBM3E: SK Hynix first (Q3 2023 development, mass production March 2024), Micron 9 months behind
• HBM4: SK Hynix first (development completed April 2025, mass production February 2026), Samsung/Micron validation stage

1. Yield Advantage: HBM manufacturing is extremely difficult. The bonding, stacking, and thermal management of 8-12 die stacks requires extraordinary precision. SK Hynix achieved mature yields 6-12 months before competitors at every generation, giving them a massive cost advantage during the critical ramp period.

1. Samsung's Stumble: Samsung had quality issues with HBM3E, failing NVIDIA's qualification process repeatedly in 2024. This gave SK Hynix what one analyst called "a multi-quarter window of near-monopoly supply." Samsung's problems were not trivial — they reflected deeper organizational issues in Samsung's memory division.

HBM Market Share

SK Hynix's share has been declining slowly from its peak of ~70%+ as Samsung and Micron ramp up, but it remains the dominant supplier with more than 50% of the market.

HBM Technology Roadmap

Key technology details:

• Thermal Management: SK Hynix uses wafer-level passive heat spreaders (Si/SiC), achieving less than 2°C temperature delta on 12-hi stacks. This is production-proven and critical for HBM3E 12-hi yields.
• Bonding: Transitioning from solder micro-bumps to copper-copper direct bonding (hybrid bonding) for HBM4, which enables tighter pitch and better thermal/electrical performance.
• Logic Base Die Partnership: SK Hynix outsources HBM4 logic base dies to TSMC through a "One-Team" alliance, using TSMC's 12nm process. This differs from Samsung, which fabricates its own logic dies in-house.

The NVIDIA Relationship

SK Hynix is NVIDIA's preferred and primary HBM supplier. The relationship goes beyond simple vendor-customer:

• SOCAMM2 Exclusive: SK Hynix began mass production of 192GB SOCAMM2 (Small Outline Compression Attached Memory Module 2) modules in April 2026 specifically designed for NVIDIA's Vera Rubin platform. SOCAMM2 delivers 2x bandwidth and 75% better power efficiency vs. conventional RDIMM. While "exclusive" is not explicitly stated in press releases, SK Hynix is the only company producing SOCAMM2.

• Sold Out Through 2026: SK Hynix's CFO stated the company has "already sold out our entire 2026 HBM supply." Given that NVIDIA is the dominant buyer of HBM (80%+ of HBM demand comes from NVIDIA GPUs), this effectively means NVIDIA has contracted for SK Hynix's entire HBM output.

• Co-Development: SK Hynix engineers work directly with NVIDIA's GPU architecture team to optimize HBM specifications for each new GPU generation. This creates switching costs — NVIDIA would need to re-validate and potentially re-design parts of its GPU to accommodate a different HBM supplier's product.

Can Samsung/Micron Close the Gap?

Samsung:

• Samsung completed HBM3E performance analysis with "encouraging feedback" in late 2025
• Planning a 50% HBM capacity surge in 2026 (170K → 250K wafers/month)
• Delivering paid HBM4 samples to NVIDIA, but still in validation
• Forecast to capture ~25-28% of HBM4 market (up from ~17% in Q1 2025)
• Samsung's manufacturing scale is a long-term threat, but their quality and yield issues have been persistent
• Assessment: Samsung will close some of the gap but is unlikely to overtake SK Hynix in HBM within the next 2-3 years. The yield gap is a structural advantage that takes years to close.

Micron:

• Micron has a larger HBM patent portfolio (621 vs. 315) but patents don't translate to manufacturing execution
• HBM3E production started ~9 months after SK Hynix
• HBM4 expected Q2 2026, about 4-6 months behind SK Hynix
• Micron's HBM3E has 30% lower power consumption, which is a genuine technical advantage
• Micron has invested $1B+ specifically in HBM capabilities
• Market share: ~21%, positioned as qualified secondary supplier
• Assessment: Micron is technically competent but structurally behind. Their US-based manufacturing may actually be advantageous for geopolitical reasons but they lack the production scale and yield maturity of SK Hynix.

Key insight from the equity research: "Patent strength does not automatically translate to manufacturing execution. SK Hynix's manufacturing lead and design win relationships remain structural advantages that patents alone cannot overcome."

Segment 2: Conventional DRAM (DDR5, LPDDR5)

Market Position

As of Q1 2025, SK Hynix became the #1 global DRAM manufacturer by revenue for the first time, overtaking Samsung which had held the top position since 1992:

• SK Hynix: 36% global DRAM market share
• Samsung: 34%
• Micron: 25%

This historic shift was driven by HBM (which is classified as DRAM revenue), but SK Hynix is also competitive in conventional DRAM:

• DDR5 server DRAM: Strong position, high-capacity 128GB+ DDR5 modules saw shipments more than double in Q3 2025
• LPDDR5/LPDDR5X: Mobile and on-device AI DRAM, growing segment
• Technology Node: 1cnm (sixth-generation 10nm-class) in stable mass production — this is among the most advanced nodes in the industry

Competitive Position vs. Samsung and Micron

Cost Structure: Detailed per-wafer cost comparisons are not publicly disclosed, but operational metrics reveal the picture:

• SK Hynix FY2025 operating margin: 49%
• Samsung Memory Division FY2025 estimated margin: 30-35%
• Micron FY2025 operating margin: ~30%

SK Hynix's significant margin advantage is primarily HBM-driven, but even in conventional DRAM they are cost-competitive thanks to the 1cnm process and high utilization rates.

The Structural Shift: HBM Cannibalization

A critical dynamic is playing out: HBM production cannibalizes conventional DRAM supply. HBM requires 3-5x more wafer area per bit than conventional DRAM (because of the TSV keep-out zones, lower yields, and 3D stacking overhead). As all three producers divert more wafer capacity to HBM, the supply of conventional DRAM tightens.

This is creating a memory shortage expected to last until at least 2027 (per industry analysts). Conventional DRAM prices are rising as a result, which benefits all three producers but SK Hynix most because they have the largest HBM allocation.

Segment 3: NAND/eSSD (Including Solidigm)

The Intel NAND Acquisition (Solidigm)

SK Hynix completed the $8.85 billion acquisition of Intel's NAND business in March 2025:

• Originally announced in October 2020 as a two-step transaction
• Step 1 (completed December 2021): Acquired Intel's SSD business, including Solidigm brand
• Step 2 (completed March 2025): Acquired remaining NAND flash wafer and component business, including the Dalian, China fab
• Total consideration: $8.85 billion over 5 years

Integration Status (as of Q1 2026):

• Solidigm is contributing positively — NAND/eSSD revenue grew ~39% YoY in FY2025 to ~₩25 trillion
• The enterprise SSD (eSSD) business is a bright spot, driven by AI server demand for high-capacity storage
• SK Hynix is expanding 321-layer NAND production and QLC-based high-capacity eSSD products
• The acquisition gives SK Hynix a stronger competitive position in enterprise storage, where Solidigm had existing customer relationships with hyperscalers

Key Risk: The Dalian NAND fab in China faces the same US export control risks as the Wuxi DRAM fab. If US restrictions tighten further, both fabs face equipment upgrade limitations.

NAND Market Position

SK Hynix + Solidigm combined is the #3 NAND producer globally behind Samsung and Kioxia/WD. NAND is the weaker business compared to DRAM:

• Lower margins (~10-15% mid-cycle vs. ~25-30% for DRAM)
• More commoditized product
• Cyclical pricing with less differentiation opportunity
• Samsung's vertical integration (controller + NAND + packaging) gives it structural advantages

However, the eSSD business changes this calculus. Enterprise SSDs for AI servers carry meaningfully higher margins than commodity NAND, and this is the fastest-growing sub-segment.

Key Business Questions & Answers

SK Group Structure and Governance

Ownership Chain:

SK Group → SK Inc. → SK Square (formerly part of SK Telecom) → SK Hynix (20.5% stake)

SK Square was spun off from SK Telecom in November 2021 to house SK Telecom's semiconductor and non-telecom investments. SK Square is the controlling shareholder of SK Hynix.

Other Major Shareholders:

• National Pension Service (NPS): 7.9%
• Capital Research & Management: 5.1%
• BlackRock: 5.0%
• Foreign institutional investors (total): ~54% — among the highest foreign ownership ratios for any KOSPI large-cap

Related Party Transaction Risks:

Like all Korean chaebols, the SK Group structure creates potential for value extraction from SK Hynix by the parent:

• SK Square could pressure SK Hynix to make investments or transactions that benefit the broader group
• Cross-subsidization risk exists (though SK Hynix is the crown jewel, making this less likely — the parent benefits from SK Hynix's success)
• SK Group has a reputation for being more transparent and minority-friendly than some other chaebols (e.g., Samsung Group has a much worse track record on governance)
• The 54% foreign ownership provides meaningful governance oversight

Board Structure: 9 members (3 inside directors, 6 independent). One-share-one-vote, no dual-class structure. NPS and large foreign holders "effectively hold the balance of power" on contentious votes.

Korean Corporate Governance Risks

The "Korea Discount" is real. Korean equities systematically trade at lower multiples than equivalent US or Taiwanese companies. SK Hynix at 3.8x forward P/E vs. Micron at 5.2x illustrates this.

Reasons for the discount:

1. Chaebol risk: Controlling shareholders can extract value in ways that don't benefit minority shareholders
2. Treasury share overhang: Korean companies historically held massive treasury stakes (SK Hynix had 20M+ shares). New regulations may force cancellation, which is actually positive
3. Dividend culture: Korean companies historically paid very low dividends. SK Hynix's 0.25% yield reflects this
4. Regulatory uncertainty: Korean labor laws (52-hour workweek), tax policies, and regulatory environment can change
5. Geopolitical risk: Korea's proximity to North Korea and exposure to US-China tensions
6. Currency risk: KRW volatility adds uncertainty for foreign investors

Recent Positive Developments:

• Proposed Commercial Act amendments would require cancellation of treasury shares within 1 year (positive for minority shareholders)
• SK Hynix's ADR listing (if approved) would bring it under US disclosure requirements and attract US institutional capital
• The company's share buyback/cancellation program signals improving capital allocation

Geopolitical Risk: US-China Tech War

This is the most significant risk to SK Hynix outside of cyclical downturn.

The China Fab Exposure:

• Wuxi DRAM Fab: Produces DRAM chips, representing approximately 40% of SK Hynix's DRAM production capacity. However, it operates at prior-generation technology (not cutting-edge). This fab is critical to SK Hynix's production volume.
• Dalian NAND Fab (Solidigm): Produces NAND flash, part of the Intel acquisition. Older-generation technology.

US Export Controls:

• Samsung and SK Hynix were removed from the "validated end-user" (VEU) list in the latest US export control tightening
• They now face case-by-case licensing reviews for each equipment shipment to Chinese fabs
• Estimated 1,000+ additional export license applications annually
• Equipment cannot be easily upgraded without US approval, meaning China fabs will gradually fall behind in technology

Impact Assessment:

• Equity research estimates 15-20% revenue loss if China operations are completely severed
• Worst case: -30% operating profit impact through fixed-cost deleverage
• The China fabs produce older-generation commodity products, so the margin impact is lower than the revenue impact
• SK Hynix is gradually shifting cutting-edge production to Korean fabs and reducing China dependency

Mitigating Factors:

• The US has not (yet) fully blocked operations — they've made it harder but not impossible
• Complete severance would hurt US companies too (lost equipment sales, supply disruption)
• SK Hynix is investing heavily in Korean capacity (Yongin, M15X) to reduce China dependence
• The China fabs don't produce HBM or cutting-edge products

Korea's K-CHIPS Act (Government Support)

South Korea passed the K-CHIPS Act in February 2025, providing:

• Tax incentives for semiconductor R&D and capital investment
• Government support for semiconductor infrastructure
• However, the act came with caveats — "lingering regulatory uncertainties could still undermine long-term competitiveness"
• The separate Semiconductor Special Act (which would have exempted chipmakers from the 52-hour workweek rule) remains stalled in the National Assembly due to political gridlock

Comparison to US CHIPS Act:

• US CHIPS Act provides ~$52 billion in direct subsidies and tax credits
• Korea's K-CHIPS Act focuses more on tax incentives than direct subsidies
• Korea's approach is less generous in direct dollar terms but the tax benefits are meaningful for capex-heavy companies like SK Hynix
• Key difference: US CHIPS Act explicitly restricts recipients from expanding in China; K-CHIPS Act does not (though US export controls achieve a similar effect)

Management Quality

CEO Kwak Noh-jung has led SK Hynix through the most significant transformation in its history:

• Navigated the brutal 2023 downturn (₩7.7 trillion operating loss) without panic
• Made the early bet on HBM that proved transformative
• Completed the Solidigm integration
• Set aggressive but credible financial targets (₩100 trillion net cash)
• Managed the NVIDIA relationship into a strategic partnership
• Capex discipline during downturn, aggressive expansion during upturn

Capital Allocation Track Record:

• FY2024 capex: ₩16.0 trillion (restrained during recovery phase)
• FY2025 capex: ₩27.5-30.2 trillion (ramping into cycle)
• FY2026 capex: ₩35 trillion+ (full expansion mode)
• The timing of capex has been well-calibrated to the cycle — they didn't overinvest at the top (2022) and didn't underinvest during the recovery
• Share buyback/cancellation program adds value without reckless spending

Assessment: Management is excellent. They have been consistently 6-12 months ahead of competitors in technology, well-timed in capital allocation, and increasingly shareholder-friendly. This is one of the best-run semiconductor companies in the world.

Cost Structure Comparison

SK Hynix's margin advantage is structural (HBM yield + mix) plus cyclical (peak pricing). If HBM margins normalize, the gap narrows but doesn't disappear because SK Hynix maintains a process technology lead.

Moat Assessment

What is the moat? SK Hynix's competitive advantage is a combination of:

1. Technology Leadership (HBM): 6-12 month head start at each generation. This is the primary moat — competitors cannot easily replicate years of yield learning and process optimization.

1. Customer Lock-in (NVIDIA): Deep co-development relationship with the dominant GPU maker. Switching costs are real — NVIDIA would need to re-validate its entire GPU platform with a different HBM supplier.

1. Oligopoly Structure: Only 3 companies in the world can make advanced DRAM (Samsung, SK Hynix, Micron). The barriers to entry are enormous — a new DRAM fab costs $15-20 billion and takes 3-5 years to build. No new entrant has successfully entered the DRAM market in decades. China's CXMT is trying but is years behind and faces US technology restrictions.

1. Scale Economics: Memory manufacturing has massive fixed costs and steep learning curves. SK Hynix's scale (₩97 trillion revenue, growing to ₩190 trillion) spreads these costs and funds the R&D needed to stay competitive.

1. Physical Assets: You cannot replicate these fabs in a data center. This is one of the most AGI-durable moats in technology — regardless of how intelligent AI becomes, it still needs physical memory chips manufactured in billion-dollar clean rooms.

Moat Durability Rating: STRONG — The oligopoly is one of the most durable structures in the semiconductor industry. Three producers, massive barriers to entry, no new entrants possible within 5-10 years. The HBM technology lead is more temporary (Samsung is closing the gap) but the oligopoly structure is essentially permanent.

Sources: SK Hynix official earnings releases and investor presentations, KEDGlobal, Counterpoint Research, Patsnap HBM technology analysis, Korea Herald, equity research from Daishin/Citi/KB/Macquarie, Introl blog, TrendForce.

SK Hynix (000660.KS) — Future Prediction

Last Updated: April 23, 2026

2-Year Outlook (2026-2028): Supply, Demand, and Pricing

HBM Revenue Trajectory

Current State (Q1 2026): SK Hynix reported ₩52.6 trillion revenue with 72% operating margins. HBM is the overwhelming driver. The company's entire 2026 HBM and DRAM production is sold out under contract.

2026 Full-Year Estimates:

• Revenue: ~₩190 trillion (+95% YoY)
• Operating Profit: ~₩125 trillion
• Operating Margin: ~65%
• Net Income: ~₩98 trillion
• EPS: ~₩140,000

2027 Estimates (extrapolated from current trajectory and analyst range):

• Revenue: ~₩230-260 trillion (+20-37% YoY, growth decelerating)
• Operating Profit: ~₩140-170 trillion
• Operating Margin: ~60-65% (slight compression as Samsung competes more effectively in HBM4)
• EPS: ~₩160,000-200,000

2028 Estimates:

• Revenue: ~₩250-300 trillion (+8-15% YoY, further deceleration)
• Operating Profit: ~₩140-180 trillion
• Operating Margin: ~55-62% (HBM pricing pressure from three-player competition)
• EPS: ~₩160,000-210,000

Can SK Hynix Sustain 50%+ HBM Market Share?

Short answer: Probably yes through 2027, then it gets harder.

Daishin Securities projects SK Hynix's HBM share will erode from ~62% (HBM3E peak) to ~55% as Samsung captures ~28% of HBM4 volume. This is a gradual decline, not a collapse.

Factors supporting sustained share:

1. First-mover advantage in HBM4 (mass production Feb 2026, 4+ months ahead of Samsung/Micron)
2. NVIDIA strategic relationship and SOCAMM2 exclusivity
3. Yield maturity advantage compounds over time — it's not just about starting first, it's about accumulating yield learning that competitors can't shortcut
4. TSMC partnership for logic base dies (outsourcing to the world's best foundry)
5. 2026 supply is already fully contracted

Factors threatening share:

1. Samsung's 50% HBM capacity increase in 2026 (170K → 250K wafers/month)
2. Samsung completed HBM4 validation with NVIDIA ("encouraging feedback")
3. Micron expected to begin HBM4 production Q2 2026
4. Non-NVIDIA customers (AMD, Google, Broadcom custom ASICs) may diversify suppliers more aggressively
5. Goldman Sachs warns of 10% HBM price decline in 2026 as supply catches up

Base case: SK Hynix maintains 50-55% HBM share through 2027, gradually declining to 45-50% by 2028 as Samsung and Micron achieve yield parity on HBM4E.

Conventional DRAM Pricing Outlook

This is actually a positive surprise story that few investors appreciate. The diversion of wafer capacity from conventional DRAM to HBM is creating a structural supply shortage in conventional DRAM:

• All three producers are prioritizing HBM production because margins are 5-10x higher
• Every wafer used for HBM is a wafer NOT used for DDR5/LPDDR5
• HBM requires 3-5x more wafer area per bit than conventional DRAM
• Industry analysts forecast the memory shortage to last until at least 2027

This means:

1. Conventional DRAM prices are rising (good for all three producers)
2. SK Hynix benefits most because they have the highest HBM allocation, meaning their conventional DRAM production is the most constrained → prices go even higher
3. Server DDR5 is seeing strong demand from AI data center expansion (servers need conventional DRAM too, not just HBM)
4. On-device AI (smartphones, PCs) is driving LPDDR5X demand higher

Bottom line: Even if HBM pricing softens 10%, conventional DRAM pricing strength could partially offset the impact. The overall DRAM pricing environment is the strongest in years and likely persists through 2027.

NAND/Solidigm Outlook

NAND is the weaker segment but improving:

• Enterprise SSD (eSSD) demand is robust — AI training/inference workloads require massive storage
• Solidigm integration is complete and the business is contributing revenue growth
• 321-layer NAND and QLC technology provide cost competitiveness
• NAND pricing has been recovering from 2023 lows

Risk: NAND is more commoditized than DRAM and lacks the HBM-like differentiation opportunity. Samsung's vertical integration (controller + NAND + packaging) gives it structural advantages. Solidigm's Dalian fab faces the same US-China export control risk.

Assessment: NAND is a contributor, not a drag, but it's not the thesis. If you're buying SK Hynix, you're buying HBM first, DRAM second, NAND third.

HBM Oversupply Risk

This is the single biggest risk to the investment thesis.

If all three producers ramp HBM simultaneously:

• Samsung: 50% capacity increase in 2026
• SK Hynix: M15X ramp + Yongin cluster first fab in 2027
• Micron: $1B+ HBM-specific investment, ramping through 2026-2027

Demand side: NVIDIA's GPU revenue was ~$57 billion in a recent quarter. Hyperscalers collectively planned $131 billion+ in capex. HBM demand is growing 50-100%+ annually as each new GPU generation requires more HBM per chip (H100: 80GB, B200: 192GB, Vera Rubin: 288GB).

The math: Right now, demand exceeds supply — all HBM is sold out through 2026. The question is when supply catches up. TrendForce and Goldman Sachs have warned of potential 10% HBM price declines in late 2026 as Samsung's expanded capacity reaches market. But 10% is manageable given HBM margins of 50%+.

Genuine oversupply risk timeline: Late 2027 to 2028, when all three producers' capacity expansions are fully online and if AI infrastructure spending growth decelerates. If hyperscaler capex plateaus or declines (as OpenAI's capex revision from $1.4T to $600B through 2030 suggests is possible), supply could exceed demand.

Probability assessment:

• Temporary HBM price softening in H2 2026: 40% probability (manageable, 10-15% price decline)
• Genuine HBM oversupply in 2027-2028: 25% probability (would compress margins from 65% to 40-50%, still very profitable)
• Severe HBM glut (margins below 30%): 10% probability — would require both massive supply expansion AND a pullback in AI spending

Where Will EPS Be?

Even in the bear case (HBM oversupply, Samsung competition, China fab disruption), SK Hynix would still be a massively profitable company at ₩100,000+ EPS. The question is whether the stock price holds at these levels in a bear scenario — it probably doesn't. A reversion to 10-12x mid-cycle P/E on bear-case earnings implies ₩1.0-1.2 million, roughly flat from here.

5-10 Year Outlook (Post-AGI)

Does the HBM Leadership Position Persist?

The technology-specific leadership (HBM3E, HBM4) is temporary. The oligopoly is permanent.

SK Hynix's generation-by-generation HBM leadership will narrow over time. Samsung has vastly more resources (3x SK Hynix's capex capacity) and will eventually close the yield gap. By HBM5 or HBM6 (2028-2030), all three producers will likely have competitive products. At that point, competition shifts to:

1. Cost efficiency (Samsung's scale advantage)
2. Customer relationships (SK Hynix's NVIDIA advantage)
3. Technology innovation speed (currently SK Hynix, could shift)

However, the critical point is that all three producers will be enormously profitable. Even if HBM market share equalizes at 33/33/33, the total HBM market will be 5-10x larger than today. A third of a much larger pie is still a very large piece.

How Does AGI Affect Memory Demand?

AGI is the ultimate bull case for memory semiconductors.

Apply the AGI worldview from the analysis guide:

1. Recursive self-improvement requires compute, which requires memory. As AI models grow to trillions of parameters and inference demand scales, HBM and DRAM demand grow proportionally. Every GPU/ASIC in a data center needs more memory than the previous generation.

1. 30+ GW of data center capacity by 2030 (OpenAI 30GW, Anthropic 15-30GW, plus Google, Meta, Microsoft, xAI). Each GW of AI data center capacity requires tens of thousands of GPUs, each requiring HBM. The raw demand numbers are staggering.

1. Agentic AI = continuous real-time inference. SK Hynix explicitly noted this in their Q1 2026 earnings — as AI moves from batch training to continuous inference (agentic AI doing real-time tasks), memory demand expands across both DRAM and NAND. Agents need persistent memory, context windows, and fast retrieval.

1. The memory bottleneck is real and persistent. GPUs are compute-limited, but increasingly they are ALSO memory-limited. The shift from 80GB (H100) to 288GB (Vera Rubin) per GPU illustrates that memory scaling is as critical as compute scaling.

1. On-device AI creates new demand vectors. As AGI capabilities trickle down to edge devices, every smartphone, laptop, and robot needs more memory. LPDDR5X demand for on-device AI is a new growth driver that barely existed 2 years ago.

Memory demand in an AGI world: If we take the analysis guide's framework literally — billions of geniuses operating in data centers, recursive self-improvement, 30GW+ of compute by 2030 — the memory demand curve is essentially exponential. Every dollar put into AGI compute generates demand for memory. And memory fabs take 3-5 years to build, creating persistent supply constraints.

Is the Memory Oligopoly AGI-Durable?

YES — this is one of the most AGI-durable moats in technology.

The analysis guide asks: "Is the moat durable when billions of geniuses operate in data centers?" For memory semiconductors, the answer is unequivocally yes:

1. Physical manufacturing cannot be replicated in software. You cannot code a DRAM fab into existence. These are among the most complex manufacturing facilities ever built, requiring:

• Clean rooms with Class 1 air quality (fewer than 1 particle per cubic foot)
• Extreme ultraviolet (EUV) lithography machines (only ASML makes them, $150M+ each)
• 1,000+ process steps per wafer
• Years of yield learning that cannot be shortcutted even with AI
• $15-20 billion per fab, 3-5 years to build

1. The AGI companies need memory — they can't bypass it. OpenAI, Anthropic, Google, etc. will design custom ASICs and optimize software, but they still need physical memory chips. AI can help optimize fab processes (and it will), but it cannot eliminate the need for physical manufacturing.

1. New entrants face 15-20 year catch-up. China's CXMT has been trying to build a domestic DRAM industry for years and remains 3-5 generations behind, with US restrictions making it harder. Even with unlimited funding and AI assistance, replicating the yield learning and process knowledge of SK Hynix/Samsung/Micron would take a decade.

1. Robotics doesn't change this calculus. Even when robotics is solved, building and operating semiconductor fabs requires extreme precision that pushes the limits of physics, not just labor. Robots could help with fab construction but the process engineering knowledge remains a moat.

However, AGI could accelerate process development. AI-driven materials science and process optimization could eventually compress the innovation cycle, potentially allowing Samsung or new entrants to close technology gaps faster. This is a very long-term risk (2030+) and would affect all three producers equally.

Samsung's Manufacturing Scale: Long-Term Threat?

Samsung has a structural advantage SK Hynix cannot match: Samsung is the world's largest semiconductor manufacturer with 3x SK Hynix's total capacity. Samsung can outspend SK Hynix on capex (₩40T+ vs. ₩35T+) and can afford to lose money longer in a downturn.

In a truly mature HBM market where technology leadership narrows, Samsung's scale advantage could allow it to:

• Undercut SK Hynix on pricing
• Capture share through sheer production volume
• Cross-subsidize memory from its other semiconductor businesses

Counter-argument: SK Hynix has a ₩600 trillion Yongin cluster plan specifically designed to close the capacity gap. And Samsung's conglomerate structure is a double-edged sword — the memory division competes with Samsung's foundry, display, and other businesses for capital and management attention. SK Hynix, as a pure-play memory company, has a singular focus that Samsung lacks.

Assessment: Samsung will eventually become a competitive HBM producer. But "competitive" is not the same as "dominant." The most likely long-term equilibrium is three roughly comparable producers competing on technology, cost, and customer relationships — similar to the conventional DRAM market that has existed for decades. The oligopoly persists; the question is only the share split.

Comparison to Micron (MU)

The Core Question: Which Is the Better Investment?

Detailed Comparison

HBM Position:

SK Hynix has a massive and indisputable lead. 3x Micron's market share. 6-9 months ahead at every generation. The NVIDIA primary supplier relationship. There is no comparison on HBM — SK Hynix wins overwhelmingly.

Micron's HBM3E does have 30% lower power consumption, which is a genuine technical advantage. Micron also has a larger patent portfolio. But these advantages have not translated to market share because manufacturing execution — yield, volume, reliability — is what matters, and SK Hynix dominates on all three.

Valuation:

SK Hynix is cheaper on every metric. Forward P/E of 3.8x vs. 5.2x. P/B of 5.2x vs. 7.5x. This reflects the "Korea discount" — the structural discount Korean equities carry vs. US-listed equivalents.

The valuation gap is unjustified on fundamentals. SK Hynix has higher margins, higher growth, and a stronger competitive position. The discount is purely structural (listing jurisdiction, governance risk, currency risk).

Risk Comparison:

SK Hynix Risks:

• China fab exposure (~40% DRAM + Dalian NAND) — existential geopolitical risk
• Chaebol governance (SK Group structure)
• KRW currency risk
• Korean regulatory/political risk
• Less liquid for US investors (until ADR listing)

Micron Risks:

• HBM market share loss if SK Hynix/Samsung dominate
• Higher valuation relative to earnings
• US CHIPS Act subsidy comes with restrictions (no China expansion for 10 years)
• Micron's fabs are primarily in the US, Japan, and Singapore — higher cost structures
• Less differentiated in the AI memory narrative

CHIPS Act vs. K-CHIPS Act:

Micron received $6.1 billion+ in direct US CHIPS Act subsidies. SK Hynix gets Korean tax credits but no comparable direct cash. This is a real advantage for Micron's US operations. However, SK Hynix's lower tax base and KRW-denominated cost structure partially offset this.

Currency Risk:

For a USD-denominated investor, SK Hynix carries KRW/USD currency risk. The won has weakened ~2.5% YTD in 2026. In a risk-off environment, the won can weaken 10-15%, which would erode USD returns even if the stock appreciates in KRW. This is a genuine additional risk that Micron does not have.

Verdict: Which to Buy?

For maximum risk-adjusted return: SK Hynix is the better value, but only if you can manage the execution complexity (international brokerage, currency, Korea discount acceptance). You are getting a demonstrably better business at a demonstrably lower valuation.

For accessibility and simplicity: Micron is the pragmatic choice. US-listed, no currency risk, no governance concerns, solid HBM exposure (even if #3). You pay a premium for the simplicity.

Optimal strategy for a sophisticated US investor: Buy SK Hynix directly on KRX through Interactive Brokers, hold for the ADR listing catalyst (expected 2026). The ADR listing alone could close some of the Korea discount, driving a 15-25% rerating. Then decide whether to hold the ADR or the KRX shares based on liquidity needs.

If the ADR listing happens: It would likely trade at a premium to the KRX shares (as TSMC's ADR trades at a premium to the Taiwan listing), providing an immediate catalyst.

10x Framework

Bull Case 5-Year Price Target (2031)

Assumptions for the bull case:

• HBM market grows 5-7x from current size by 2031 (driven by AGI compute buildout)
• SK Hynix maintains 45-50% HBM market share
• HBM revenues reach ₩200-250 trillion annually
• Total revenue reaches ₩400-500 trillion
• Operating margin stabilizes at 40-50% (competition compressed from current 72% but still very high)
• Operating income: ₩160-250 trillion
• Net income: ₩130-200 trillion
• EPS: ₩185,000-285,000
• At 12-15x P/E (re-rated closer to global semiconductor peers thanks to ADR listing):
• Bull case price: ₩2,200,000 - ₩4,300,000 per share
• In USD: $1,500 - $3,000 per share (assuming ₩1,450/USD)

Working Backwards: What Entry Price Gives 10x in 5 Years?

For 10x in 5 years, you need the stock at one-tenth of the 5-year target:

• Bull case target: ₩2,200,000 - ₩4,300,000
• 10x entry price: ₩220,000 - ₩430,000 per share
• In USD: $150 - $295 per share

Current Price vs. 10x Entry Price

Current price: ₩1,216,000 per share

10x entry price: ₩220,000 - ₩430,000 per share

The stock is currently 2.8x - 5.5x ABOVE the 10x entry price.

This means from the current price:

• Bull case upside: 1.8x - 3.5x over 5 years (80% - 250% return)
• A 10x return requires buying at ₩220,000-430,000, which would only happen in a severe downturn (comparable to the ₩75,000 - ₩176,000 range seen in 2023)

Reality check: SK Hynix is not a 10x opportunity at current prices. The massive appreciation in 2024-2026 (+572% in 12 months) has priced in much of the HBM thesis. You are buying a great business at a fair price — not a cheap one.

What Would Make It a 10x from Here?

For the stock to 10x from ₩1,216,000 to ₩12,160,000:

• Revenue would need to reach ₩800T-1,000T (5-6x current)
• This implies HBM + memory becoming a $500-600 billion annual market
• And SK Hynix maintaining 40%+ share
• And P/E expanding to 15-20x

This is conceivable only in the most extreme AGI bull case where:

• Data center compute scales to 50-100GW globally by 2031
• Each GW requires 100,000+ GPUs with 288GB+ HBM each
• Memory demand becomes the primary bottleneck in AGI infrastructure
• The three producers can't expand capacity fast enough

Probability of 10x from current price in 5 years: 5-10%. Not impossible given AGI dynamics, but you'd need everything to go right simultaneously.

Risk/Reward at Current Price

Expected value: Weighted average ~₩2,000,000 - ₩2,400,000 (65-97% upside from current)

The risk/reward is positive but not overwhelmingly so. You are paying for a lot of good news, and the downside scenarios are real (20-25% probability of significant loss).

Key Catalysts and Risks (Next 12-24 Months)

Catalysts (Positive)

1. ADR listing (2026): Would unlock US institutional demand and potentially close the Korea discount. Single biggest near-term catalyst.
2. HBM4 ramp execution: Continued first-mover advantage with NVIDIA Vera Rubin validates the technology leadership moat.
3. Q2/Q3 2026 earnings: If margins sustain at 65%+, it confirms the supercycle is intact.
4. Conventional DRAM price increases: Memory shortage through 2027 provides pricing tailwind beyond HBM.
5. Share buyback/cancellation: ₩12.2T program is EPS-accretive.
6. Yongin cluster first fab launch (2027): Demonstrates long-term capacity scaling.

Risks (Negative)

1. Samsung HBM4 qualification by NVIDIA: If Samsung passes NVIDIA's quality bar and captures 25%+ share, SK Hynix's margins compress.
2. HBM price decline: Goldman Sachs warns of 10% price decline in H2 2026 as supply ramps.
3. US-China export controls tightening: Loss of China fab operations = 15-20% revenue hit.
4. Hyperscaler capex pullback: If AI spending disappoints (OpenAI revising capex from $1.4T to $600B is a warning signal).
5. Currency: KRW weakening further erodes USD returns.
6. Cyclical risk: Memory is cyclical. Current margins (72%) are extreme and will normalize.

Watch Triggers

• Quarterly earnings (every 3 months — next: Q2 2026 in late July)
• ADR listing announcement and pricing
• Samsung HBM4 NVIDIA qualification result
• US export control updates affecting China fabs
• NVIDIA GPU demand signals (NVIDIA earnings as proxy for HBM demand)
• Conventional DRAM spot pricing trends
• KRW/USD exchange rate
• Hyperscaler capex guidance (Meta, Microsoft, Google, Amazon quarterly results)

Investment Conclusion

SK Hynix is an exceptional business at the center of the most important technology trend of the decade. The HBM franchise is genuinely unique — a memory company with GPU-like margins. The NVIDIA relationship is deep. The technology lead is real. The AGI demand thesis is as strong as it gets for a physical infrastructure company.

But the stock has already moved. +572% in 12 months. A forward P/E of 3.8x looks cheap, but it's pricing in peak-cycle earnings that are 3-4x normalized levels. You are not buying a hidden gem — you are buying the most obvious AI beneficiary in the memory space at a price that reflects that.

For a US investor considering significant capital:

1. The ADR listing (expected 2026) is the optimal entry event — it will provide a liquid, US-listed vehicle and may temporarily depress KRX shares as ADR supply is created, or conversely may attract new demand that lifts both.
2. If buying before the ADR, use Interactive Brokers for KRX access.
3. Position sizing should reflect the genuine downside risk (20-25% probability of -18% to -75% drawdown in a cycle turn or geopolitical shock).
4. The currency risk is real and unhedgeable for most retail investors.

vs. Micron: SK Hynix is the better business. It trades at a cheaper valuation. But Micron is the easier investment. For maximum expected return: SK Hynix. For simplicity and risk management: Micron. For a portfolio approach: own both, weighted toward SK Hynix.

Sources: SK Hynix earnings releases (Q1 2026, FY2025, Q3 2025), equity research (Daishin, LS Securities, Citi, KB, Macquarie), TrendForce, Goldman Sachs HBM market analysis, Counterpoint Research, Korea Herald, Seoul Economic Daily, StockAnalysis.com, Patsnap, Introl, LambdaFin.