The Day SK Hynix Overtook Samsung: New Rules of the Semiconductor Supercycle

In late October 2025, the Q3 earnings reports for both companies were released side by side.

For the first time since its founding, SK Hynix recorded a quarterly operating profit of 11.3834 trillion KRW. In the same quarter, Samsung Electronics’ Device Solutions (DS) division posted an operating profit of 7 trillion KRW.

It wasn’t that Samsung’s semiconductor business had a poor quarter. In fact, it was a quarter of perfect rebound, generating the highest memory revenue in its history. Yet, it fell behind SK Hynix in profit. This was the moment the name that had held the top spot in the memory market for nearly 20 years moved down to the second line.

I believe these two figures summarize the nature of this boom most concisely: the rules of the game have changed from a competition where the company that produces the highest volume wins, to one where the company that makes the most necessary chips first wins.

What made the 2025 Semiconductor Supercycle different?

The engine of this supercycle is corporations, not consumers.

Previous booms—driven by laptops in the early 2000s, mobile/smartphone upgrade cycles in the 2010s, and the cryptocurrency mining boom—were mostly B2C demand generated by individuals purchasing devices.

Demand in 2025 comes from data centers running Generative AI. Big Tech and cloud operators are racing to build ‘digital power plants’ to run the massive brain that is AI, and the core components of these plants are AI accelerators (GPUs) and High Bandwidth Memory (HBM). The two components exploded together, relying on each other; GPU performance depends on how quickly HBM can transfer data.

This difference has changed the economics of semiconductors.

Previous cycles were battles over how much, and how cheaply, one could make general-purpose products like DDR and LPDDR. It was a price-sensitive market.

In contrast, companies entering the AI service competition do not strictly mind the cost if it guarantees peak performance. Since demand is insensitive to price, the company that excels at making high-unit-price HBM rakes in the profits.

The center of gravity has shifted from economies of scale to economies of technology.

The ‘Tilting Effect’: How HBM pushed up total DRAM prices

The HBM boom didn’t just make HBM expensive; it pushed up the prices of general-purpose DRAM as well. The industry calls this the Tilting Effect.

The principle is as follows: Since the profitability of HBM is overwhelmingly higher than that of standard DRAM, Samsung Electronics and SK Hynix concentrated their limited production capacity and core talent on HBM. HBM consumes significantly more wafers to produce the same capacity. Consequently, the production of general-purpose products like DDR5 for PCs and mobile devices decreased relatively.

The result was immediately reflected in price tags. According to The Guru’s report, the spot price of DDR5 16Gb jumped more than 100% in one month, from 7.68 USD at the end of September 2025 to 15.5 USD at the end of October. With supply tightening even for older products like DDR4 and LPDDR4X—which had been pushed out by the shift to advanced processes—a rare scene emerged where legacy product prices rose in the second half of the year.

The new rules of the semiconductor supercycle where SK Hynix surpassed Samsung

The reason SK Hynix moved ahead can be summarized in one sentence: It was the first company to put the highest-performing HBM into NVIDIA.

As a first-mover in the HBM market, SK Hynix virtually monopolized the supply of HBM3 and HBM3E to NVIDIA, climbing to the top of the DRAM market in Q3 2025. Conversely, Samsung Electronics lagged behind as its NVIDIA quality validation for HBM3E was delayed. While a king of production volume, it failed to become the first-mover.

At the root of this gap lies stacking technology. HBM involves stacking DRAM chips vertically like an apartment building.

SK Hynix uses the MR-MUF method, where a liquid protective material is poured between chips and hardened all at once. The process is relatively simple and advantageous for heat dissipation, which was decisive in boosting yields.

Samsung Electronics and Micron have used the TC-NCF method, which involves placing a thin film between each layer and pressing it with heat and pressure, a process known to be more complex.

That is why Samsung’s gaze is fixed on the next generation. It stated that it is in “close discussions” for HBM4 supply to NVIDIA and has significantly increased its HBM production plan for 2026. For Samsung, HBM4 is not just about market share recovery, but about reclaiming lost technological leadership.

Why HBM4 ties memory and foundry into a single product

Starting with HBM4, it will become difficult for a memory company to complete HBM alone. The ‘base die’ at the bottom of the chip stack acts as the brain, and this die must be manufactured using advanced foundry processes.

Until now, what connected chip to chip were small bumps called ‘micro bumps.’ However, in generations requiring higher stacks—like the 16-layer HBM4 or 20-layer HBM5—the bump method hits physical limits. The alternative is hybrid bonding. By eliminating bumps entirely and bonding copper to copper directly, chip thickness is reduced, and transmission speed and heat dissipation are both improved.

An even bigger change is the base die. The HBM4 base die is not just a support base; it contains customized logic circuits desired by customers like NVIDIA.

These circuits require advanced foundry processes of 5nm or below, not memory processes. This is the point where the two pillars of the semiconductor industry, memory and foundry, are forcibly merged into one product.

The landscape is divided accordingly. SK Hynix, which only possesses memory processes, has partnered with foundry leader TSMC. In this structure, TSMC makes the base die and stacks SK Hynix’s memory on top, but TSMC is also eyeing leadership in HBM stacking through this process.

For Samsung, this is both a crisis and an opportunity. Samsung is the only company in the world that possesses both advanced memory and advanced foundry capabilities. If both business units are properly integrated to launch a ‘one-stop’ HBM4, it becomes a card worth playing against the SK-TSMC alliance.

The history of chicken games: Winter always follows the semiconductor supercycle

The history of the semiconductor industry is summarized by cycles of alternating extreme booms and harsh busts, and the ‘chicken games’ in between.

Like a game where two cars drive toward each other and the one that swerves first loses, companies engage in bloodletting competition until someone collapses.

In the 1980s, the chicken game was fought between nations.

When Japanese firms effectively dominated the DRAM market through aggressive investment, the U.S. pushed for the U.S.-Japan Semiconductor Agreement twice, in 1986 and 1991. The agreement aimed to boost the market share of U.S.-made semiconductors in Japan to 20% and prevent dumping of Japanese DRAM. Price controls and investment contraction gradually eroded the competitiveness of Japanese semiconductors.

The biggest beneficiary of that void was Samsung Electronics. In 1987, when cumulative semiconductor losses reached 200 billion KRW, Samsung made a counter-intuitive investment by breaking ground on a 3rd-line factory, anticipating the PC boom the following year. While Japan hesitated, bound by regulations, Samsung reaped the fruits of the DRAM shortage and rose to be a major player.

In the 2000s, the chicken game was fought with money.

The expansion race led by Taiwanese firms in 2007 turned into sales below cost, and Qimonda of Germany, the world’s 5th largest DRAM maker, filed for bankruptcy in January 2009. Elpida of Japan, which combined the DRAM businesses of NEC and Hitachi, also collapsed in 2012. Samsung’s weapon was counter-cyclical investment—maintaining or increasing investment even when competitors cut production—and it pulled out the same formula during the 2022 memory bust, stating “there will be no artificial production cuts.” The three companies that survived this bloodletting—Samsung, SK Hynix, and Micron—created the current DRAM oligopoly.

When these histories are overlaid, 2025 becomes clear.

The 1980s was a game of geopolitics, and Samsung was the beneficiary.

The 2000s was a game of capital, and Samsung was the leader.

Now, both games are happening simultaneously. In HBM, it is a game of technology and capital; in foundry, it is a game of geopolitics with the U.S. government involved.

The 2nm foundry war and TSMC’s moat

As of 2025, the foundry market is the stage for a single company: TSMC.

According to TrendForce’s Q2 2025 data, TSMC’s market share is 70.2%, while Samsung Electronics’ is 7.3%, a gap that has widened to approximately 63 percentage points. According to Counterpoint Research, the figures are 71% and 8%, respectively.

Q2 2025 Foundry Market Share Rankings

Rank	Company	Q2 2025 Share	Q2 2024 Share	Note
1	TSMC (Taiwan)	71% (TrendForce 70.2%)	65%	All-time high due to AI chip demand
2	Samsung (Korea)	8% (TrendForce 7.3%)	10%	Share decline
3	SMIC (China)	5%	6%	Supported by Chinese gov
4	UMC (Taiwan)	5%	6%	-
5	GlobalFoundries (US)	4%	5%	-

Note: Based on Counterpoint Research; figures in parentheses are from TrendForce

TSMC’s moat is not about which nanometer process it has, but in the advanced packaging (CoWoS) that ties GPUs and HBM into a single unit.

An AI chip is not a single chip but a system where logic dies and HBM are intertwined. From NVIDIA or Apple’s perspective, going to TSMC means manufacturing and packaging can be completed at once. Even in the ‘Foundry 2.0’ market, which Counterpoint recalculated to include packaging, IDM, and mask companies, TSMC was 1st with 38%, while Samsung fell to 6th with 4%.

The next battleground is 2nm. TSMC is known to have secured Apple and Qualcomm volume as it enters 2nm mass production in the second half of 2025, but a major 2nm technology leak occurred in August of the same year. Samsung struggled with yield issues of the GAA structure it introduced early in 3nm, but it recorded its largest-ever quarterly orders in Q3 2025 and won a contract for Tesla’s next-gen AI chip at its Taylor plant in the U.S. Intel announced mass production of its 1.8A (18A) process in October 2025, highlighting its GAA structure, RibbonFET, and PowerVia, which delivers power from the back of the chip.

Why the CHIPS Act and tariffs make geopolitics the second front for semiconductors

Beyond technological competition, Korean and Taiwanese firms are standing on a second front: U.S. politics.

The U.S. attracted domestic production through CHIPS Act subsidies, and Samsung (Taylor), SK Hynix (Indiana), and TSMC (Arizona) have promised tens of trillions of KRW in U.S. investments.

The problem is that the conditions keep changing. Subsidies came with tricky requirements like excess profit sharing.

After the Trump administration took office in 2025, the repeal of the CHIPS Act was discussed, and in August, there were even remarks about imposing tariffs of up to 100% on imported semiconductors, with exceptions only for companies investing in the U.S. For companies that had already begun building factories, their irreversible investments were effectively turned into negotiation chips.

Meanwhile, Intel, a U.S. company, is preparing for a revival while receiving massive subsidies.

The structure of attracting investments from allied companies while growing domestic firms is reminiscent of the scene in the 1980s when the U.S. suppressed the semiconductor industry of its ally, Japan, through agreements. The trend of Korean semiconductor firms starting to take direct equity stakes in AI model companies was covered in an analysis of Samsung/SK Hynix’s Anthropic investment, and the data center power bottleneck beyond the chips was covered in [the Naver-NVIDIA AI Factory alliance article](https://phoue.co.kr/posts/ naver-nvidia-ai-factory-sovereign-ai-alliance-analysis/).

2026 Semiconductor Supercycle: Where do optimism and the ‘winter’ theory diverge?

The 2026 outlook diverges sharply depending on which market you look at. My assessment is that the top tier faces supply shortages, while the middle and lower tiers face a slowdown.

Representing the optimism is Samsung Electronics.

In its Q3 2025 earnings call, Samsung projected that even with increased facility investment and maximum production, customer demand in 2026 would exceed it. The grounds for this are observations that HBM4 prices could rise by more than 30% compared to HBM3E, the prospect of continued shortages in mobile, PC, and legacy products due to the tilting effect, and demand for high-capacity SSDs for AI.

On the other side is the ‘winter’ theory.

In September 2024, Morgan Stanley reversed its ‘AI supercycle’ outlook from three months prior and issued a report stating “Winter Looms.” The analysis was that inventory for general-purpose DRAM (excluding HBM) was at an all-time high of 62 weeks, and NAND at 67 weeks. The industry countered that this was preposterous, but the market was shaken. In October 2025, iM Securities also predicted that the DRAM/NAND market would enter a moderate slowdown in 2026 and that HBM growth would peak due to intensifying competition.

Both outlooks are looking at different layers.

Samsung’s optimism is about the top-tier market like HBM4 and server DRAM, while the winter theory is about mid-tier products like HBM3E and below, and general-purpose DRAM/NAND. As all three memory companies have embarked on HBM3E/HBM4 expansion, specific product groups like HBM3E may undergo price normalization. Ultimately, the battle will not be won by who makes HBM, but by who captures the highest floor—HBM4 and hybrid bonding—first.

As of September 2026, when I revised this article, winter has not yet arrived for general-purpose DRAM.

In November 2025, the fixed transaction price for PC DDR5 16Gb hit 19.5 USD, a 2.2-fold increase in one month, and the German retail DDR5 price index tracked by TechRadar has risen by 544% as of September 2026 compared to July 2025.

What will happen to DRAM prices? My personal hope is that they stabilize a bit, easing the economic burden on general consumers purchasing computers.