No Matter Who Wins, Memory Wins

MEMORY · SOVEREIGN AI · 2026.06.14

Whoever wins,

memory wins

Samsung Electronics · SK Hynix

005930 · 000660

No matter who wins the chip war — every accelerator needs fast memory

June 14, 2026 · Narrative constructed after research and multiple discussions by the Goguma Team
Transition to a contract industry · Essential sovereign AI commodity · Balancing between the U.S. and Korea · Bull/Ultra-Bull take-profit targets

Onion (Macro/Trade) · Acorn (Industry/Substance) · Peanut (Risk/Value) · Carrot (Chart/Flow) · Goguma (Summary)

Prologue

Who is the real winner of the chip war?

The world is currently fighting over 'who makes the AI chips.' From NVIDIA's graphics processing units (GPUs) and custom semiconductors (ASICs) designed by Broadcom and Marvell, to the self-developed inference chips that countries are trying to build themselves. However, all these chips have one thing in common — no chip can operate without fast memory (HBM). No matter who wins the chip war, memory wins.

This is the starting point of this report. The market is focused on the market share battle between NVIDIA and Broadcom, but the picture changes when you take a step back. The performance of an AI accelerator is increasingly dependent not on the computing chip, but on how quickly the memory can feed it data. Therefore, High Bandwidth Memory (HBM) is the 'heart pump' that must be attached to every computing chip. And there are effectively only two companies that make that pump — SK Hynix and Samsung Electronics.

Even more significant changes are overlapping. First, the memory industry itself is shifting from a 'market-price business where prices are negotiated every quarter' to a 'contract industry where demand and supply are committed for years.' Second, as countries declaring 'AI sovereignty' begin building their own chips and data centers, memory demand is expanding from a few big techs to dozens of nations. As the types of chips become more diverse, the demand for the memory that goes into all of them becomes more dispersed and robust.

This report answers three things in order. Why is memory no longer a cyclical business? (Transition to a contract industry). Why does sovereign AI add 'additional upside' to memory? (Common essential commodity). And amidst U.S. trade pressure and the international dynamics of 'U.S. production vs. Korean production,' how high can the valuations of these two companies go — up to their final take-profit targets. We answer these questions based on multiple team discussions and data from brokerage and industry experts.

Chapter 1

The old belief — "Memory is ultimately cyclical"

Forward price-to-earnings (P/E) ratios were at one point around 6x for Samsung and 5x for SK Hynix. Even as they headed toward record-high profits, the market only assigned single-digit multiples. This is not evidence of undervaluation, but a trace of an old belief.

For decades, memory has been defined by the 'repetition of boom and bust.' When prices rise, everyone builds factories; when supply overflows, prices crash; and companies endure losses until the next boom—a chicken game. Therefore, the market learned — "When memory profits are at their best, it is the most dangerous time." This belief is etched into valuations. The paradox where the P/E ratio is lowest when profits are at their peak — this is because the market discounts them in advance, assuming "these profits will soon collapse."

However, at the bedrock of the industry, the premise of that belief is shaking. The subjects of demand have changed (consumers → AI data center infrastructure), the transaction methods have changed (quarterly bargaining → 5-year long-term contracts), and the yardstick the market uses to measure memory has begun to change (PBR based on assets → P/E based on earnings). When these three changes happen simultaneously, the premise that "memory is cyclical" itself collapses.

Chapter 2

Where the belief breaks — From market-price business to contract industry

The essence of the change is this: Memory is moving from a 'market-price business where prices are negotiated every quarter' to a 'contract industry where demand and supply are committed for years.' The very structure that created the amplitude of the cycle is changing.

AI customers now see 'securing volume' as a bigger risk than 'price.' Therefore, they sign 5-year Long-Term Agreements (LTAs) with Samsung, SK Hynix, and Micron to lock in future supply. HBM volumes for 2026 are already sold out, and margins are 3 to 5 times higher than general-purpose DRAM. The industry where prices fluctuated every quarter is changing into an industry where performance is visible in advance through contracts.

The securities industry began to read this transition first — by changing the valuation method from asset-based (PBR) to earnings-based (P/E). This is not just a target price upgrade, but a declaration that "we will no longer view memory as a cyclical industry." And the most powerful witness to that transition appeared in Seoul in June.

Jensen Huang · NVIDIA CEO

"SK Hynix has been NVIDIA's largest memory partner, and it will continue to be. We are already procuring billions of dollars worth every year, and it will only grow larger." — At the same venue, he discussed HBM4, HBM5, and foundry cooperation with Samsung Electronics, and formalized a multi-year technical partnership between NVIDIA and SK Hynix for next-generation memory for AI factories.

Key takeaways of Chapter 2. Demand is shifting from replacement to infrastructure (trending), transactions from bargaining to contracts (reduced volatility), and the yardstick from PBR to P/E (revaluation). These three changes are the logical basis for resolving the P/E 5-6x undervaluation. However, the cycle hasn't disappeared, just postponed — revaluation will occur while inventory is normal and contracts support earnings (until 2027-2028).

Chapter 3

An arms dealer to the arms dealers — Sovereign AI as a new demand axis

Added to this, beyond the transition to a contract industry, is a second driver. It is 'AI sovereignty (Sovereign AI).' And the core of this driver is paradoxical — the more types of chips there are, the broader the demand for memory becomes.

The world is currently trying to break free from its dependence on U.S. models and U.S. chips. The 2026 incident where frontier AI models became subject to U.S. export controls accelerated this. Countries have begun building their own models, data centers, and chips, with investment scales reaching $500 billion for the U.S. Stargate, $1 trillion for Saudi Arabia, €200 billion for the EU, and 150 trillion KRW for South Korea. Yet, no matter which chip these massive funds flow into — whether NVIDIA GPUs, custom chips from Broadcom/Marvell, or indigenous inference chips built by each country — HBM is the component that all of them commonly consume.

$54.6 billion

2026 HBM market size forecast (+58% YoY) — A common essential commodity consumed by all chips regardless of AI accelerator type

This is the meaning of the 'arms dealer to the arms dealers' analogy. Just as those who sell pickaxes in a gold rush profit more stably, whoever wins the AI chip war, the company that makes the memory that goes into those chips wins. Moreover, sovereign AI disperses demand — whereas memory demand was previously concentrated in a few U.S. big techs, it now expands to data centers in dozens of countries. The structure that was once at the mercy of one customer's whims is being replaced by robust, broad demand spread across many nations.

Key takeaways of Chapter 3. Sovereign AI adds additional upside separate from the 'contract industry transition.' ① Demand sources are dispersed and expanded from big tech to dozens of countries, ② HBM is commonly required regardless of which computing chip wins, and ③ the importance of memory capacity and bandwidth grows as we move into the era of inference. Demand that thickens regardless of the chip battle — that is the place of memory.

Chapter 4

SK Hynix — The King of HBM, and 'hedging its bets'

SK Hynix is the most direct protagonist of this story. It is the overwhelming #1 in the HBM market, NVIDIA's largest memory partner, and the biggest beneficiary of the transition to a contract industry.

Numbers speak to its stature. SK Hynix is estimated to supply about 60~70% of the next-generation HBM4 and has signed a multi-year technical partnership with NVIDIA. Net profit paths of approximately 101 trillion KRW in 2026 and 142 trillion KRW in 2027 are being discussed, with long-term contracts supporting that earnings visibility. HBM volumes for 2026 are already sold out.

Indicator	Value	Significance
HBM4 Market Share (Est.)	Approx. 60~70%	Effective oligopoly in next-gen memory
Net Profit Path	2026 ~101T / 2027 ~142T	Visibility supported by contracts
HBM Margin	3~5x of general DRAM	Core of product mix improvement
NVIDIA Relationship	Largest memory partner · Multi-year contract	"Will continue to be" (Jensen Huang)

However, the real issue in 2026 is not market share, but 'where it is made.' While the U.S. is pressuring South Korea over its massive trade surplus, it is difficult to impose tariffs on HBM itself — because the U.S. has almost no domestic HBM production, and tariffs would immediately mean higher costs for U.S. AI data centers. Therefore, U.S. pressure is not directed at tariffs, but at "produce it in the U.S." SK Hynix is responding by building an advanced packaging and HBM hub in Indiana (operational in 2028).

The structure of hedging. For SK Hynix, the U.S. base (Indiana) is a 'hedge' to soothe trade tensions and avoid tariffs, while the Korean base (Icheon/Cheongju) is the main headquarters that generates actual profits. Furthermore, the large-scale GPU infrastructure being secured by the Korean government (national data centers/AI factories) acts as captive demand, laying a defensive line against a down-cycle. In short, the core valuation variable for SK Hynix is not its 'U.S. fab,' but the 'HBM margin of its Korean fab.'

Chapter 5

Samsung Electronics — HBM4 pursuit + the second card of Foundry

Samsung Electronics' story is 'two options within one company.' Memory (HBM4 pursuit) and Foundry (semiconductor contract manufacturing). And the trade climate of 2026 elevates the value of this second card — foundry — in a new way.

Option 1 — HBM4 Pursuit

Samsung is assessed to be meaningfully re-entering the NVIDIA supply chain by securing about 25~30% supply for the next-gen HBM4. It is the company with the greatest room to resolve its deepest undervaluation (where forward P/E was once stuck at 6x) through HBM4 entry and memory revaluation. The nearly 8% surge in its stock price in a single day in June also reflects this expectation.

Option 2 — Foundry, a bargaining chip in the 232 trillion era

Samsung is a memory company and simultaneously a foundry operator second only to TSMC. The trade climate of 2026 amplifies this card. As the U.S. pressures companies to "build in the U.S. to avoid tariffs," Samsung is constructing an advanced process (2nm) fab in Taylor, Texas. This 'U.S. production base' is in itself a card to appease trade issues and avoid tariffs, and at the same time holds the potential to emerge as a 'second foundry' for NVIDIA, Tesla, etc. (having already won some orders for NVIDIA-affiliated inference chips). While outsourcing cutting-edge main chips is difficult in the short term, entry via inference/segment chips and advanced packaging (I-Cube) is a realistic path.

Option	Details	Significance
HBM4 (Memory)	Approx. 25~30% share re-entry	Key to resolving P/E 6x undervaluation
Foundry (Contract Mfg)	U.S. Taylor 2nm + Inference chip orders	Option barely reflected in current value
U.S. Fab	Tariff avoidance/trade appeasement card	Bargaining power in the 232T tariff era
Korean Fab	HBM/Memory stronghold	Actual profit leverage

Key takeaways of Chapter 5. If SK Hynix is a 'pure HBM bet,' Samsung Electronics is a dual option of 'Memory (HBM4 pursuit) + Foundry (hedging).' In particular, because foundry is currently valued at nearly zero due to losses, the turnaround to profit and the bargaining value of its U.S. fab remain as pure upside — which is why Samsung's risk-reward structure is more asymmetric in the same AI cycle.

Chapter 6

U.S. or Korea — The economics of hedging

The deepest issue in 2026 memory investment is not technology, but geopolitics. With Korea's trade surplus with the U.S. exploding due to the HBM supercycle, the U.S. will not stand idly by. However, they cannot impose tariffs on memory. This contradiction shapes the production strategies of both companies.

The team dissected this structure through multiple discussions. The key was the 'separation of two goals' — the path to appease trade surpluses and the path to secure AI computing sovereignty cannot be solved with the same move.

Onion · Macro/Trade (Bullish)

"The U.S. cannot impose tariffs on HBM — domestic self-sufficiency is effectively zero, and tariffs would immediately mean higher costs for their own AI data centers. Therefore, the pressure is not on tariffs, but on 'produce it in the U.S.' Korea has no choice but to resolve this by hedging between U.S. fabs (to appease trade) and Korean fabs (for profit)."

Acorn · Industry/Substance (Balanced)

"Regardless of which computing chip wins, HBM will be used by all — that is the structural position of memory. However, buying back US-made chips to Korea is blocked by 'supply priority' regulations during a supercycle. Compute sovereignty must ultimately be secured through domestic production in Korea."

Peanut · Risk/Value (Cautious)

"Do not forget the paradox. The more Korea achieves self-sufficiency, the weaker the bargaining leverage that 'US AI stops without HBM' becomes. Also, US fabs are 'cost centers' due to subsidy cuts and margin dilution. Note that the cycle has been delayed, not disappeared — if inventory exceeds 8 weeks, that is the signal."

Carrot · Chart/Supply-Demand (Practical)

"The key variable for value is not the US fab, but the HBM margin of Korean fabs. The large-scale domestic guaranteed demand set by the government creates a defensive line for the downcycle. SK Hynix has already priced this in significantly, while Samsung is more attractive in terms of risk/reward as its foundry option remains unpriced."

The conclusion of the discussion was clear. Covering deficits through US production and investment in the US, while maintaining compute sovereignty and actual profits through Korean production — the answer is a 'two-pronged' strategy that operates both pillars separately. The key is to build both pillars simultaneously within the golden time (approx. 3 years) when the leverage that "US AI stops without HBM" is valid. In the meantime, the domestic guaranteed demand set by the government will support the bottom of the downcycle.

Chapter Seven

Valuation — Bull/Ultra-Bull Take-profit Targets

When the yardstick shifts from cycle (PBR) to contract industry (PER), and the common demand for Sovereign AI is added — the values of these two companies cannot remain at single-digit multiples. The take-profit targets below are three paths drawn based on brokerage price targets, earnings forecasts, and industry facts.

SK Hynix (000660) — Current approx. 2,150,000 KRW

Scenario	Take-profit Target	Rationale
Base	2,600,000 KRW +21%	Transition to contract industry + PER yardstick normalization to ~101 trillion KRW in 2026 net profit.
Bull	3,000,000 KRW +40%	60~70% market share in HBM4 + multi-year partnership with NVIDIA. Brokerage target 3 million KRW.
Ultra-Bull	3,800,000 KRW +77%	2027 net profit ~142 trillion KRW + diversification/expansion of Sovereign AI common demand. Top of consensus 3.8 million KRW.

Top pick for pure HBM betting. Indiana (US) serves as a tariff hedge; the core of value is the HBM margin of Korean fabs. Government captive demand provides downside protection.

Samsung Electronics (005930) — Current approx. 322,500 KRW

Scenario	Take-profit Target	Rationale
Base	400,000 KRW +24%	Re-entry into HBM4 + Memory PER normalization (6→10x range).
Bull	500,000 KRW +55%	PER normalization + profit contribution from HBM4 and foundry. Brokerage target 500k KRW.
Ultra-Bull	650,000 KRW +102%	HBM5 + foundry turnaround + US Taylor origin card (simultaneous re-evaluation of memory and foundry).

Dual option of memory+foundry. Foundry is barely priced in — turnaround and US fab negotiation value are pure upside. Asymmetric risk/reward.

Two stocks, one conclusion. SK Hynix is the most direct beneficiary of Sovereign common demand through a 'pure HBM bet' (already significantly priced in), while Samsung Electronics has a more asymmetric upside through unpriced options of 'HBM4 catch-up + foundry two-pronged strategy.' Both stand on three pillars: ① transition to contract industry, ② Sovereign common essential goods, and ③ two-pronged strategy (tariff hedge + domestic captive). Execution should be done by selling not all at once — partially at Bull, and additional at Ultra-Bull.

Chapter Eight

Signals that the story is deviating

The transition to a contract industry has not eliminated the cycle, but delayed it. Therefore, we pre-define the signals for the return of that cycle and for potential reversals in trade conditions.

Signal	Meaning	Response
HBM/DRAM inventory exceeding 8 weeks	Currently 2~4 weeks. Crack in the contract industry logic — warning of return to oversupply cycle.	Accelerate partial profit-taking at Bull/Ultra-Bull. Re-examine transition logic.
Slowdown in LTA renewals	Weakening of the premise that 'contracts support performance'.	Reduction in visibility premium — conservative re-evaluation.
US Section 232 Phase 2 (Direct HBM tariffs)	Currently not directly targeting HBM. Margin pressure if expanded to specific Korean tariffs.	Emphasis on the value of origin hedging via US fabs (Indiana/Taylor).
Bargaining power paradox	Increased Korean self-sufficiency → weakening leverage that "US AI stops without HBM".	Check whether both prongs are being built simultaneously within the golden time.

Closing. On the stage of the chip war, everyone asks for the winner of the computing chip, but off-stage, all those chips consume the same memory. The yardstick for memory is changing from 'market-driven trade' to 'contract industry,' Sovereign AI is expanding that demand to dozens of countries, and trade conditions have positioned both companies on a two-pronged strategy of 'US fabs (hedge) and Korean fabs (profit).' SK Hynix is the HBM king at the center of it, and Samsung Electronics is an asymmetric option holding two cards: memory and foundry. Whoever wins, memory wins — but during the time the cycle is delayed, the discipline of watching for signals and scaling in and out is the way to walk the double path to the end.

This report is a narrative analysis summarizing the research and multiple discussions of the Goguma Team (Yangpa, Dotori, Ttangkong, Danggeun, Goguma).
The proposed targets are scenarios based on brokerage consensus and the latest industry trends, and do not constitute investment advice. Investment decisions and responsibilities rest solely with the investor.
As of June 14, 2026 · Samsung Electronics (005930) · SK Hynix (000660)