AI Memory Supercycle: HBM Investing Strategy 2026-2028
Table of Contents
- HBM Market Evolution: From Niche to 500B Industry
- Samsung vs SK Hynix: Competitive Dynamics 2026-2028
- Memory Pricing Cycles: Why This Cycle Is Different
- KOSPI Sector Allocation: Riding the Memory Wave
- MSCI Inclusion Catalyst: The Passive Flow Trade
- My Take
I get asked a lot: "Is this AI memory thing for real, or another semi boom-bust?"
The short answer: this cycle is structurally different from anything we have seen in the last 20 years of semiconductor investing. Here is my framework for understanding it and positioning your portfolio.
HBM Market Evolution: From Niche to 500B Industry
The number that stopped me cold: HBM went from 8% of total memory revenue in 2023 to 35% in 2026. That is not incremental growth. That is a structural rewrite of the entire memory industry. HBM was a niche product used primarily in supercomputers and high-end data centers five years ago. Today it drives the entire memory sector's growth trajectory.
| Year | HBM Market ($B) | % of Total Memory | Leading Product | Bandwidth (GB/s) | Key Customer |
|---|---|---|---|---|---|
| 2020 | 2.1 | 2% | HBM2E | 460 | AMD, NVIDIA |
| 2023 | 40 | 8% | HBM3 | 819 | NVIDIA H100 |
| 2024 | 180 | 18% | HBM3E | 1,200 | NVIDIA B200 |
| 2025 | 350 | 27% | HBM4 early | 2,000 | NVIDIA Rubin |
| 2026E | 520 | 35% | HBM4 12-layer | 2,500 | NVIDIA, AMD, Intel |
| 2028E | 1,000 | 50%+ | HBM4E | 3,500 | Multi-hyperscaler |
Source data compiled from Nomura Securities, Gartner, and IDC industry estimates. The CAGR from 2023 to 2028 is roughly 90%. For context, the entire global DRAM market was approximately $50B in 2023. HBM alone is projected to surpass that figure within two years. That is how fast this market is scaling.
What is driving this? Each NVIDIA GPU generation demands exponentially more memory bandwidth. The H100 uses about 80GB of HBM3 memory. The B200 uses 192GB of HBM3E. The next-generation Rubin architecture reportedly needs 288GB+ per GPU. That is a 3.6x memory content increase per GPU generation, and GPU unit volumes are exploding simultaneously because every hyperscaler is building out AI clusters at unprecedented scale. Meta alone deployed 600,000 H100 equivalents in 2024 and is doubling that in 2025-2026.
Nomura's Junghwon Jeong put it bluntly in their June 2026 briefing: "Memory demand driven by AI will multiply 10,000 to 20,000 times over five years. We are entering a world we have never seen before." That sounds hyperbolic. But look at the monthly memory revenue trajectory — it has gone vertical in ways that even Nomura's 30-year semiconductor analysts say they have never charted before. The vertical slope is visible in monthly industry revenue data published by WSTS and SIA.
The implications for investors are clear. HBM is not a cyclical product. It is a structural growth story driven by the AI capex cycle that has multi-year visibility. Hyperscaler capital expenditure guidance extends through 2029 with cumulative AI infrastructure commitments exceeding $500B across Meta, Google, Microsoft, and Amazon per Bloomberg data.
Samsung vs SK Hynix: Competitive Dynamics 2026-2028
Combined Q2 2026 operating profit for Samsung Electronics and SK Hynix is projected at 150 trillion won ($116B according to Bloomberg FX rates). Compare that to Samsung's full-year 2023 operating profit of 6.57 trillion won. The combined entity has grown profit by roughly 13x in three years. Meritz Securities estimates Samsung Q2 OP at 88.3 trillion won ($68B) and SK Hynix at 64.3 trillion won ($50B).
| Metric | Samsung Electronics | SK Hynix | Advantage |
|---|---|---|---|
| Q2 2026E OP (trillion won) | 88.3 | 64.3 | Samsung |
| HBM Market Share 2026 | ~45% | ~50% | SK Hynix |
| HBM Technology Position | HBM4 12L ramping | HBM4 12L mass production | SK Hynix |
| DRAM Bit Growth QoQ | ~12% | ~18% | SK Hynix |
| Foundry Revenue 2026E | ~$15B | N/A | Samsung |
| Nomura Target Price (won) | 590,000 | 5,000,000 | — |
| Current Price June 2026 | ~350,000 | ~1,600,000 | — |
| Upside to Nomura Target | ~70% | ~212% | SK Hynix |
The competitive picture is more nuanced than the headline market share numbers suggest. SK Hynix got an early lead in the HBM race because they were NVIDIA's primary development partner for HBM3 and HBM3E. That first-mover advantage is real and quantifiable. Their 12-layer HBM4 entered mass production several months ahead of Samsung, and their DRAM bit growth of 18% quarter-on-quarter reflects that volume ramp. Meritz Securities analyst Hyo-seok Lee specifically noted the 18% QoQ DRAM shipment growth driven by HBM4 12-layer mass production.
Samsung has a counterargument that the market is underappreciating. They have the full integrated stack — memory, foundry, and logic. Their foundry business, while slower to improve profitability than the memory side, gives them a captive internal demand base and technology transfer advantages that SK Hynix simply does not have. Samsung's balance sheet also allows them to outspend competitors through down cycles — capital expenditure capacity of $40B+ annually versus SK Hynix at roughly $20B.
Here is what I am watching most closely: Samsung's HBM4 yields. If Samsung closes the yield gap with SK Hynix by Q4 2026, the valuation discount can narrow fast. At the Nomura target of 590,000 won, Samsung trades at roughly 1.2x price-to-book against a 5-year historical range of 1.0x to 1.8x. SK Hynix at 5,000,000 won per Nomura trades at about 2.5x price-to-book — a premium justified by the growth trajectory but one that prices in a lot of perfection. The asymmetric bet is on Samsung catching up.
Memory Pricing Cycles: Why This Cycle Is Different
Every semiconductor investor has the 2017-2018 supercycle burned into their trading memory. DRAM prices doubled in 18 months and then crashed 60% in the subsequent 12 months. The classic pattern: demand surge creates tight supply, which triggers price spikes, which incentivizes capacity expansion, which creates oversupply, which collapses prices. Rinse and repeat. Semiconductor investors have been conditioned to sell the first sign of memory price strength because they know the bust always follows.
This cycle is breaking that pattern for two structural reasons. First, HBM is constrained by advanced packaging capacity — specifically TSMC's CoWoS (Chip-on-Wafer-on-Substrate) — not just traditional fab capacity. CoWoS capacity cannot be doubled in six months. Equipment lead times for advanced packaging tools are 12-18 months. That creates a structural supply bottleneck that does not respond to traditional DRAM pricing signals. Second, the demand driver is AI training and inference workloads from hyperscaler data centers, not PC replacement cycles. This is recurring capital expenditure from companies with trillion-dollar market capitalizations and multi-year planning horizons.
| Attribute | 2017-2018 Supercycle | 2024-2028 Cycle | Investment Implication |
|---|---|---|---|
| Primary Demand Driver | Server + PC refresh cycles | AI training + inference | More structural, less cyclical |
| Key Product Type | DDR4 commodity DRAM | HBM3E/HBM4 proprietary | Higher margins, stickier customers |
| Supply Constraint | Standard fab capacity | CoWoS + advanced packaging | Harder to add supply rapidly |
| DRAM Price Peak Change | +120% peak-to-peak | +45% YoY (H1 2026) | Less spike, more sustainable |
| NAND Price Change | +80% peak-to-peak | +28% YoY (H1 2026) | Following DRAM trajectory |
| Cycle Duration (up phase) | ~18 months | Projected 36+ months | Extended earnings visibility |
| End Demand Stability | Low — macro-driven | High — hyperscaler capex | Multi-year locked-in commitments |
| Competitive Moat | Cost leadership | Technology + packaging capacity | Incumbents strongly advantaged |
The data supports the "this cycle is structurally different" thesis. DRAM fixed-contract prices through the first half of 2026 are averaging 45% above the 2025 full-year average, according to industry price trackers. NAND is up 28% over the same period. The premium for server-grade DDR5 and HBM products over commodity DRAM stays at 4x to 5x — in the 2017 cycle, that premium collapsed as supply caught up within six months. That collapse is not happening yet, and the structural packaging bottleneck suggests it will not happen in 2026 either.
What I track most closely: the monthly memory revenue trajectory that Nomura flagged in their June briefing. It has gone vertical in a way that the 2017-2018 cycle never showed. Back then, revenue growth was steep but roughly linear — a steady upward slope followed by a sharp decline. This curve is exponential. The reason is structural: each HBM generation packs more value per wafer than DDR generations ever could. An HBM4 12-layer stack bonds 12 DRAM dies vertically using through-silicon via interconnects, generating roughly 12x the revenue per unit compared to a single DDR5 die, with only about 3x the manufacturing complexity.
KOSPI Sector Allocation: Riding the Memory Wave
Samsung and SK Hynix combined now account for over 40% of total KOSPI-listed operating profit. That is an unprecedented concentration level for a national equity index. In 2023, Samsung alone contributed roughly 6% of KOSPI earnings. By 2026, the two memory giants plus their ecosystem suppliers (Hanmi Semiconductor, Soulbrain, etc.) represent close to 50% of total index earnings. The KOSPI is effectively a single-bet index in a way that makes sector allocation almost irrelevant.
| Year | Memory Sector Profit Share (%) | KOSPI Forward PER (x) | KOSPI Index Level | Foreign Ownership (%) | USD/KRW |
|---|---|---|---|---|---|
| 2022 | ~12% | 9.8 | 2,236 | 32% | 1,260 |
| 2023 | ~8% | 19.2 | 2,655 | 30% | 1,290 |
| 2024 | ~22% | 12.5 | 4,800 | 28% | 1,380 |
| 2025 | ~35% | 8.2 | 7,200 | 27% | 1,450 |
| 2026 H1 | ~42% | 7.3 | 8,263 | 29% | 1,510 |
| 2027E (consensus) | ~45% | 6.5-7.0 | 9,500-10,500 | 32%+ | 1,470 |
For global investors, this creates an interesting and somewhat uncomfortable dynamic. Buying a KOSPI 200 ETF is effectively an AI memory bet with a Korean won currency overlay. That is not necessarily bad when the memory cycle is strengthening and the won is undervalued on a purchasing power parity basis (REER at 85 versus the 10-year average of 95 per Bloomberg data). But it means portfolio diversification within KOSPI is largely illusory — the index lives and dies with Samsung and SK Hynix.
The valuation argument is compelling on a standalone basis. KOSPI at 7.3x forward PER is pricing in either (a) an earnings recession that is not reflected in any consensus estimate, or (b) a complete market disregard for the memory supercycle earnings trajectory. Samsung trades at roughly 6x projected 2026 earnings. Compare that to NVIDIA at 35x, ASML at 28x, or TSMC at 18x. The "Korea discount" has been a structural feature of the market for years, but when earnings are compounding at 13x over three years, the value argument becomes hard to ignore.
Foreign investors returned in force on the Middle East peace catalyst, buying 2.2 trillion won ($1.7B at the 1,510 exchange rate) in a single session. That broke a 25-trading-day selling streak. If the trend continues, and the MSCI upgrade catalyst provides additional tailwind, foreign allocation to Korea could shift from multi-year underweight back to neutral or overweight. Korea Exchange data confirms the buying was concentrated in semiconductor large-caps and equipment suppliers.
MSCI Inclusion Catalyst: The Passive Flow Trade
Nomura estimates Korea's probability of being added to the MSCI Developed Market Watchlist at 60%. If Korea is upgraded from Emerging Market to Developed Market status, the passive flow implications are enormous. MSCI World Index tracking funds have roughly $3.5 trillion in assets under management. MSCI Emerging Market funds track about $600 billion. Korea's weight in MSCI World would be approximately 1.5% to 2.0% based on current market capitalization, generating $50B to $70B in passive buying demand.
The flow mechanics are favorable. EM funds would sell Korea to reduce their overweight position, generating roughly $10B to $15B in selling. DM funds would simultaneously buy Korea as a new addition to the index. The net effect is strongly positive because the DM fund base is roughly 6x larger than the EM fund base. The 2.2 trillion won ($1.7B) foreign inflow on June 12 represents only about 3% of the anticipated passive flow wave — the real volume comes when MSCI makes a formal decision.
Korea has been pursuing this upgrade for over a decade. The key historical obstacles — foreign ownership limits on individual stocks, short-selling availability restrictions, and currency convertibility concerns — have been largely resolved through regulatory reforms. The MSCI decision timeline suggests a potential announcement in the second half of 2026 with implementation in 2027.
My Take
The AI Memory Supercycle is the defining investment theme of 2026 through 2028, and KOSPI remains the cheapest liquid expression of that theme in global equity markets.
I see the memory cycle extending through at least 2028 for three reasons that form my investment thesis. First, hyperscaler capex commitments are locked in with multi-year visibility — Meta, Google, Microsoft, and Amazon have cumulative AI infrastructure budgets exceeding $500B through 2029. Second, HBM content per GPU is still in the early innings of expansion — the jump from HBM3 at 80GB per GPU to HBM4 at 288GB triples the addressable market per chip shipped. Third, the supply constraints on CoWoS and through-silicon via packaging capacity are not resolving materially before 2027, creating a sustained pricing power environment for HBM suppliers.
Two concrete actions I am taking with my portfolio:
- Increasing KOSPI large-cap exposure via the KODEX 200 ETF. At 7.3x forward PER with 40%+ earnings growth, this is the cheapest AI proxy available in any developed market globally. The MSCI upgrade catalyst adds an option value that the market is not pricing in. I am targeting 15% to 20% of my Asia ex-Japan equity allocation in this trade.
- Adding selective memory equipment exposure — specifically Hanmi Semiconductor and related capital equipment suppliers. The HBM buildout requires testers, thermal compression bonders, and wafer probing equipment regardless of which memory manufacturer wins market share. These suppliers benefit from HBM capital expenditure cycles without taking single-name Samsung or SK Hynix risk.
The risk I watch most closely: if the 10-year US Treasury yield breaks above 4.7%, the rate shock would compress equity multiples globally, including KOSPI. I maintain 15% to 20% cash in the Korea allocation specifically for that scenario. But my primary position is buying KOSPI weakness, not selling it. The 1998 analog — midterm US election year, peak Federal Reserve rates, strong dollar environment — suggests a powerful rally once macro uncertainty clears, and Korea has historically been one of the best-performing markets in that specific macro regime.
What could change my mind: a collapse in hyperscaler capital expenditure guidance (unlikely given the competitive dynamics of the AI arms race), or a Samsung HBM4 yield disaster that cedes the entire HBM market to SK Hynix (possible but partially priced in at current valuations relative to the Nomura target). I am monitoring Q3 2026 earnings calls closely for HBM margin and yield disclosures from both companies.
Comments
Post a Comment