Asian exports are being rearranged according to their position in the AI industry chain

This article may contain an automated translation. The original source is linked for verification.

Asian exports are being rearranged according to their position in the AI industry chain

AGI-Signal

2026.11.11 09:23

· From Beijing

Full text 5347 words

Turning to the latest monthly trade data in Asia (South Korea to October, Taiwan to October), two sets of records were set in the two months before and after. South Korea's exports in October were US$115.28 billion, a year-on-year increase of 93.7%, the largest monthly increase since January 1974, slightly lower than the record of US$121.1 billion set in September. From January to October, cumulative exports reached US$930 billion, exceeding US$900 billion for the first time in history. The industrial sector predicts that the annual export volume is expected to reach US$1 trillion within November.

Taiwan's exports in October were US$96.76 billion, exceeding US$90 billion for the first time in a single month, an annual increase of 58.9%, the 36th consecutive month of growth, and once again setting a new record for the longest continuous growth in history.

Comprehensive official trade data from South Korea, Taiwan, and Japan showed that South Korea's exports in the first half of the year were US$496.3 billion, Taiwan's exports were US$416.6 billion, and Japan's US$384.4 billion. For the first time, the exports of South Korea and Taiwan exceeded Japan's. Taiwan's financial authorities estimate that its annual exports will surpass Japan's for the first time, and its ranking in global export statistics is expected to rise to eighth. Japan has long been at the forefront of Asia's exports and is the benchmark for high-end manufacturing exports. Being surpassed is not only related to rankings, but also reflects changes in the regional division of labor structure: the source of export competitiveness is shifting from overall manufacturing capacity to its link in the AI supply chain.

The two sets of data have one thing in common. It is not the traditional main force such as automobiles and machinery that drives growth, but chips. The final demand side is a handful of cloud computing companies on the other side of the Pacific. But it can be seen from the data that prosperity is real and dependence is real.

Record-breaking numbers, let’s look at the caliber first

The number itself is not controversial. South Korea's semiconductor exports in October were US$56.44 billion, a year-on-year increase of 259.4%. At least three amplifying factors are superimposed: memory prices are at a high level; the Mid-Autumn Festival holiday fell in September this year and in October last year. The base number in October last year was therefore low, and the number of working days in October this year also increased accordingly. According to the Ministry of Industry, Trade and Energy, the average daily export growth rate after adjusting for working days was 76.1%, which was significantly lower than the nominal value.

The situation in Taiwan is similar. The problem of low base numbers also exists, and there is a clear cause. Taking September as an example, Meekita Gupta, an Asian economist at the research institution Pantheon Macroeconomics, explained in her comments that the market had generally ignored the low base effect. Server-related shipments fell sharply in September 2025, raising this year's year-on-year reading. Her team's prior forecast for Taiwan's export growth in September was only 34%, which was more than half of the actual value of 60.9%.

After excluding base and price factors, the trend growth rate of chip exports is roughly estimated to be between 30% and 40%. This is still strong growth, and it reflects the actual level of demand, but 259% includes the amplification effect of statistical caliber. If you extrapolate 259% to make production capacity decisions, there is a high probability that you will face a surplus in one year. Judging the prosperity based on 30% to 40% may underestimate the heat of the inventory replenishment cycle and the intergenerational switching of technology.

Price factors are also at play. This round of surge in chip exports not only comes from the expansion of volume, but also includes the contribution of price. Storage manufacturers have simultaneously raised prices on both HBM and conventional DRAM. TrendForce has revised up the annual growth forecast for HBM hybrid average selling price in 2027 to 121%, and expects DRAM contract prices to increase by about 20% in the fourth quarter of 2026.

Data for October released by the General Administration of Customs of China on November 10 showed that integrated circuit exports that month were US$39.59 billion, a year-on-year increase of 137.1%, but the export volume fell by 2.5% year-on-year. The increase in amount was almost entirely due to price and product structure.

More telling is the structure. Semiconductors accounted for 49% of South Korea's exports in October, compared with 26.4% a year ago; semiconductors contributed about 73% of the US$55.8 billion year-on-year increase in exports. The total growth of all other categories is 34%, which is not low, but not at the same level as chips. This is not a full-blown export recovery, with growth concentrated in one category: semiconductors, while the rest of the economy is flat.

How orders become customs data

Exports are highly concentrated, and their roots lie in the supply chain itself. An order for an AI accelerator started from the budget meeting of the California board of directors: Most of the advanced logic chips are manufactured in Taiwan, which accounts for more than 90% of the world's most advanced manufacturing processes; the chips need to be paired with high-bandwidth memory (HBM), and HBM is currently produced by only three companies in the world: SK Hynix, Samsung and Micron. The first two together account for more than 80% of the share (in terms of revenue); the finished accelerators are assembled into servers in Taiwan and Vietnam and sent to data centers in the United States.

Almost every dollar of accelerator demand translates into exports of Korean memory, Taiwanese wafers and Southeast Asian assembly. On the demand side, the total capital expenditure guidance of Microsoft, Google, Amazon, and Meta in 2026 is approximately US$730 billion (see the figure below, and the calculation caliber is shown in the figure note). The third quarterly report at the end of October was the latest confirmation. The total cash capital expenditure in a single quarter is approximately US$203 billion, which has exceeded the operating cash flow of approximately US$175.5 billion in the same period. Three of them have turned negative free cash flow.

Amazon said expenditures will increase significantly in 2027. Google maintained its full-year guidance of US$195 billion to US$205 billion and expected significant growth in 2027. Microsoft cited delays in third-party production capacity delivery and slightly lowered its forecast for the 2026 calendar year to US$170 billion, while forecasting an investment of more than US$50 billion in the next quarter. The revenue growth rate of the cloud business is also rising, with Azure growing by 48%, AWS growing by 39%, the fastest since 2021, and Google Cloud growing by 103%.

This money will not be invested evenly in global manufacturing, but will only flow to a small number of factories that can produce bottleneck components. A quarterly decision in California translates into trade data in Busan or export statistics at the Port of Kaohsiung in a quarter.

Scarcity is well documented at every link in this transmission chain. HBM's mass production certification cycle is measured in years. TSMC's advanced process production capacity will be fully booked in the next two to three years. Vietnam's assembly capacity is accumulated from years of investment by giants such as Samsung. These are assets that cannot be replicated by a sudden capital expenditure, and they are also the micro-foundation of the sustainability of this round of prosperity.

This chain also explains the changes in Japan’s export trade. Japan’s export strengths are automobiles, machine tools, and consumer electronics. Its advantages lie in precision manufacturing and supply chain collaboration, while AI hardware demand is concentrated in advanced process nodes and the highest bandwidth memory. These assets cannot be converted from traditional industries. Japan has not been completely excluded from this cycle. Semiconductor equipment and materials are also benefiting, and exports of complete machines and finished products have fallen.

The buyer has changed, and the nature of the cycle has also changed.

Asian exporters are no stranger to such booms, but the mix of buyers this round is different. In the 2010s, the buyers at the peak of smartphones were consumers. The demand for mobile phones, laptops, and TVs was based on the replacement decisions of hundreds of millions of families. Consumers could use their mobile phones from two to three years, and demand shrank accordingly. The buyers in this round are a handful of cloud vendors. They are competing with each other on computing power deployment. Postponing server orders means giving up share to their opponents.

The sensitivity of orders to interest rates and inflation is therefore likely to be lower than in past Asian export cycles. This is an inference derived from the buyer structure and has not yet been tested in the full cycle. It is both a source of resilience and can easily give people the illusion that “this time is different.”

The precedent exists. During the peak period of optical fiber construction around 2000, telecom operators also believed that bandwidth demand would never be met, and optical equipment orders were also scheduled for several years. After the bubble burst, many electronics export economies in Asia experienced several years of destocking. The smartphone cycle in the 2010s was milder, ending with penetration peaking and demand gradually flattening. What the two histories have in common is that the behavioral logic of buyers at the peak of the boom is completely different from the behavioral logic of buyers when demand weakens.

However, there are two aspects of this cycle that are different from the past, and there are data to follow. The most direct one is the demand structure, which is shifting from training to inference. Inference demand grows with actual usage rather than construction announcements. The AI-related revenue disclosures of major cloud vendors initially support this trend.

Another place is order stock. Even if the growth rate of capital expenditure slows down, shipments of signed orders and projects under construction will still be maintained for a period of time. SK Hynix judged at the latest financial report that the shortage of storage may continue until 2030. Calculations of the supply chain order structure by multiple seller organizations show that if expenditures shrink, the first to be affected will be the network, memory and advanced packaging links, rather than scheduled chip orders.

There is another physical constraint on the demand side that cannot be circumvented, electricity. It directly determines whether capital expenditure can be implemented. Sightline Climate estimated in a February 2026 analysis that 30% to 50% of the 140 large-scale data center projects planned to be put into operation in the United States that year may be delayed or canceled due to power shortages, grid connection delays, insufficient supply of transformers, and approval and community resistance.

This estimate is being realized. Public statistics show that at least 47 large data center projects in the United States have been postponed or canceled in 2026; the midterm election on November 4 has further tightened the policy environment. The governor-elect of Texas advocated suspending the approval of new projects before the state assembly enacted legislation, and the governor-elect of Ohio proposed a conditional suspension. The backlog of grid connection applications in Texas had reached 474 GW when the audit was frozen in August.

There are reports that OpenAI management discussed whether to lower its data center capacity forecast for 2027 on the day after the election. If construction is halted due to power outages, order volume will shrink, regardless of how strong the AI usage trend is.

The True Exposure to Concentration: Market Capitalization, Exchange Rates, and the Two-Speed Economy

This concentration has corresponding manifestations in asset prices, exchange rates and competitive landscape.

The reaction in asset prices has already come out, and preceded the weakening of export data. The third-quarter reports of Samsung and SK Hynix pushed the boom to its peak. Samsung's single-quarter operating profit was 107.4 trillion won, becoming the world's first technology company to have a single-quarter operating profit exceed 100 trillion won. SK Hynix's revenue exceeded 100 trillion won for the first time, and its operating profit margin reached 80%. After the financial report was released, the stock prices of the two stocks fell instead. On October 30, SK Hynix fell 7.08% in a single day. Both stocks fell from their highs during the year. The market is no longer pricing current profits, but whether this round of expenses can be recovered.

U.S. stocks reacted earlier, with the Nasdaq 100 index retracing more than 10% from its high on October 6, and Google and Meta falling about 28% and 19% respectively from their respective highs. Nvidia reacted the most, falling by more than 9% intraday on November 5. This wiped out the gains during the year, and its market value evaporated by nearly US$1.4 trillion from its October high. Every dollar of chip exports in the export data from South Korea and Taiwan is traced back to the source as an order from this company.

This repricing will be even heavier in South Korea, where Samsung and SK Hynix together account for about half of the market capitalization of the country’s benchmark stock index. The employment rate of the latest graduates from Chungbuk Semiconductor High School reached 96.4%. The financial assets and talent flows of the entire economy are being reallocated to a single industry, and the valuation of this industry has just undergone a round of drastic repricing.

The contradiction between exchange rate and competitiveness is deeper. The chip boom has pushed up the Korean won and the New Taiwan dollar. The monetary authorities of South Korea and Taiwan are facing the choice of intervention under the pressure of appreciation. The appreciation benefits imported energy and equipment, alleviates imported inflation, and also reduces the profits of small and medium-sized exporters. The strengthening of the currency due to chips is also weakening the price competitiveness of the remaining approximately 51% of the export basket, and it is these sectors that the recovery depends on to spread. This is a structure that lacks a homeostatic mechanism.

The competitive landscape is also changing. HBM is still divided among the top three, but there is rapid internal redistribution: Counterpoint data shows that Samsung’s share has increased from 21% in the first quarter to 33% in the second quarter, and the gap between it and SK Hynix The gap has narrowed from 37 percentage points to 17 percentage points; UBS predicts that Samsung will overtake 40% to 39% in 2027 based on bit shipments (in terms of revenue, SK Hynix still leads 50% to 33% in the second quarter). The data from the third quarter report also confirms this change. Samsung's self-reported HBM market share in the conference call is about 40%, and said that HBM production capacity in 2027 has been sold out.

Another layer of change is in the pricing method. SK Hynix has signed long-term supply agreements with 14 major customers, locking in more than half of its production capacity from 2027 to 2028. Samsung has signed multi-year agreements with 10 major customers, with the goal of increasing the proportion of production capacity covered by long-term agreements to 60 to 70% by 2030. Competition in the bottleneck link is intensifying, but orders are also locked in advance by long-term contracts. How much of today's excess profits can be extended depends on the bargaining power when these contracts expire.

This dependence has been dispersed geographically. The Vietnamese government disclosed at the end of October that Samsung Vietnam exported approximately US$42 billion in the first nine months of this year, accounting for approximately 12% of Vietnam's total exports. Data released by China's General Administration of Customs on November 10 showed that integrated circuit exports in October were US$39.59 billion, a year-on-year increase of 137.1%. Chips and computer parts together contributed about 40% of China's export growth in the first 10 months. The export destinations of each economy are different, but the product composition of growth is the same type. The diversification of destinations cannot replace the dependence on the same type of demand.

Judgment and three signals

Based on the previous data, this round of prosperity is more like a cyclical upward trend, superimposed on a structural uplift of limited magnitude. Cyclicity comes from the investment return constraints of capital expenditures. The two previous histories have shown how buyer behavior changes when the economy reverses; the structural rise comes from inference demand. Deployed models require continuous computing power as long as someone uses them, which provides a bottom for demand that does not exist before 2023. As for how high this bottom is, there is currently only directional evidence and no reliable quantitative estimates.

The qualitative debate is difficult to answer in the short term, but there are three data windows that can be tracked directly.

The most intuitive thing is the chip export growth rate of South Korea and Taiwan in the next two months. Maintaining above 100% indicates that demand is still accelerating. Falling back to the center of 30% to 40% indicates that the low base effect has been exhausted and growth has returned to the normal range.

The capital expenditure guidance of cloud vendors is the second window. The just-concluded third quarter report shows that the four companies' expenditure expectations for 2027 are still all pointing to growth. However, Microsoft has slightly lowered its expenditure expectations for this year due to delays in production capacity delivery. Constraints have begun to shift from the demand side to the supply side. Any reduction in the expenditure pace will be transmitted to orders in Asia within a quarter.

The last window is whether non-chip exports can keep up. The remaining categories in South Korea increased by 34% in October, still stagnating around 30%, and the two-speed structure shows no signs of loosening.

Following this set of signals, there is sufficient short-term momentum in the coming year. The long-term supply agreement signed by SK Hynix with 14 customers has already locked more than half of the production capacity from 2027 to 2028 in advance, and Samsung's 2027 HBM production capacity has also been sold out; the key variable in the mid-term is the fulfillment of the 2027 spending guidance. Any disappointment will first fall on Asian exports and then be transmitted to the financial reports of U.S. technology stocks; the long-term depends on whether reasoning demand can support a higher platform.

The AI boom has indeed reinvigorated Asia’s export model, making it more reliant than ever on a geographically distant customer with extremely fast decision-making tempo. As long as the construction cycle continues, this investment will pay off; the real question is, when these buyers reduce spending, how much of this industrial zone that has reallocated resources to a single industry will be left does not rely on the growth of this cycle. (This article was first published on Titanium Media APP, author | AGI-Signal, editor | Jiao Yan)

Respect for the original work, if you are qualified, you should be appreciated.

According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.

Express