The global semiconductor sell-off is not simply another correction in an expensive sector. It is the moment when two assumptions that sustained the artificial-intelligence boom have begun to weaken simultaneously. The first was that demand for advanced chips, memory and computing infrastructure would expand almost without limit. The second was that China would remain structurally dependent on Western and Asian technology for the equipment required to manufacture them. Both assumptions are now being questioned. The immediate market reaction has been brutal. Asian semiconductor shares suffered their sharpest decline since March, South Korea’s Kospi fell by more than 10% at one stage and trading was temporarily suspended. Samsung Electronics and SK Hynix each lost more than 12%, while technology-heavy indices in Japan and Taiwan fell by more than 4%. The correction spread rapidly through the entire semiconductor ecosystem. Memory manufacturers fell. Equipment suppliers fell. Lithography companies fell. The shares of several Chinese equipment producers rose.

The catalyst was a report that a Chinese state-backed company had started mass production of deep-ultraviolet immersion lithography machines. These tools are less advanced than the extreme-ultraviolet systems used to manufacture the most sophisticated chips, but they remain essential to large parts of semiconductor production. The reported output is initially modest, with only a handful of machines expected this year and perhaps twenty next year. Chinese equipment remains years away from reproducing the full performance, reliability, and industrial-scale capabilities of leading Western systems. That should have limited the market reaction. It did not.

ASML fell sharply. Nikon and Tokyo Electron declined by more than 9%. Applied Materials, Lam Research, ASM International and BE Semiconductor Industries also suffered heavy losses. The reason lies less in the immediate commercial threat than in the strategic implication. Markets had assumed that export controls would preserve a durable technological gap between China and the West. China could manufacture less advanced semiconductors, adapt older equipment and increase production through less efficient processes, but it would remain dependent on foreign tools for the most critical stages of chipmaking. The latest report suggests that dependence may be declining faster than expected. China does not need to surpass ASML immediately to alter the industry’s economics. It merely needs to become good enough. A domestic DUV machine with lower productivity, lower yields, or lower precision may still be strategically valuable if it allows Chinese manufacturers to expand capacity without relying on restricted foreign imports. Beijing is not seeking the highest short-term return on capital. It is seeking technological sovereignty. That distinction is fundamental.

Western semiconductor companies operate according to commercial logic. China increasingly operates according to strategic logic. A Western company abandons a project when expected returns become insufficient. The Chinese state may continue to finance it because national security, industrial independence, and geopolitical resilience matter more than immediate profitability. This asymmetry gives China time. It can accept lower yields. It can tolerate duplication. It can subsidise inefficient production. It can build capacity before demand fully exists. It can absorb losses that private shareholders would reject. For years, investors interpreted these characteristics as evidence of poor capital allocation. Increasingly, they must also recognise them as instruments of industrial power.

The history of Chinese manufacturing offers a clear warning. China did not initially dominate solar panels because its technology was superior in every respect. It dominated by building capacity, reducing costs and forcing competitors to operate against a state-supported industrial system. The same pattern appeared in batteries, electric vehicles, telecommunications equipment and critical minerals. Semiconductors are vastly more complex, but the strategic method is familiar. Subsidise domestic production. Import knowledge. Recruit experienced engineers. Replicate foreign equipment. Create a protected home market. Expand capacity. Reduce prices. Make foreign dependence politically unacceptable.

The market is no longer valuing semiconductor scarcity alone. It is beginning to price the possibility of future overcapacity. That change is especially dangerous for memory chips. Memory has always been cyclical. Prices rise when demand accelerates, and supply remains constrained. High prices generate extraordinary margins, encouraging producers to invest. New capacity eventually arrives, supply catches up, and prices collapse. Artificial intelligence appeared to have interrupted this familiar cycle. High-bandwidth memory became essential for advanced AI processors. SK Hynix established an early lead, allowing it to capture exceptional pricing and profit growth. Its shares surged, and the company became one of the purest expressions of the AI infrastructure boom. Yet the logic that made the company attractive also made the position increasingly crowded. Its valuation may have appeared modest relative to earnings, but those earnings reflected extraordinarily favourable pricing. When investors capitalise peak margins as though they were permanent, apparently inexpensive cyclical companies become expensive in disguise. SK Hynix now embodies this contradiction. The company is expected to report another quarter of record profitability, with sales potentially more than tripling from a year earlier and operating profit increasing severalfold. Yet its shares have fallen sharply from their June peak, erasing an extraordinary amount of market value. The market is no longer asking whether the next quarter will be strong. It asks whether current profits reflect a sustainable new regime or the most profitable stage of another semiconductor cycle. This is a much more difficult question.

The same phenomenon is beginning to appear across the wider AI complex. Investors originally treated artificial intelligence as a software revolution with limited marginal costs and almost unlimited scalability. The reality has become much more capital-intensive. Data centres require land, electricity, cooling, transmission infrastructure, servers and advanced chips. The largest cloud companies are committing trillions of dollars to the build-out, supported by corporate bonds, private credit and increasingly complex project-finance structures. The five leading hyperscalers are expected to require an extraordinary amount of AI-related investment by the end of the decade, forcing them to use multiple funding channels as conventional bond markets gradually approach saturation. Nvidia’s latest infrastructure agreements, reportedly totalling more than USD 750 billion, have intensified concerns that the scale of financing is beginning to outstrip the visibility of future returns. The question is no longer whether AI will matter. It undoubtedly will. The question is whether every dollar being spent today will generate an adequate return. Markets often confuse technological importance with investment profitability. A technology can be revolutionary and still produce poor returns for investors who pay too much, finance too much capacity or choose the wrong part of the value chain. Artificial intelligence will not escape this rule.

The current semiconductor sell-off suggests that investors are beginning to understand the distinction. This is a profound change in market psychology. The Chinese breakthrough in lithography, assuming it proves commercially viable, also exposes a deeper contradiction in Western export policy. Controls were designed to slow Chinese technological progress by restricting access to advanced equipment. In the short term, they succeeded. But restrictions also transformed technological independence from an economic ambition into a national-security imperative. Before export controls, Chinese companies could buy advanced foreign tools. After export controls, they had no strategic alternative but to build their own. This is the paradox of technological containment. It can slow progress. It can also accelerate determination. This does not mean ASML’s dominance is about to disappear. Extreme-ultraviolet lithography remains one of the most complex engineering achievements in modern industry. Its ecosystem includes optics, lasers, precision mechanics, software and thousands of specialised suppliers accumulated over decades. China cannot reproduce that system quickly. However, ASML’s valuation reflected not only technological excellence but also the expectation of durable scarcity. Any evidence that scarcity may weaken deserves a lower multiple, even when the immediate earnings impact remains limited. This principle applies across the sector.

Strong earnings will confirm what markets already know: current demand remains robust and recent memory pricing has been exceptionally favourable. They will not settle the central debate. The more important questions concern future expenditure, customer behaviour and capacity. Will hyperscalers continue raising their AI budgets at the same pace? Will memory prices begin to constrain deployment? Will Chinese suppliers qualify for more global customers? Will semiconductor equipment orders remain strong despite geopolitical restrictions? Will returns on AI investment become visible before capital requirements accelerate again? These are questions that one quarter cannot answer. The market’s unusually high threshold for positive surprises reflects this uncertainty. Companies may beat earnings estimates and still fall because expectations are no longer contained in published forecasts.

The semiconductor industry has crossed an important psychological threshold. Investors no longer believe that more AI spending is always better. They no longer assume that technological leadership is permanent. They no longer treat China as a technologically static competitor. They no longer believe that record earnings guarantee rising share prices. The artificial-intelligence revolution remains real. The investment cycle supporting it remains enormous. But the market has begun to distinguish between the importance of the technology and the profitability of financing it. That distinction is where every boom ultimately meets reality. For two years, semiconductor investors feared missing the future. They are now beginning to fear that they have already paid for too much of it.