Currency dealers watch monitors as an electronic screen shows South Korea’s benchmark stock index (KOSPI) in a foreign exchange dealing room at the Hana Bank headquarters in Seoul on July 28, 2026.
Jung Yeon-je/AFP via Getty Images
South Korea’s semiconductor industry turned in the strongest two-month stretch in memory-chip history in July, even as the country’s stock market was staging what may be its most volatile month since the Asian financial crisis. Official trade data released Saturday by the Ministry of Trade, Industry and Energy showed July outbound shipments reached $98.89 billion — up 62.8% from a year earlier — with semiconductor exports surging 179% to $41.01 billion. That figure marks the second consecutive month above the $40 billion threshold, a feat no major exporting nation has previously achieved in a single product category, according to Korea JoongAng Daily reporting on the milestone.
The data, published the morning after the KOSPI staged a record single-day recovery, offers the first direct measurement of South Korea’s underlying chip shipment volumes across a period when investor confidence all but collapsed. In July, the KOSPI fell 22.19% — its steepest monthly decline since 1997’s Asian financial crisis — as back-to-back circuit breakers, a missed SK Hynix earnings consensus, and fears about circular AI financing sent nearly three decades of index progress unwinding in a matter of days, as documented in TechTimes’ coverage of the historic market break. The July trade data says the fundamentals kept shipping regardless.
The July total was the second-highest monthly export figure in South Korean history — beaten only by June’s all-time record of $102.25 billion — and exceeded the median Reuters poll forecast of a 59% year-on-year gain, according to the ministry.
What the Trade Data Says About the Bear Case
The question dominating global semiconductor markets in July was whether AI infrastructure demand was real, durable, and end-user-driven, or whether it was partly a financing loop circulating among a small group of companies — a concern crystallized when Nvidia’s reported $250 billion guarantee to back OpenAI’s Ohio data center drew comparisons to circular bookkeeping as reported by TechTimes. SK Hynix’s own record Q2 operating profit — ₩60.54 trillion (approximately $41.9 billion; exchange rate as of August 1, 2026; conversions are approximate), a 557% year-over-year increase — missed analyst consensus by 6.6%, igniting a selloff that eventually wiped approximately $270 billion off the combined market values of Samsung Electronics and SK Hynix within two trading sessions, as detailed in TechTimes’ circuit-breaker coverage.
The July trade data provides one straightforward answer: the actual physical shipment of chips did not slow. It accelerated. Computer exports — a category dominated by enterprise solid-state drives bound for AI server deployments — surged 404% year-over-year. That is not a figure consistent with the circular-financing theory: chips sent to AI data centers show up in customs data as physical exports to physical destinations. They showed up in July in unprecedented volume.
“The popularisation of AI technology, with the industry entering the era of agentic AI, continued to drive demand for memory chips used in servers,” the Ministry of Trade, Industry and Energy said in its monthly statement.
On Friday — one day before the trade data was released — the KOSPI itself delivered a similar verdict when it surged a record 17.91% to close at 6,595, its largest single-day gain in history, driven by a global technology rally after stronger-than-expected earnings from Microsoft, Amazon, and Meta, as reported by CNBC. Samsung Electronics gained 26.81% in that session; SK Hynix rose 29.95%.
Rolf Bulk, semiconductor analyst at Futurum Group, attributed Friday’s recovery to improving confidence that the AI investment cycle remains intact. “There are no signs of the AI infrastructure build-out slowing down, and we remain positive on the sustainability of the cycle,” Bulk told CNBC.
Samsung Posts Record Quarter, Galaxy Posts First-Ever Loss
When Samsung Electronics released its full second-quarter results on July 30, the numbers deepened a paradox that the July trade data now extends: the same chip boom that is setting export records is simultaneously destroying margins in the consumer electronics products that carry those chips.
Samsung’s Device Solutions semiconductor division posted ₩89.2 trillion (approximately $61.8 billion) in operating profit for Q2 — another all-time quarterly record for any memory chipmaker — on revenue of ₩171.5 trillion (approximately $118.8 billion), also a record, according to the Samsung Electronics Q2 2026 results release. The company’s memory business set all-time highs in both DRAM and NAND shipment volumes, fueled by hyperscaler demand that Samsung said it could not fully satisfy.
Simultaneously, Samsung’s mobile and networks division — which sells Galaxy smartphones and buys DRAM and NAND flash as components — reported an operating loss of ₩0.7 trillion (approximately $485 million), its first-ever quarterly deficit since segment-level reporting began in 2011. Revenue climbed year-over-year, lifted by strong Galaxy S26 sales, but surging component costs across the memory market eliminated the margin. The same capacity decisions that benefit Samsung’s chip division — redirecting wafer capacity from commodity DRAM to high-bandwidth memory — raise the input costs for its handset operation. The AI boom has split Samsung’s own business in half.
Samsung told analysts it expects “robust server demand from continued AI infrastructure investment and agentic AI adoption” in the second half of the year, and projected that memory supply constraints will persist through 2028, with multi-year supply agreements becoming standard practice.
The Architecture Behind the Numbers
The reason South Korea’s memory exports have reached levels no exporting nation has previously achieved in a single product category is an engineering constraint that neither pricing nor politics can easily route around.
High-bandwidth memory, or HBM, is the specialized chip at the center of the AI infrastructure buildout. Unlike conventional DRAM — which sits flat on a circuit board, connected to a processor through a 64-bit data pathway that delivers roughly 60 gigabytes per second of bandwidth — HBM stacks multiple DRAM dies vertically, bonding each layer to the next through tens of thousands of microscopic copper pillars called through-silicon vias, or TSVs, as described in detail by the JEDEC HBM4 standard. Each die is thinned to roughly 30 micrometers — about one-third the width of a human hair — so that twelve or more layers fit within the standard height limit. The completed stack sits directly adjacent to the AI processor on a silicon interposer, eliminating the long circuit board traces that slow conventional memory.
The result, in HBM4 (now in mass production), is a 2,048-bit interface delivering up to 2 terabytes per second of bandwidth per stack — approximately 33 times the bandwidth of a commodity DDR5 memory module, per the JEDEC JESD270-4 specification. That gap is the architectural reason Nvidia, AMD, and Google cannot simply substitute standard DRAM for their AI accelerators: switching to a different memory type requires redesigning the entire chip package.
What makes the July trade figures particularly striking is a secondary consequence of HBM production that the headline numbers obscure. Every HBM stack requires approximately three times the manufacturing capacity of an equivalent DDR5 memory module, because the stacking, bonding, and via-etching process consumes orders of magnitude more fab time than flat-die production, a dynamic explained in depth by TechTimes’ HBM supply analysis. As Samsung and SK Hynix redirect wafer capacity toward HBM, they simultaneously remove conventional DRAM from the market. Commodity DDR4 prices reached approximately $21 per unit by end of June 2026 — roughly eight times their level a year earlier. AI infrastructure demand is not just consuming HBM supply; it is tightening commodity memory supply as a structural consequence of the same investment.
SK Hynix commands approximately 54% of the global HBM market and holds roughly 70% of Nvidia’s HBM4 orders for the Vera Rubin platform, per TechTimes’ coverage of the AI memory crunch. Its 2026 output of HBM, DRAM, and NAND is entirely sold out. Nvidia’s current Blackwell B200 GPU carries eight HBM3E stacks totaling 192 gigabytes; the next-generation Rubin platform moves to twelve HBM4 stacks totaling 288 gigabytes, with HBM already accounting for more than 50% of the GPU’s bill of materials, as detailed in TechTimes’ structural reckoning coverage.
The supply crunch in the AI tier has had collateral effects on consumer hardware. Nvidia reportedly cut production of its RTX 50-series consumer gaming GPUs by 30% to 40% earlier this year to redirect memory allocations toward AI data center customers.
Broad Gains Across All Categories
Despite the semiconductor industry commanding the headline, July’s trade data revealed unexpected breadth. South Korea’s exports excluding semiconductors still rose 26% year-over-year, and 19 of the country’s 20 major export categories posted gains — the sole decliner being home appliances, which shed 4.1% to $600 million on sluggish US demand tied to a housing market slump, according to ministry data.
Automobile exports climbed 7% to $6.2 billion, propelled by global demand for hybrid and eco-friendly models. Petroleum product shipments jumped 34.1% to $5.68 billion, partly on higher oil prices linked to continued Middle East uncertainties. Petrochemical exports rose 10.3% to $4.18 billion, while mobile device shipments — buoyed by strong Galaxy S26 sales despite the MX division’s margin squeeze — climbed 51% to $1.81 billion.
“Amid the strong performance of semiconductors, 19 of the country’s 20 major export categories posted gains,” Industry Minister Kim Jung-kwan said in a statement, describing the data as evidence that South Korea had “successfully diversified its export portfolio,” as quoted by Korea Herald.
By destination, shipments to China nearly doubled to $21.68 billion, driven by demand for chips, nonferrous metals, and petroleum goods. Exports to the United States surged 68.7% to $17.43 billion, fueled by AI data center projects rolled out by major US technology companies. Shipments to Southeast Asia climbed 73.7% to $18.80 billion, while exports to the European Union rose 55.7% to $9.38 billion, according to ministry data.
On the import side, South Korea’s inbound shipments rose 26.5% to $68.56 billion, resulting in a trade surplus of $30.32 billion — down from June’s record $36.09 billion but historically elevated. Imports of crude oil, natural gas, and coal spiked 50.1% to $14.48 billion, reflecting both higher energy prices and the country’s own industrial expansion. Non-energy imports climbed 21.4% to $54.08 billion, driven partly by demand for semiconductor manufacturing equipment.
Is the Supply Crunch Getting Worse?
Both chipmakers’ management teams pushed back sharply on any suggestion that AI infrastructure demand is plateauing. SK Hynix President of Corporate Center Song Hyeon-jong told analysts on the Q2 earnings call: “Our major customers are still requesting more memory supply,” as covered by TechTimes. The company has completed long-term supply agreements with approximately ten major customers, typically spanning five years, with volume commitments and price-stabilization mechanisms that provide multi-year demand visibility.
Meritz Securities analyst Kim Sunwoo warned in mid-July that memory chip suppliers were meeting only 75% to 80% of current demand — and that the demand fulfillment rate could fall to just 60% by 2027, even after excluding speculative orders, as reported by TechTimes’ supply crisis analysis. SK Hynix CEO Kwak Noh-jung separately told investors that demand would continue to exceed the company’s full production capacity well beyond 2030 despite aggressive expansion plans, including accelerating its Yongin fabrication complex from a 2045 completion target to 2033, also detailed by TechTimes.
Bank of America has characterized 2026 as “a supercycle similar to the boom of the 1990s,” forecasting global DRAM revenue to surge 51% for the full year, with average selling prices climbing materially across the board, per TechTimes’ rate-hike coverage.
The July cumulative picture underscores the pace: South Korea’s exports for 2026 through late July reached $551.2 billion, up 48.7% year-over-year, according to Korea Customs Service data — a run that would have been unthinkable before the AI-driven memory supercycle began reshaping global semiconductor markets, as confirmed by Korea Herald’s August 1 reporting.
How South Korea Became a Bellwether for Global AI Infrastructure
South Korea’s monthly trade data carries outsized global significance for a structural reason: the country typically reports preliminary trade figures faster than any other major exporting economy, making its customs service print one of the most closely watched early-warning signals for global chip demand, memory pricing trends, and the pace of AI infrastructure buildout.
The July figures arrived with an unusual amount of interpretive weight. In recent weeks, Samsung reported a record quarterly profit that beat operating-profit estimates — ₩89.5 trillion (approximately $62.0 billion), up 1,814% year-over-year — while the broader KOSPI benchmark fell 22.19% in July, its worst monthly performance since the 1997 Asian financial crisis, as reported by Samsung’s newsroom. The divergence between record corporate earnings and catastrophic stock performance reflects a dynamic well-documented in equity research: stocks priced at high forward multiples can fall even on genuinely strong results if the results trail the embedded expectations. That is what happened to SK Hynix. It is not, as July’s customs data makes clear, what happened to actual chip shipments.
Risks Remain Real
Despite the extraordinary momentum, South Korea’s trade outlook is not without headwinds that the monthly record numbers can obscure. Industry Minister Kim cautioned that exporters face ongoing risks from “growing protectionism in major economies and persistent geopolitical uncertainties in the Middle East,” according to Korea Herald reporting on the trade release.
Trade policy remains unpredictable. The US-China tariff truce struck in May 2025 — which reduced US duties on Chinese goods from 145% to 30% — has provided some relief and opened incremental demand from Chinese hyperscalers, but the longer-term policy environment remains contingent on US-China relations and any renewed escalation, as covered by The Hill. Export controls on chipmaking equipment, shifting alliance dynamics, and the pace at which China develops domestic alternatives to Korean memory chips — illustrated by ChangXin Memory Technologies’ $8.6 billion IPO on July 27 — are variables that could alter the trade picture over a multi-year horizon, per TechTimes’ circuit-breaker analysis.
China’s CXMT poses a credible medium-term threat to commodity DRAM pricing, but independent analysts have confirmed it cannot yet challenge South Korean chipmakers’ position in the high-bandwidth memory tier that drives their record margins — HBM manufacturing requires a level of through-silicon via precision and die-stacking yield that CXMT does not currently possess, as analyzed by TechTimes. As of mid-2026, approximately 2% of CXMT’s wafer capacity targets HBM production, and at the HBM3E generation — already one behind what Samsung and SK Hynix are currently shipping, per TechTimes’ DUV lithography coverage.
Some analysts also note that the memory price environment, while still elevated, may eventually face correction pressure as Samsung, SK Hynix, and Micron accelerate production capacity expansion into 2027. Whether the structural shift in AI infrastructure demand is large and durable enough to absorb that supply is the defining question for what has — for now — been one of the most extraordinary export booms in South Korean economic history.
Frequently Asked QuestionsWhat does “second consecutive month above $40 billion in chip exports” mean, and why is it significant?
It means South Korea shipped more than $40 billion worth of semiconductors in both June and July 2026 — a threshold no major exporting economy has crossed even once in a single month for a single product category before the AI memory supercycle. June’s figure was $44.82 billion; July’s was $41.01 billion. The significance is straightforward: these are physical shipments of physical chips confirmed by customs data, not earnings forecasts or stock valuations. The consecutive-month feat establishes that the volume of AI-bound memory chips leaving South Korea is not a one-month spike but a sustained new demand level. For enterprise buyers of AI infrastructure, it signals that the supply crunch is ongoing and unlikely to ease soon. For equity investors who sold Korean chip stocks in July on fears of demand weakness, the trade data directly contradicts the underlying thesis.
Why did the KOSPI fall 22% in July when Samsung and SK Hynix were posting record profits?
The KOSPI’s July decline was driven primarily by three factors that are distinct from actual chip shipment fundamentals. First, both Samsung and SK Hynix had been priced for outcomes even more extreme than their record results — when SK Hynix’s Q2 operating profit of approximately $41.9 billion missed analyst consensus by 6.6%, the gap between expectation and result triggered forced selling despite the result being historically extraordinary. Second, three structural threats arrived simultaneously: China’s CXMT completed Asia’s largest IPO of 2026 with capital designated for DRAM expansion, a state-backed Chinese group began domestic production of immersion lithography machines, and renewed scrutiny fell on Nvidia’s $250 billion financing guarantee to OpenAI. Third, the KOSPI’s concentration — Samsung Electronics and SK Hynix together represent more than 40% of the index’s total market capitalization — meant that selling pressure in two stocks produced a market-wide collapse. Friday’s record 17.91% rebound reflected the same mechanism in reverse: when confidence returned after strong Microsoft, Amazon, and Meta earnings, the same concentration amplified the recovery.
How does South Korea’s high-bandwidth memory boom make other electronics more expensive?
When Samsung and SK Hynix redirect wafer capacity from standard DRAM production to high-bandwidth memory production, they remove roughly three units of conventional memory from the market for every one unit of HBM they produce — because the stacking and via-etching process required for HBM consumes approximately three times the manufacturing capacity of flat-die DRAM. This is not a pricing strategy; it is an architectural consequence of how HBM is made. The result is that commodity DRAM supply tightens as HBM demand rises. By end of June 2026, commodity DDR4 prices had reached approximately $21 per unit — roughly eight times their level a year earlier. Samsung’s own Galaxy smartphone division posted its first-ever quarterly operating loss in Q2 2026 precisely because the component costs its handset business pays are elevated by the same memory market dynamics that its chip division is profiting from.
Will the AI memory shortage get worse before it gets better?
Current industry guidance points toward 2027 as the year the shortage intensifies most severely before new capacity from Samsung, SK Hynix, and Micron begins to provide meaningful relief. Meritz Securities analyst Kim Sunwoo estimated in July 2026 that suppliers are currently meeting only 75% to 80% of demand, and projects that figure could fall to 60% by 2027 even after excluding speculative orders. SK Hynix CEO Kwak Noh-jung has told investors that demand will continue to exceed full production capacity well past 2030. Samsung’s Q2 results confirmed that it cannot produce HBM fast enough to meet current orders and projected that multi-year supply agreements — rare before 2025 — are now becoming standard practice. Whether that supply-demand gap proves durable beyond 2028, when meaningful new fab capacity is expected to come online, depends on whether AI infrastructure investment continues growing at the pace that July’s customs data suggests it has sustained.