Taiwan gains a semiconductor premium as AI demand shifts from memory cycles toward durable bottlenecks in logic, packaging and HBM markets.
A forward price-to-earnings (P/E) multiple is a market expectation about future earnings written in shorthand. It says less about what a company earns today than about what investors believe those earnings are worth—and how long they can survive competition. Read that way, the most revealing number in Asian equities this autumn is a ratio. Taiwan's benchmark TAIEX (Taiwan Stock Exchange Capitalization Weighted Stock Index) trades at nearly 18 times its one-year forward earnings. South Korea's KOSPI (Korea Composite Stock Price Index) trades at about 5.5 times, according to market estimates reported on October 7, 2026. Both markets ride the same artificial intelligence (AI) hardware boom. The gap between them is a market judgment about which form of semiconductor scarcity will prove more durable.
That judgment is striking because Korea's earnings and export data have rarely looked stronger. On October 1, Korea's Ministry of Trade, Industry and Resources (MOTIR) reported September exports of US$120.94 billion, up 83.5 percent from a year earlier and the first month in which exports exceeded US$120 billion. Semiconductors alone reached US$60.30 billion, up 262.8 percent, and memory chips accounted for US$54.07 billion of that. Samsung Electronics subsequently estimated third-quarter operating profit at ₩107.4 trillion (about US$80.2 billion at the company's reported exchange-rate conversion), nearly nine times the ₩12.17 trillion reported a year earlier and slightly above the London Stock Exchange Group (LSEG) SmartEstimate of ₩106.1 trillion based on 21 analysts. Yet KOSPI fell 19.3 percent in the July–September quarter while TAIEX rose 3.9 percent. That 23.3-point gap is reportedly the widest since the turn of the century. By the close on October 7, TAIEX stood at 49,806.37, up 72.0 percent for the year, according to Taiwan Stock Exchange data. KOSPI stood at 6,803.90, up 61.5 percent, but 25.4 percent below its June 22 peak of 9,114.55.
Investors offer a ready explanation. M&G Investments portfolio manager Vikas Pershad put it plainly: Taiwan's earnings "come from volume," while Korea's, "in the near-term, are coming from price." A sharper version is that markets are repricing scarcity. Korea's core strength, memory, has historically been cyclical: capacity additions eventually discipline prices. Taiwan's strength sits in an ecosystem anchored by leading-edge logic, contract foundry and advanced packaging, where long qualification and limited capacity create longer-lived bottlenecks. The semiconductor cycle has not disappeared; the critical bottleneck has shifted, along with expectations about how long it will last. That hypothesis is attractive, which is exactly why it should be tested against three alternative explanations and valuation evidence before it is accepted.
The first alternative explanation is concentration: perhaps Taiwan's outperformance is just Taiwan Semiconductor Manufacturing Company (TSMC). The company carried a 41.12 percent weight in TAIEX on September 30, according to the Taiwan Futures Exchange. If TSMC were the story, the rest of the index would lag. It does not. TSMC shares rose from NT$1,550 (New Taiwan dollars) at the end of 2025 to NT$2,480 on September 30, a 60.0 percent gain, while the index rose 65.5 percent over the same period. Using the September 30 weight as a simplifying proxy, we at IndraStra Global estimate that TSMC's return would account for about two-fifths of the index's gain on a simplified basis, implying a substantially stronger gain for the remainder of the index. About 10 percent of TAIEX stocks had at least doubled this year, against 4.1 percent of KOSPI stocks. In Korea the pattern runs the other way. Samsung Electronics (common and preferred shares) and SK hynix together averaged 52.82 percent of KOSPI in September, according to the Korea Financial Investment Association (KOFIA). On those weights, the calculations suggest that Samsung Electronics and SK hynix together accounted for approximately four-fifths of the KOSPI's 62.3 percent year-to-date gain through September 30, leaving a substantially smaller gain for the remainder of the index. Because the calculation uses September weights rather than daily constituent weights and does not perform formal index attribution, these figures should be treated as directional proxies rather than precise contribution measures. The concentration story is more consequential in Korea than in Taiwan.
The second alternative explanation is that Taiwan earned more. It did not. Samsung reported second-quarter 2026 operating profit of ₩89.5 trillion, and SK hynix reported ₩60.54 trillion at a 76 percent operating margin. TSMC reported NT$766.6 billion at a 60.3 percent operating margin. Converted at the US Federal Reserve's H.10 second-quarter average rates (₩1,500.48 and NT$31.60 per US dollar), that is US$59.7 billion for Samsung, US$40.4 billion for SK hynix and US$24.3 billion for TSMC. Together, the two Korean companies increased quarterly operating profit by nearly US$90 billion from a year earlier, versus about US$9 billion for TSMC. The market’s skepticism is therefore not about whether Korea can generate extraordinary semiconductor profits, but about how much of those profits can survive when memory-price growth normalizes. That matters because the strongest cyclical pricing power in semiconductors currently sits in memory, particularly in AI-linked HBM and high-value DRAM.
The third test is valuation, and the evidence supports the hypothesis, albeit in a more nuanced form. If Korean profits are larger and growing faster, a 5.5-times multiple suggests that investors are heavily discounting their durability, rather than simply rewarding their current earnings. Earnings estimates for TAIEX rose about 19 percent in the September quarter against about 15 percent for KOSPI, and Taiwan's upgrades outpaced Korea's for the first time since March 2025. In the Bank of America fund-manager survey for September, 40 percent of respondents were overweight Taiwan against 25 percent for Korea. Asked which market would benefit most from the next phase of the AI cycle, 35 percent chose Taiwan and 5 percent chose Korea. Societe Generale strategists expect memory price appreciation to slow before normalizing in 2028 as Chinese competition rises. Part of the valuation gap is the long-standing "Korea discount." A portion of the remaining gap can therefore be interpreted as a durability premium, although the Korea discount also reflects longer-standing structural factors: markets pay 18 times for profits they expect to persist and 5.5 times for profits they regard as closer to a cyclical peak.
| Measure | Taiwan | South Korea |
|---|---|---|
| Year-to-date change through October 7, 2026 | TAIEX +72.0% | KOSPI +61.5% |
| July–September 2026 | +3.9% | −19.3% |
| Largest index weight(s) | TSMC 41.1% | Samsung + SK hynix 52.8% |
| Approximate contribution proxy using cited weights (to September 30) | About 39% | About 81% |
| Q2 2026 operating profit (US$) | TSMC US$24.3 billion | Samsung + SK hynix US$100.0 billion |
| Increase from Q2 2025 (US$) | +US$9.3 billion | +US$90.1 billion |
| One-year forward P/E (reported) | About 18× | About 5.5× |
Sources: Taiwan Stock Exchange, Korea Exchange closing data, Taiwan Futures Exchange, KOFIA, company filings, Federal Reserve H.10 (Q2 averages), reported market estimates; Contribution figures are IndraStra Global proxy estimates based on cited weights and constituent returns; they are not formal index-attribution calculations
The price data show why markets hesitate. TrendForce expects conventional DRAM (dynamic random-access memory) contract prices to rise 10 to 15 percent in the fourth quarter. NAND flash prices are expected to rise 15 to 20 percent, and the market remains undersupplied. That follows a second-quarter DRAM surge of about 60 percent. MOTIR's own fixed-price series shows a 16-gigabit DDR5 (fifth-generation double data rate) chip at US$48.00 in September, up from US$37.50 in May. That is deceleration, not collapse. In a highly cyclical market, however, investors often price the turn before it arrives, and memory makers know it. Long-term agreements with ceiling-price mechanisms, now spreading across the industry, trade some upside for predictability.
Why should Taiwan's supply advantages last longer? The answer lies in how an AI accelerator is built. A modern AI processor is not simply a single chip. It is a system: logic dies made on 3-nanometer or 2-nanometer processes, stacks of HBM (high-bandwidth memory) and a silicon interposer that joins them. For many leading AI accelerators, that integration is performed through TSMC's CoWoS (chip-on-wafer-on-substrate) packaging, which TSMC describes as integrating multiple system-on-chip (SoC) dies with high-bandwidth memory (HBM) stacks. Each added layer raises complexity. Complexity can lengthen qualification, because a customer must prove a design on a specific process and packaging flow before committing volume. Long qualification cycles can concentrate orders in suppliers that have already passed. Concentration during a demand surge creates a supply imbalance. A bottleneck then confers pricing power for as long as it takes a rival to qualify. Here, "bottleneck" means control over production capacity that is difficult to replicate, not a legal monopoly. TSMC held 73 percent of Counterpoint's pure-play foundry market in the second quarter, a tracker that counts Samsung's foundry business alongside dedicated foundries. In Counterpoint's broader "Foundry 2.0" measure, which adds non-memory chipmakers, outsourced assembly and test (OSAT) firms and photomask makers, its share was 42 percent. The two percentages are not directly comparable because Foundry 2.0 includes non-memory IDMs and OSAT revenue. TrendForce described 3-nanometer supply as "a temporary single-supplier dynamic dominated by TSMC." It also noted that the CoWoS shortage running since 2023 now extends to equipment, substrates and materials. TSMC is qualifying a 5.5-reticle version of its CoWoS-L packaging this year, and its SoIC (System on Integrated Chips) 3D stacking entered volume production at 3 nanometers in 2025.
| Supply-chain layer | Leading supplier(s) | Latest share or indicator | Diversification / erosion signal |
|---|---|---|---|
| Pure-play foundry (Q2 2026 revenue) | TSMC | 73%; Samsung Foundry 7% | Samsung 2nm wins; Rapidus 2nm plan |
| CoWoS-class 2.5D packaging | TSMC-led ecosystem; ASE and other OSAT expansion | Supply-demand gap about 20% in 2026, about 10% in 2027 | ASE, Amkor expansion; Intel EMIB-T (embedded multi-die interconnect bridge) |
| HBM (Q2 2026 revenue) | SK hynix, Samsung | 50%, 33%; Micron 18% | SK hynix down from 64% a year earlier |
| Conventional DRAM (Q2 2026 revenue) | Samsung, SK hynix, Micron | 38%, 25%, 24% | ChangXin Memory Technologies (CXMT) 10%, up from 4% |
| Extreme ultraviolet (EUV) lithography | ASML | Sole supplier | No rival EUV supplier |
| T-glass fiber cloth for substrates | Nittobo | About 90% | New capacity mid-2027 at the earliest |
Sources: Counterpoint (foundry, Foundry 2.0, DRAM and HBM trackers), ASML, TrendForce. HBM shares may not sum to 100 because of rounding.
The comfortable version of this argument stops there, casting Korea as the cyclical memory supplier and Taiwan as the structural chokepoint. HBM breaks that picture. HBM is itself a bottleneck, and Korea holds most of it. According to Counterpoint, SK hynix earned 50 percent of second-quarter HBM revenue and Samsung 33 percent, which means Korean suppliers accounted for roughly five-sixths of global HBM revenue. Korea therefore controls a major part of the AI supply chain. The East Asia Institute (EAI) identifies two bottlenecks in the AI supply chain: HBM and advanced foundries, one held mainly by each economy. What distinguishes them is not whether they are scarce but how quickly their ownership moves. The investment question is therefore not who owns a bottleneck today, but who can retain pricing power while the next supplier qualifies. SK hynix held 64 percent of HBM revenue a year earlier; Samsung rose from 15 percent to 33 percent as its products qualified. A market in which the leading supplier can lose 14 percentage points of share in a year behaves differently from one in which the leading supplier retains about 73 percent of pure-play foundry revenue across consecutive quarters. HBM's value chain also runs partly through Taiwan. SK hynix adopted TSMC's logic process for the base die of its sixth-generation HBM4 and agreed to optimize integration with CoWoS. Counterpoint observes that value is shifting toward logic base dies, bridges and substrates. The bullish counterargument deserves its weight too. Citigroup's Peter Lee argued in late September that markets underestimate how much HBM will be needed in 2027.
The strongest objection to this thesis is that bottlenecks erode—and some of Taiwan's current supply advantages are already easing. Counterpoint estimates TSMC's CoWoS supply-demand gap at about 20 percent in 2026, narrowing to roughly 10 percent in 2027 as new capacity comes online. TrendForce says TSMC plans to expand CoWoS capacity by more than 60 percent by 2027. ASE, the largest packaging-and-test supplier, raised its 2026 capital spending to US$10.5 billion, with 13 greenfield and 8 brownfield projects under way. Its chief operating officer, Tien Wu, said the company would welcome any alternative technology offering the same yield, given severe capacity constraints. Amkor is expanding 2.5D capacity, and Intel's EMIB-T bridge packaging has been adopted for Google's next tensor processing unit. Counterpoint adds that neither Intel nor Samsung "currently serves as a perfect alternative to TSMC's combination of advanced-node scale, yield and packaging execution." Japan has committed about US$15 billion in state support to Rapidus, which plans 2-nanometer production in the second half of its next fiscal year with 17 design partners. One prospective customer executive said that shifting work from TSMC to Rapidus would leave no easy way back if Rapidus failed: qualification is a moat in both directions. China's CXMT now holds 10 percent of DRAM revenue, up from 4 percent a year earlier. EAI concludes that both economies' bottleneck status is "most likely" to erode gradually over the next 12 to 18 months rather than collapse. It names silicon photonics and copper interconnects as the next constraints.
That objection does not defeat the thesis; it changes its form. Taiwan's advantage is better described as a dynamic bottleneck position than as permanent ownership of a single chokepoint: the repeated capture of the next constraint before the previous one fully commoditizes. TSMC's 3-nanometer process supplied 30 percent of wafer revenue in the second quarter, while 2-nanometer, at 3 percent, is just beginning. TrendForce expects CoWoS-L to remain the mainstream advanced-packaging solution through 2028 despite competition from Intel's EMIB-T, largely because of its technology maturity and yield. TSMC's SoIC stacking is next in line, which provides another layer of advanced integration beyond conventional 2.5D packaging. The constraint keeps moving, and so far the same supplier has kept arriving first. Even materials matter: TrendForce reports that Nittobo controls approximately 90 percent of the global T-glass market used in advanced substrates and is not expected to bring new capacity online until mid-2027 at the earliest. ASML describes EUV lithography as "unique to ASML." Korea has a route to the same premium. Samsung aims to place about two-thirds of its memory output under long-term contracts, and SK hynix has signed agreements with about ten customers. If those contracts turn peak prices into durable volume, the 5.5-times multiple will be hard to defend. Korea is not losing this cycle. It is being paid differently for winning it.
Geopolitics follows the same logic. What gives Taiwan strategic weight is not simply how many chips it makes, but how difficult its role is to replace. Replacing conventional memory capacity generally requires replicating large-scale fabrication for products that are more standardized than leading-edge logic and advanced packaging. HBM complicates that distinction because its value depends not only on memory fabrication but also on advanced stacking, packaging, qualification and integration with compute. Replacing a qualified leading-edge process and packaging flow means rebuilding an ecosystem of process knowledge, yield learning, interposers and customer validation. That is why diversification policies increasingly target the layers where the Taiwanese ecosystem is hardest to replace. Japan's subsidies to Rapidus, Intel's packaging push and the expansion of TSMC's overseas manufacturing footprint are different responses to the same problem: reducing the difficulty of replacing concentrated semiconductor capacity. But geographic diversification does not automatically replicate Taiwan's accumulated process knowledge, yield learning, supplier density and customer-qualification ecosystem. The same concentration that earns Taiwan its premium also invites the policies designed to dilute it.
The contest between Taiwan and Korea is no longer mainly about which semiconductor products they make. It is about which economy controls the bottleneck that cannot easily be substituted, and for how long. If the next binding constraint forms where qualification is slow and suppliers few, in packaging, base dies or optical interconnects, the premium is likely to stay in Taipei. If it forms in HBM, where Korean firms already account for roughly five-sixths of global revenue, Seoul's 5.5-times multiple will read less as a verdict than as a discount waiting to close.
Apple News, Google News, Feedly, Flipboard, and WhatsApp Channel
