The iShares MSCI South Korea ETF (EWY) is on pace for its best week since December 2008. The fund has gained +16.6% in five sessions — a move that would rank among the top 10 weekly returns in EWY's 25-year history. The multi-timeframe scoreboard:

This is not a Korea story. It is a memory story — and Korea owns the memory market.

The Bottleneck Nobody Saw Coming

The AI buildout has a constraint that most analysts underweighted until this year: every GPU needs memory, and the world is running out of it.

Nvidia's flagship H200 and B200 GPUs — the units that Microsoft, Google, Amazon and Meta are ordering in the hundreds of thousands — each require High Bandwidth Memory (HBM). HBM is not conventional DRAM. It is a stacked 3D memory architecture that sits directly on the same package as the GPU die, connected by thousands of micro-bumps, providing 10-15x the bandwidth of standard DDR5. Without HBM, the GPU cannot function at data centre scale.

The problem: HBM cannot be manufactured at speed. The production process requires advanced packaging — specifically TSMC's CoWoS (Chip on Wafer on Substrate) technology, which bonds the HBM stacks and the GPU logic die onto a silicon interposer. CoWoS capacity is sold out through 2027. Samsung's CEO is now seeking multi-year contracts of three to five years with major customers, a direct acknowledgment that the shortage is structural, not cyclical.

TrendForce called it in January 2026: "Memory Wall Bottleneck: AI Compute Sparks Memory Supercycle." Fortune in February: "Rampant AI demand for memory is fueling a growing chip crisis." Bitget research identified "RAMageddon" — DRAM prices surged 75% from December to January alone as hyperscaler procurement teams exhausted spot market inventory.

The supply side cannot respond quickly. Building a new HBM-capable fab requires 18-24 months of construction and 12 months of qualification. The companies that already have HBM capacity — SK Hynix and Samsung — are the only game in town. And customers are now booking supply not for 2026, but for 2027 and 2028. SanDisk reported customers booking storage capacity two years forward. Kioxia guidance showed multi-year forward bookings. This is not a quarterly demand spike. It is a structural realignment of the memory supply chain.

Why South Korea Wins

Korea's dominance in this trade is not coincidental — it is the product of 40 years of industrial policy, chaebol capital allocation and a willingness to invest through cycles when Western competitors retreated.

SK Hynix is the dominant HBM supplier globally, holding approximately 50% of all HBM3E production capacity. The company has been supplying every Nvidia H100, H200 and B200 GPU with its HBM stacks. No other company has the yield rates, packaging technology and supply relationships to displace SK Hynix in the near term. Samsung is aggressively catching up — its HBM3E yields were initially lower than SK Hynix's, but Q1 2026 earnings confirmed the gap has narrowed materially, with Samsung now delivering HBM at scale to multiple hyperscalers.

The combined result: SK Hynix (000660.KS) now represents 23.17% of EWY — its highest weight ever in the fund — and Samsung Electronics (005930.KS) represents 21.38%. Together, the two memory champions control 44.55% of EWY's NAV. When both stocks move vertically, EWY moves vertically. There is no diversification buffer.

EWY's AUM has surged to $23.67 billion — up $2.8 billion in 48 hours as institutional inflows chased the performance. This is not passive rebalancing. This is active capital allocation into the most concentrated expression of the memory supercycle available in ETF form.

Nvidia's own supply chain data reinforces the Korea thesis: Asian suppliers now represent approximately 90% of Nvidia's production costs, up from 65% a year ago. The semiconductor supply chain has become more Korea-dependent with every generation of AI infrastructure, not less.

The Valuation Case: Cheapest Large-Caps in the Market

The most analytically striking feature of the memory trade is the valuation profile. The companies at the centre of the most disruptive technology cycle in decades are priced like mature, cyclical commodity producers.

SK Hynix at 5.1x NTM earnings is one of the cheapest large-cap technology companies in the world. For context: the S&P 500 trades at 21.5x forward P/E, Nvidia at approximately 35-40x, and the MSCI ACWI at 18.07x. A company that supplies the irreplaceable memory component for every Nvidia GPU being deployed in AI data centres globally is priced at a quarter of the market's multiple.

The explanation is structural and partially rational: memory is a cyclical business. The decade of commodity DRAM pricing — where Samsung, SK Hynix and Micron competed purely on cost and destroyed each other's margins — left institutional memory (no pun intended) of what these stocks do in downturns. In 2022-2023, Samsung and SK Hynix both posted operating losses as memory prices collapsed. The market is pricing some probability of that cycle repeating.

The counter-argument, gaining traction: this time, the demand driver is not PC DRAM or mobile storage — it is HBM for AI accelerators, a market that did not exist at scale three years ago. The customers are hyperscalers with multi-year capex commitments measured in hundreds of billions of dollars. Samsung is now signing 3-5 year supply contracts precisely because the old quarterly spot market dynamic no longer applies. Benzinga's headline this week captures the tension: "Micron Stock Logs Best Week Since 2008: Why The Price Tag Is Still Cheap."

At 5.1x and 5.3x NTM P/E, SK Hynix and Samsung trade as if their earnings will be cut in half within two years. If the HBM supercycle persists — and the forward booking data suggests it will through at least 2027 — those multiples would compress to 2-3x on current run-rate earnings, making the stocks extraordinarily cheap even after vertical moves.

Micron: The US Signal

Micron Technology's +30% weekly gain — its best week since December 2008 — is the US market's confirmation of what Korean stocks have been pricing for months. Micron is SK Hynix's closest Western analog: a DRAM and NAND manufacturer with growing HBM exposure. Micron at 8x NTM P/E is more expensive than SK Hynix (5.1x) but still cheaper than virtually any other major technology company.

The Micron signal matters for EWY because the US market's validation of the memory trade removes the "emerging market discount" narrative from SK Hynix's valuation. When Micron — listed in the US, followed by hundreds of analysts, owned by every major fund — trades at 8x forward earnings and posts its best week since 2008, the argument that SK Hynix at 5.1x is "structurally cheap because it's Korean" becomes harder to sustain.

SanDisk's +450% YTD performance is the most extreme expression of the NAND component of the trade. SanDisk — spun off from Western Digital, focused entirely on NAND flash storage — has benefited from the same AI data centre buildout that is driving HBM demand. Every AI training run requires massive amounts of flash storage for datasets; every inference deployment requires persistent storage for model weights. SanDisk at 9.5x NTM P/E, despite its +450% YTD gain, is being bid by buyers who believe NAND prices have structurally reset higher for multiple years.

The Technical Warning: Extreme Distance from Moving Averages

This is where the bull case requires the most intellectual honesty. The fundamental argument for the memory trade is compelling — supply bottleneck, structural demand, record earnings, cheap multiples. The technical picture is a different conversation.

These are not normal technical conditions. A stock trading 157% above its 200-day moving average — SK Hynix — has moved three standard deviations or more above its long-run price trend. In the history of large-cap equities, moves of this magnitude above the 200-day SMA have been associated with two distinct outcomes: sustained fundamental re-ratings (where the old SMA becomes permanently obsolete as earnings reset higher) or violent mean-reversion corrections (where the price catches down to the moving average, not vice versa).

The 295% gap for SanDisk is extraordinary. For context: SanDisk's 200-day moving average reflects a year of price history — including the low-single-digit share price the stock traded at before the memory supercycle began. The gap has been created by a near-vertical price appreciation that the moving average, by construction, cannot immediately reflect. This creates a mechanical setup: any pause in the fundamental narrative — a single quarter of earnings disappointment, a hyperscaler reducing AI capex guidance, or simply profit-taking after a 450% year — would send the price toward the moving average with the same velocity it moved away from it.

For EWY specifically, the implication is direct. SK Hynix at +157% above its 200-day SMA is EWY's largest position (23.17%). Samsung at +112% is the second largest (21.38%). The fund is technically stretched at every level — individual stock, sector and country. The fundamental case for holding does not require a technical argument. But risk managers who set position limits using moving average distances are today looking at EWY and seeing a fund that has never been this extended.

The historical precedent that markets consistently reference: December 2008 was the last time Micron had a week like this. What followed was not a sustained trend — it was a violent reversal that gave back most of those gains within months as the financial crisis deepened. The parallel is imperfect — the 2008 context was macro-driven, while the 2026 context is fundamentally driven. But the technical signal is real and warrants acknowledgment.

The $2.8 Billion Inflow Problem

EWY's AUM jumped from $20.91 billion to $23.67 billion in 48 hours — a $2.76 billion inflow in two sessions. This matters because late-cycle inflows into an already-extended fund create a mechanical risk: the ETF must buy the underlying Korean stocks at elevated prices, increasing the float-weighted ownership of SK Hynix and Samsung at precisely the moment they trade at maximum distance from their moving averages.

When retail and institutional capital chases performance into an ETF, the ETF's buying pressure supports the very prices that make the technical overextension worse. The self-reinforcing dynamic works on the way up — it can reverse equally fast on the way down if fund flows reverse.

Track EWY live at countryetftracker.com/country-detail?ticker=EWY. Full valuation data at countryetftracker.com/valuation.

The Core Tension

The memory trade in 2026 presents one of the most acute tensions in modern equity analysis: a genuine, documented, structural supply bottleneck — with multi-year forward bookings, record earnings, CEO-level supply agreements and analysts explicitly calling a supercycle — coexisting with the most extreme technical overextension in the memory sector's history.

The fundamental investors are correct that SK Hynix at 5.1x earnings is cheap. The technical analysts are correct that a stock trading 157% above its 200-day moving average is dangerous. Both are true simultaneously. The resolution will be determined by whether Q2 2026 earnings — due in July from Samsung and SK Hynix — confirm or disappoint the supercycle thesis.

If Samsung and SK Hynix deliver Q2 results that match or exceed Q1's record levels, the moving averages become irrelevant — they will simply reset higher as the price anchors at elevated levels long enough for the 200-day to catch up. If either company guides conservatively for H2 2026, the distance from the moving average becomes the dominant risk factor, and the correction will be proportional to the extension.

Frequently Asked Questions

Why does SK Hynix trade at 5.1x earnings if it is the dominant HBM supplier?

Memory stocks have historically traded at low multiples because memory is a cyclical commodity — in 2022-2023, Samsung and SK Hynix both posted operating losses when DRAM prices collapsed. The market prices some probability of that cycle repeating. The counter-argument: HBM for AI accelerators is not conventional commodity DRAM. Hyperscalers are now signing 3-5 year supply contracts, fundamentally changing the demand visibility that previously justified single-digit multiples. If the AI supercycle persists, 5x earnings on record profits implies the stock is worth substantially more — or the earnings need to be cut by 70%+ for the valuation to look expensive.

What is the HBM bottleneck exactly?

High Bandwidth Memory (HBM) is a stacked 3D memory architecture where multiple DRAM dies are vertically connected and bonded directly onto the same silicon package as the GPU compute die. The bonding process — CoWoS advanced packaging — can only be done by TSMC, which has limited capacity for the silicon interposers required. As of May 2026, TSMC's CoWoS is sold out through 2027. SK Hynix controls approximately 50% of global HBM3E production. Samsung controls approximately 35-40%. Micron has less than 10%. There is no alternative supplier. There is no technology substitute. The bottleneck is physical and cannot be resolved in less than 18-24 months even with unlimited capital.

What would cause the memory trade to reverse?

Three scenarios could reverse the move: (1) a major hyperscaler — Microsoft, Google, Amazon or Meta — reduces AI capex guidance in Q2 earnings, reducing HBM demand below current trajectory; (2) Samsung or SK Hynix guides Q2 2026 below consensus, signalling that Q1's record was a peak rather than a floor; (3) a broader market risk-off episode that forces institutional investors to reduce exposure regardless of fundamentals. None is the current base case. The forward booking data — customers reserving capacity into 2027-2028 — provides the most structural evidence that scenario (1) is unlikely in the near term.

Why is SanDisk up +450% YTD at 9.5x earnings?

SanDisk is a NAND flash memory pure-play spun off from Western Digital. NAND flash is used for AI dataset storage, model weight persistence and inference deployment infrastructure — the storage layer of the AI stack that runs below the GPU compute layer. As AI training run sizes increase exponentially, the storage requirements scale proportionally. SanDisk's 9.5x NTM P/E reflects the market's belief that NAND prices have structurally reset higher through at least 2027, as Kioxia guidance showed customers booking capacity into 2028. At 295% above its 200-day moving average, SanDisk is the most technically extended large-cap equity in the memory sector.