Asian Semiconductor Sell-Off: Is the Market Overreacting?

• CXMT and China’s domestic DUV technology have made tangible progress, but neither is going to disrupt memory market or advanced-chip supply in the near term.
• TSMC and Samsung both beat expectations, while cloud revenue continued to grow strongly, suggesting that AI demand remains resilient.
• The recent correction appears more consistent with a market adjustment and deleveraging than a cyclical peak. We therefore maintain our positive view on the Asian semiconductor sector.

Asian equity markets have just been through a brutal correction. From its June high, the KOSPI fell as much as 34% at one point, and circuit breakers were triggered several times by single-day plunges. Samsung Electronics and SK Hynix both dropped more than 40% at one stage. Taiwan's market was not spared either.

Sentiment in China's stock market, by contrast, could hardly have been more different — as if it belonged to a parallel universe. ChangXin Memory Technologies (CXMT) listed on the STAR Market and closed its debut session up around 466%; at almost the same time, news broke that a domestically developed DUV lithography system had entered production.

The two headlines were quickly stitched together into a simple and frightening story: China now has both its own lithography machines and the world's fourth-largest DRAM maker. Is the next step a flood of cheap memory into the market, and then a breakthrough past the technology bottlenecks in AI chips and HBM?

Elsewhere, Google raised its capital expenditure guidance again and posted a rare negative free cash flow in the second quarter, prompting investors to ask whether AI spending on this scale can be sustained.

With all of this landing at once, have the fundamentals of Asian semiconductors genuinely turned — or is the market simply overreacting in a moment of panic?

A Decade of Backing from Hefei, but Still a Long Road Ahead

As we discussed in “CXMT, Zhipu AI and Unitree: Three Paths in the Rise of China's Hard Tech”, CXMT may be the most representative case of China's high-tech industry coming of age. But CXMT did not start from nothing — it had timing, location and people on its side.

In 2009, the German DRAM maker Qimonda went bankrupt, leaving behind roughly 7,000 patents, a memory cell architecture already proven in volume production, and around 2.8TB of technical documentation. That inheritance meant CXMT did not have to work out the basic architecture and process roadmap from scratch.

Technical documents, however, are not the same as the ability to manufacture at scale. Turning designs on paper into chips still requires experienced engineers. CXMT recruited Karl-Heinz Kuesters, who had spent 24 years at Siemens, Infineon and Qimonda, and hired extensively from Micron, Samsung, Taiwan and South Korea. Those engineers brought with them the hardest thing in the industry to replicate: hands-on experience of memory manufacturing at volume.

CXMT has also benefited from long-term support from the Hefei municipal government. Even though the company lost money throughout its first ten years, with accumulated losses reaching RMB36.65 billion at one point, funding never dried up. With strong government backing, packaging and testing houses, specialty gas suppliers, wafer reclaim services and equipment vendors gradually set up around the fab, forming a complete local industrial ecosystem. At the same time, China's own large memory demand gave CXMT a deep pool of potential customers.

A decade on, CXMT has grown from a locally sponsored industrial project into the world's fourth-largest DRAM maker — an unambiguous breakthrough. Look more closely, though, and the gap to the leading edge remains clear. CXMT's main products are still built on a roughly 1z-equivalent G4 node, broadly comparable with where the industry stood in 2017. Its shipment share was around 9% in 2025 and, even with aggressive capacity expansion, is expected to reach only about 12% by 2027.

The more important structural disadvantage lies in lithography. The three leading memory makers have already adopted EUV for their most critical mask layers. Without EUV, CXMT can only reach comparable line widths through DUV and complex multi-patterning — at the cost of higher expense and lower effective capacity. Its DDR5 still costs more than 30% more than that of the three leading memory makers.

That is the opposite of the market's assumption that “domestic means cheap”. Because its process density is lower, CXMT's DDR5 die is around 40% larger than Samsung's at the same specification, and costs more. In China, a 64GB DDR5-5600 module using CXMT chips sells for more than comparable Samsung or SK Hynix products. Add the fact that current output is not yet enough to cover domestic demand in full, and new supply is likely to be absorbed at home first — so a “flood of cheap DRAM into the global market” does not stand up for now.

Domestic DUV Reaches the Production Line — How Far Is a Breakthrough to EUV?

The DUV news shook the market no less than CXMT's listing. According to media reports citing people familiar with the matter, a state-backed equipment maker in Shanghai has begun small-scale production of a domestically developed immersion DUV lithography system, targeting around five units in 2026 and about 20 in 2027, with the first machines going to SMIC, Hua Hong and CXMT. Moving from prototype to small-scale production is undoubtedly an important breakthrough for China's self-sufficiency in semiconductor equipment.

Starting production, however, is not the same as being able to manufacture reliably at scale. A lithography system is not judged solely on whether it can complete an exposure; it must also meet demanding requirements on overlay accuracy, exposure throughput, defect rates, tool uptime, service turnaround and the stability of component supply. Reports also note that the domestic system still lags ASML clearly on overall performance and reliability, and that some critical components still come from Japan.

This is therefore an important milestone in China's push to control more of its own chip capacity, and it is worth watching closely — but it is far from a fatal blow to ASML or to the global semiconductor industry. More to the point, of the five conditions that allowed CXMT to catch up, only capital and domestic demand are clearly in place in lithography equipment; the rest have yet to be built.

Table 1: China's Memory Industry Compared with Its Lithography Industry

Condition for catching up	Memory (CXMT)	Domestic lithography	Assessment
Source of technology	Able to take over Qimonda's patents and technical documentation	There is no other bankrupt lithography giant from which a complete, production-proven technology base could be acquired	Clearly absent
Volume manufacturing talent	Able to recruit engineers on a large scale from Qimonda, Micron, Samsung and Taiwanese and Korean firms	Individual hires are possible, but replicating the engineering teams of ASML and its suppliers wholesale is far harder	Partly in place
Local ecosystem	Packaging and testing, specialty gases, wafer reclaim and equipment suppliers could cluster around the fab, gradually forming a complete industrial ecosystem	Core optics, lasers and precision metrology sit with highly specialised global suppliers such as Zeiss and Trumpf, and cannot be replicated quickly	Clearly absent
Patient capital	Long-term commitment from the Hefei government allowed the company to absorb years of losses	Central and local governments are equally willing to invest for the long term	In place
Demand	China has huge DRAM demand and anchor domestic customers	Export controls reinforce the case for domestic substitution; fabs have both policy and supply-security incentives to adopt	In place, if anything stronger
Source: iFAST compilations.

China's semiconductor ecosystem today is certainly more mature than it was in 2016. But in lithography there is no second Qimonda from which China can inherit a complete, production-proven technology base, and the core optics, light sources, precision mechanics and metrology systems are spread across a number of highly specialised global suppliers. Even if immersion DUV is successfully localised, that is not the same as breaking through to EUV — still less being able to manufacture the most advanced AI chips competitively at scale. It took CXMT close to ten years to enter the commodity DRAM market; catching up in lithography equipment will only be harder.

Does Being Able to Make DRAM Mean CXMT Has Already Broken into HBM?

HBM is not simply a matter of stacking layers of DRAM: it also involves through-silicon vias, thermal management, advanced packaging and controller integration — all difficult processes — as well as co-design and qualification with AI chip customers. A problem at any one step can drag down the yield of the entire stack, which makes HBM far harder to manufacture than conventional DRAM. As a result, HBM is no longer a fully standardised memory product; its prices and profits will still fluctuate, but the cycle should be far less pronounced than in traditional memory.

The three memory giants understand HBM's strategic importance, and have therefore directed capital expenditure and capacity toward it first. That has tightened supply of conventional DDR and LPDDR. In conventional memory, then, the AI boom has actually opened a window for CXMT. But treating it as a near-term shock to HBM supply would overstate the real impact.

As of the end of 2025, CXMT's overall HBM yield is estimated at around 25%, with related wafer capacity of roughly 5,000 wafers per month — less than 2% of its total. Around 99% of its revenue that year still came from conventional memory products. The proceeds of its listing are earmarked mainly for wafer production lines, DRAM process upgrades and forward-looking R&D, with no separate large-scale HBM expansion programme. CXMT may gradually work its way into China's domestic HBM ecosystem, but in the near term it is not enough to move global AI memory supply and demand.

Back to the Market

Unpacking the China supply story alone, however, does not explain the size of this sell-off. The market's other worry is whether AI capital expenditure has gone too far.

On 22 July, Google raised its full-year capital expenditure guidance from US$180–190 billion to US$195–205 billion; its shares fell about 7% the following day, reflecting investors' doubts about whether spending on that scale can keep generating adequate returns. Yet in the same quarter Google Cloud revenue grew 81.8% year on year to US$24.768 billion, the third consecutive quarter of acceleration; the operating margin rose from 20.7% to 35.6%; backlog reached US$513.9 billion; and management said supply remains constrained. On a year-on-year basis, roughly 54 cents of every additional dollar of cloud revenue converted into operating profit. Judged on revenue growth, incremental margins and order visibility at least, there is still no evidence that returns on AI infrastructure are deteriorating quickly.

Amazon's results pointed the same way, with AI demand still strong. Cloud revenue grew 37% year on year, the fastest in 18 quarters, and backlog rose from US$364 billion in the previous quarter to US$496 billion. Management added that even if full-year capital expenditure were pushed to around US$220 billion, the incremental compute available in 2026 would still fall short of customer demand.

The supply chain shows no sign of weakening demand either. TSMC, on the back of strong results and confidence in the outlook, raised its full-year revenue growth guidance to slightly above 40% and lifted capital expenditure for the second time this year. Management has gone as far as checking AI data-centre construction and rack installations one by one, to make sure the chips it produces really are being deployed and generating revenue.

In memory, TrendForce expects the quarter-on-quarter increase in conventional DRAM contract prices to narrow from 90–95% in the first quarter and 58–63% in the second to 13–18% in the third. A slower rate of increase, however, does not mean demand has peaked. The main reasons are:

1.     After two consecutive quarters of steep increases, prices are already on a high base, making it naturally harder to sustain the same rate of quarterly gains.

2.     Consumers' tolerance for high prices is approaching its limit, and some of the cost is starting to show up as lower shipments of smartphones and other consumer electronics, which depresses the related memory demand in the near term.

3.     Large cloud customers have generally signed long-term supply agreements, which limits suppliers' room to raise prices.

In other words, what we are seeing looks more like a slower rate of price increases than a turn to falling prices — and it does not mean that memory demand or industry profits have peaked. According to SemiAnalysis, even counting all the new wafer capacity already announced by CXMT, Samsung, SK Hynix and Micron, its model still points to a supply shortfall in the high single digits in percentage terms in 2026, potentially widening to the low-to-mid double digits in 2027. Management teams and industry forecasts also generally expect the memory shortage to last at least until 2028.

Does SK Hynix's Miss Mean Asian Semiconductors Are Weakening?

On 29 July, SK Hynix reported its second-quarter 2026 results. Although both revenue and operating profit set quarterly records, they still came in below market expectations, and the market quickly read the shortfall as a signal that the AI memory cycle had peaked. The shares fell sharply.

On the miss itself, analysts have generally attributed the shortfall to slower revenue recognition on HBM4, product mix and long-term agreement pricing, while the company stressed that long-term contract demand and the HBM ramp remain solid. On that reading, one quarter below consensus does not necessarily mean the fundamentals have weakened. SK Hynix has now signed long-term supply agreements with around ten major customers, HBM4 has entered volume production and shipment, and the company plans to expand output in the second half. On order visibility, management's outlook, and its technology and competitive position in HBM, there is no clear sign of deterioration.

Samsung's results offer a useful contrast. Second-quarter revenue was around KRW171.5 trillion and operating profit around KRW89.5 trillion, both all-time highs, and in line with or slightly ahead of market expectations — yet the share price reaction was relatively muted. The semiconductor division accounted for almost all of the profit, while the device business was held back by rising component costs. Memory sales grew strongly quarter on quarter, management expects tight supply to last at least until 2028, and HBM4 shipments are set to accelerate in the second half.

Samsung has signed long-term supply agreements with five of the world's largest AI and data-centre customers, with around five more still in discussion, and expects long-term contracts ultimately to cover roughly 60% to 70% of its DRAM and NAND capacity; some agreements also include prepayments and price floors. Like SK Hynix, Samsung is gradually shifting part of its memory business away from spot and short-term contracts toward a model that locks in demand, pricing and cash flow through long-term agreements.

Put the two companies side by side — one posting record results that nonetheless fell short of an elevated consensus, the other posting equally record results that met or slightly beat expectations — and this looks more like the market recalibrating valuations and expectations than a broad weakening of memory demand.

More importantly, SK Hynix is only one part of the Asian semiconductor industry and of the related indices. For investors worried about single-stock risk, we think an ETF is a sensible way to gain exposure to the Asian semiconductor sector, reducing the impact of any single company event on the overall portfolio.

Take the FactSet Asia Semiconductor Index, which the Global X Asia Semiconductor ETF tracks. TSMC, Samsung Electronics, SK Hynix, MediaTek and Sony each have different revenue sources, end markets and investment cases. With TSMC's and Samsung's results and outlook still solid, and the core investment assumptions for the other major constituents intact, one quarterly miss at SK Hynix is not enough to conclude that the fundamentals of the entire Asian semiconductor index have turned.

Table 2: Top Five Constituents of the FactSet Asia Semiconductor Index

Constituent	Weight	Results vs. consensus	Guidance and demand	Investment case
TSMC	17.62%	Revenue at the top of guidance; gross margin above the top end	Raised full-year revenue growth and capital expenditure; advanced packaging and testing still constrained	AI and advanced-node demand has not weakened; case strengthened
Samsung Electronics	12.86%	Record revenue and operating profit; in line with or slightly ahead of expectations	LTAs signed with five major AI customers, five more in discussion; HBM4 shipments accelerating in H2	LTAs expected to cover 60% to 70% of capacity, improving visibility; case strengthened
SK Hynix	12.80%	Record revenue and operating profit, but below an elevated consensus	Around ten customers have signed LTAs; HBM4 capacity expanding in H2	One quarter below expectations; competitive position in HBM not reversed
MediaTek	5.84%	June revenue NT$58.01bn; Q2 NT$152.18bn, above top of guidance	AI ASIC volume production expected in Q4	Google TPU/AI ASIC path intact
Sony	5.59%	Imaging and sensing (I&SS) operating profit up 37% year on year last quarter, a segment record	I&SS guidance for this fiscal year is conservative and the shift to larger smartphone sensors is slowing; signed an MOU with TSMC to form an image sensor joint venture	Not a pure AI memory exposure; sensor demand tracks the smartphone cycle
Source: company announcements, Global X and iFAST compilations. Data as of 30 July 2026.

Investment Implications: Perfection Was Priced In, but the Fundamentals Have Not Weakened

In our view, although SK Hynix's second-quarter results came in slightly below market expectations, both TSMC and Samsung met or beat consensus, and we see no sign yet that the fundamentals of the Asian semiconductor sector are weakening. China's CXMT is still focused mainly on commodity DRAM, and domestic DUV has only just moved from prototype to small-scale production, leaving it some way from qualification at the major fabs and from stable volume manufacturing. At the same time, the revenue, earnings and order visibility of Google Cloud, Amazon, TSMC and the Korean memory makers continue to show that demand for computing power is very strong.

This correction therefore looks more like the market repricing an over-optimistic “perfect narrative”, together with concentrated deleveraging amid panic, than the end of the AI investment cycle. At this stage there is still not enough evidence that the earnings cycle for Asian semiconductors has peaked, and we maintain our positive view on the sector.

For investors who are positive on AI computing power, advanced process nodes and long-term memory demand, this sharp sell-off offers an attractive entry point. We maintain our target price of HK$334; based on the closing price on 30 July, that implies potential upside of around 130%.

Chart 1: After the correction, the forward P/E of the FactSet Asia Semiconductor Index has fallen below one standard deviation from its historical mean, leaving valuations highly attractive

Table 3: FactSet Asia Semiconductor Index Earnings Forecast and Potential Upside

FactSet Asia Semiconductor Index	2025A	2026E	2027E	2028E
EPS (HKD)	19.3	51.7	70.6	79.7
Earnings growth (YoY)		168.5%	36.4%	12.9%
Target price (based on a fair P/E of 20x)				1,594
Potential upside (%)				125%
Source: Bloomberg and iFAST compilations. Data as of 30 July 2026.