Taiwan Outlook 4Q26 | From chips to memory, Taiwan can’t stop winning

Taiwan’s exports reached a record US$82.4bn in August, up 41% year-on-year. Export orders, a leading indicator, exceeded US$100bn in a single month for the first time in August, rising 71.4% year-on-year and extending their growth streak to nineteen consecutive months. Despite a higher comparison base, growth momentum remains strong.

Figure 1: Taiwan’s export orders have grown for nineteen consecutive months

A few months ago, the market was still questioning whether the huge capital spending of the large cloud companies could earn a reasonable return. Yet revenue growth accelerated at all three major cloud providers in the second quarter and operating margins stayed high, with Google Cloud’s margin doubling in six quarters, gradually easing concerns about investment returns.

Figure 2: All three major cloud providers grew revenue faster in the second quarter than a year earlier

Figure 3: Operating margins at the three major cloud providers remain high

An interview with Google Cloud chief executive Thomas Kurian in September offered a glimpse of the returns on AI investment. He said: “Our payback period on AI servers in aggregate is less than two years, and on our own silicon is half that.” With demand this strong, cloud companies can recover their hardware costs within a short period, showing that AI infrastructure is already delivering clear commercial returns and giving them reason to keep expanding investment.

Beyond the three major cloud providers, growth at the newer specialist cloud providers is equally strong, reflecting continued tightness in AI compute supply. Nebius’s chief executive said in August that “we could sell today our entire 2027 capacity”; its first capacity auction cleared 15% above the previous highest rate, and most new contracts included prepayments, showing that customers are paying higher prices and committing funds early to secure limited AI compute.

Sustained strong demand is driving cloud companies to invest more. Amazon and Google raised their 2026 capital expenditure guidance after second-quarter results, taking combined guidance for the four largest technology companies to US$720–745bn, around 80% higher than last year; 2027 spending by the five largest hyperscalers is expected to reach US$1.3tn. Including chips, servers, networking, land, power and buildings, cumulative global infrastructure investment between 2024 and 2029 is projected to exceed US$10tn.

Figure 4: Global AI infrastructure investment is projected to keep growing

Strong demand for AI hardware continues to spread up the supply chain. Nvidia’s first full-year outlook, issued in August, already implied growth of around 70%, but chief executive Jensen Huang said demand growth was well above 70% and that the growth the company is confident of delivering depends on the supply it can ship; its chief financial officer described it as “a supply-constrained outlook”. In other words, the main constraint on AI hardware shipments is still supply, not a lack of demand.

The GPU rental market shows the same trend. Rental prices for current and previous-generation Nvidia GPUs have held firm: H100 and B200 rental prices are 20% and 33% above their January lows. CoreWeave said that A100 GPUs, launched in 2020, are still being contracted “at full freight” on long-term agreements running through 2029, and Nvidia’s chief financial officer said these A100s are “still running at full utilization today”. Even as GPU generations keep turning over, the market is still using large amounts of five- and six-year-old hardware, a sign of how strong hardware demand is.

Figure 5: Nvidia GPU rental prices have held firm

Customers add orders as TSMC accelerates expansion

TSMC, the largest constituent of the TWSE Index, reported August revenue above US$16bn for the first time, up 44% year-on-year in US dollar terms. July and August together already equalled 92% of revenue for the whole of last year’s third quarter; at the current run-rate, there is a good chance that third-quarter revenue will again beat company guidance.

In Taiwan Outlook 2H26, we argued that as the economic value created by advanced processes rises in the AI era, TSMC is in a position to raise prices gradually. TSMC reportedly told major customers in July that advanced-process prices will rise by 5–10% across the board from 2027; including surcharges related to high-performance computing, the total increase on some orders could exceed 10%.

Even with higher prices, customer demand keeps growing. Apple, Nvidia, AMD, Qualcomm and MediaTek have reportedly added to their 2nm capacity reservations, prompting TSMC to accelerate its expansion: monthly capacity could reach 120,000 wafers by the end of this year, a level the market had originally expected only in 2027. As detailed in TSMC 2Q26, the company plans to build 13 wafer and advanced packaging facilities in Taiwan over the next several years. By mid-year, equipment orders were nearly double the plan set last December, showing that management is sharply increasing investment to build capacity ahead of demand. With customers continuing to add orders, TSMC may need to accelerate investment further, even though its expansion has already stepped up significantly.

Memory brings new demand for logic chips

In Memory Supercycle: A Structural Shift Rewriting Supply and Demand Paradigm, we noted that as HBM bandwidth and power requirements rise, base dies are gradually moving from conventional memory processes to advanced logic processes to improve signal transmission, energy efficiency and overall system performance. SK hynix’s HBM4 base die uses TSMC’s N12 process, Micron has also given its HBM4E base die to TSMC to manufacture, and TSMC supports custom HBM4E designs on N3P. This means HBM is becoming a new source of AI foundry demand alongside GPUs, ASICs and advanced packaging; and as the process moves up from N12 to N3P, the foundry value of each base die is likely to rise as well.

At the same time, HBM is becoming the main bottleneck to expanding AI accelerator output, and tight supply has even begun to affect GPU product design. Nvidia, for example, cut the HBM capacity per GPU of its next-generation Rubin Ultra from 288GB to 192GB to raise the number of units it can ship. That even a company with Nvidia’s grip on its supply chain has had to change its hardware configuration because of limited memory supply shows how severe the shortage is.

For TSMC, lower HBM content per GPU means the same memory supply can support more GPU shipments, increasing demand for advanced-process wafers and CoWoS packaging.

Packaging alternatives broaden opportunities across Taiwan’s supply chain

CoWoS capacity remains tight, with lead times reported at around 78 weeks; as customers adopt more packaging technologies, other Taiwanese suppliers continue to participate. Part of Google’s next-generation TPU is reportedly moving to Intel’s EMIB-T, while AMD has developed a silicon bridge solution with ASE. The reported EMIB-T supply chain includes silicon capacitors from Powerchip, silicon bridges from UMC and substrates from Unimicron; TSMC is also developing a silicon bridge package with Kinsus. The relevant accelerator chips still rely on TSMC’s advanced processes. Alternative packaging may reduce TSMC’s share of that manufacturing step while creating business for other Taiwanese suppliers.

Taiwan’s participation also extends to chip design and system assembly. MediaTek’s role in Google’s TPU programme provides another route into custom AI chips; production of its first TPU ASIC is scheduled to begin in the fourth quarter, and Nvidia has agreed to buy US$3.5bn of its convertible bonds. Foxconn expects AI rack shipments to double this year, with a global market share of around 40%, and says its cloud and networking business, led by AI servers, now accounts for 51% of revenue, more than iPhone-related revenue. Its chief executive said chip supply will limit shipments of next-generation racks in 2027.

Earnings upgrades have outpaced the market

Since our June upgrade, the TWSE Index has risen only about 4%, while consensus earnings forecasts have been revised up 20%, bringing the forward price-to-earnings ratio down from 20.6 times to 17.7 times, around one standard deviation above its five-year average. Taiwan’s earnings base has improved markedly over the past three months, yet the valuation multiple has contracted: the improvement in fundamentals is not yet fully reflected in share prices.

Figure 6: Earnings have risen far faster than the index, bringing the forward P/E back to around one standard deviation above its average.

We acknowledge that the global semiconductor sector may remain volatile in the short term. During the third quarter, the TWSE Index recorded both its largest one-day points fall and its largest one-day points gain on record. However, we see this kind of volatility as largely noise, unlikely to change the long-term direction of AI infrastructure investment.

We see the current risks mainly in short-term sentiment and valuation swings, not in a reversal of AI demand itself. We therefore maintain our 4-star “Very Attractive” rating on Taiwan. At a fair price-to-earnings ratio of 20 times, the Taiwan market has potential upside of around 55%, and the risk-reward is more attractive than when we upgraded it in June.

Table 1: Projections for the TWSE Index and Upside Potential

TWSE Index	2025A	2026E	2027E	2028E
EPS (TWD)	1,494.5	2,308.9	3,058.7	3,732.0
Earnings Growth (%)		54.5%	32.5%	22.0%
PE Ratio (x)		20.9	15.7	12.9
Target Price (based on fair PE of 20x)				74,640
Upside Potential by end-2028 (%)				55%
Source: Bloomberg and iFAST Compilations Data as of 23 September 2026

Related article: Korea and Taiwan arrive on SGX: the first ETFs for both ends of the AI supply chain

Related article: Fund Spotlight: Capturing Taiwan’s AI opportunities beyond TSMC

Related products:

Market / Sector	Fund	ETF
Taiwan	LionGlobal Taiwan Equity Fund	IFAST-Xtrackers MSCI Taiwan ETF (SGX: TWD) (SGX-listed, in SGD) The Initial Offer Period (IOP) for the iFAST Xtrackers MSCI Taiwan UCITS Index ETF runs from October 1 to October 14, 2026, with each share priced at S$1.
Asian Semiconductors		Global X Asia Semiconductor ETF (HKEX: 3119)