The Chinese and South Korea flags fly outside the Suzhou Samsung Electronics Company facility in Suzhou on September 5, 2012.
CN-STR/AFP via Getty Images
Samsung Electronics shipped more chips to China than to the United States in the first half of 2026 — a reversal captured in precise numbers by the company’s semiannual report filed August 14 that gives the clearest view yet of how AI-fueled demand is redistributing the world’s most important semiconductor flows. Chip exports to China reached ₩88.6004 trillion (approximately $62.5 billion USD — exchange rate as of August 15, 2026; conversions are approximate), a 207.7% year-on-year jump. Exports to the United States came in at ₩70.6466 trillion (approximately $49.8 billion USD), more than doubling from ₩33.4759 trillion in the first half of 2025. Both streams of shipments — to rival superpowers running competing AI buildouts — flow through manufacturing operations that depend on a US government annual export license scheduled to expire December 31, 2026.
What the bilateral breakdown reveals is not simply that Samsung’s business is booming. It is that the world’s largest memory chipmaker is simultaneously the primary AI chip supplier to the two countries whose governments are trying hardest to reduce economic dependence on each other — and that the regulatory arrangement keeping Samsung’s China production running is renewed one year at a time.
How HBM Turned Samsung’s Chip Division Into a Profit Engine Unlike Any Before
The export numbers are a downstream consequence of a technology shift that happened inside Samsung’s products. High-bandwidth memory — HBM — is the engineering answer to a bottleneck that would otherwise cap AI model performance at current scale. In conventional memory architectures, data must travel from memory chips to a processor through a relatively narrow interface. For large-scale AI training and inference, that path cannot move data fast enough. HBM solves this by stacking multiple DRAM dies vertically, connecting them through thousands of microscopic channels called through-silicon vias (TSVs), and placing the resulting stack directly adjacent to the processor on a silicon interposer. The resulting structure delivers data across a bus 32 times wider than standard DDR5.
Samsung’s sixth-generation HBM4, which entered mass production in February 2026, implements the JEDEC JESD270-4 standard. Its interface is 2,048 bits wide — double the 1,024-bit interface of its predecessor, HBM3E — and delivers at least 2 terabytes per second of bandwidth per stack. Each Nvidia Vera Rubin GPU carries eight such stacks, producing an aggregate bandwidth exceeding 16 terabytes per second. That is the bandwidth frontier-scale AI models — the large language and multimodal systems that underpin today’s most capable AI applications — require to run at full speed. Nothing cheaper substitutes for it.
Because HBM cannot be replaced by a lower-cost alternative at the frontier AI level, and because Samsung and SK hynix together produce approximately 90% of the world’s HBM supply, Samsung’s Device Solutions (DS) semiconductor division generated operating margins of roughly 70% in the second quarter — a figure confirmed in Samsung’s Q2 results coverage that surpasses Nvidia’s own margins in the same period. The DS division accounted for 68.5% of Samsung’s total first-half sales of ₩305.4 trillion (approximately $215.4 billion USD) and was responsible for 97.4% of total operating profit — approximately ₩142.9 trillion (approximately $100.8 billion USD).
In the first quarter of 2026, DS operating profit reached ₩53.7 trillion (approximately $37.9 billion USD). In the second quarter, it rose to ₩89.2 trillion (approximately $62.9 billion USD). Both figures are quarterly records not just for Samsung but for any company’s semiconductor division globally. Samsung’s DRAM market share rose 5.4 percentage points to 39.4% in the first half of the year, as demand for HBM products redirected procurement toward the company’s highest-technology offerings.
The one part of Samsung that HBM is actively harming is Samsung’s own finished-products business. The same memory price surge that powers DS division margins drove up the cost of mobile memory bought by Samsung’s Device eXperience (DX) division — the unit that makes Galaxy smartphones — by 211% compared with last year’s annual average. As a result, the DX division reported an operating loss in the second quarter, the first in the division’s history. The same product line that makes Samsung the most profitable semiconductor company on earth is simultaneously the reason its smartphone business is struggling.
Why China Received $12.7 Billion More Samsung Chips Than the US Did
The $12.7 billion gap between Samsung’s China and US export totals in the first half does not reflect a strategic choice to prioritize Beijing over Washington. It reflects product mix and geography. Samsung’s China exports include not only server-oriented products — HBM, DRAM for AI data centers — but also the mobile product categories that remain the backbone of Chinese consumer electronics: LPDDR DRAM, NAND flash, image sensors, and display driver ICs. The US market, by contrast, receives primarily Samsung’s highest-value AI memory products, routed to Nvidia, AMD, and the major cloud providers that dominate American AI infrastructure spending.
The 207.7% year-on-year growth in China exports reflects several overlapping forces. Chinese cloud companies and AI infrastructure builders have been aggressively acquiring server memory, accelerating procurement ahead of potential further tightening in US export controls on AI-adjacent components. Commodity DRAM and NAND prices surged roughly 90-95% in the first quarter of 2026 alone — a price inflation that inflates the dollar value of each unit shipped without a corresponding increase in volume. And Samsung’s Xi’an NAND facility, Samsung’s largest NAND production site outside South Korea, continued to operate at scale throughout the first half, producing a meaningful share of the memory flowing into China.
For context: in the first half of 2023 — the last period before US export controls on advanced semiconductor equipment significantly tightened — Samsung’s chip exports to China totaled approximately $44.6 billion for the full year. In the first six months of 2026, the comparable figure has already reached $62.5 billion. The scale of the expansion underscores both the magnitude of the AI-driven demand wave and the degree to which Samsung has benefited from being one of the few companies in the world capable of supplying it at volume.
Xi’an and Wuxi: Running on Borrowed Permission
Samsung’s Xi’an NAND flash facility in China is one of the largest memory production sites in Asia. Its continued operation depends not on permanent authorization but on a one-year US government export license that was granted on December 30, 2025 and that expires December 31, 2026. That license replaced Samsung’s prior status as a “validated end user” (VEU) — a permanent pre-approval that had allowed Samsung’s China fabs to receive US-controlled chipmaking equipment without applying for individual export licenses. Washington revoked VEU status in 2025, converting the arrangement to annual renewals under what the Bureau of Industry and Security framed as closer oversight of advanced semiconductor tool flows into China.
SK hynix operates under identical terms — its DRAM fab in Wuxi and NAND plant in Dalian both run on the same annual license framework. SK hynix’s Dalian facility restarted construction and equipment installation in August 2026, with mass production targeting H1 2027, expanding NAND output by approximately 50%. Across both Korean chipmakers, the annual license structure means that a US government renewal decision in late 2026 — a decision that has not yet been made, and that will be weighed against the geopolitical conditions prevailing at year-end — will determine whether these fabs can continue to import the US-origin tools they require to maintain production quality and replace equipment that breaks down.
The practical risk is not an immediate shutdown. Samsung’s Xi’an facility could continue operating existing equipment for some time without new US-origin tool imports. But over months, the inability to import replacement parts, upgrade lithography equipment, or bring in new deposition and etch systems would degrade production yield and capacity. A denial — or even a delayed renewal — would create a supply gap in NAND flash that would ripple through data center procurement globally, since Samsung’s Xi’an plant produces a significant share of the world’s enterprise NAND.
The December 31 binary decision point is the structural risk this article’s headline numbers cannot fully convey on their own. Samsung shipped $62.5 billion in chips to China in the first half of 2026. Whether the second half of 2026 flows uninterrupted into 2027 at anything like that volume depends on an annual approval that no company controls and that the US government has not yet committed to granting.
Samsung’s H1 Investment: A Record That Bets the Supercycle Continues
The same semiannual report that disclosed the bilateral export breakdown also revealed that Samsung invested a record ₩55 trillion (approximately $38.8 billion USD) in research and development and capital expenditure in the first half of 2026. R&D spending reached ₩27.36 trillion (approximately $19.3 billion USD) — the highest first-half R&D total in Samsung’s history. Capital expenditure hit ₩28 trillion (approximately $19.8 billion USD), also a record.
This investment program funds the engineering capabilities required to maintain competitive advantage in HBM. On May 29, 2026, Samsung shipped the industry’s first HBM4E samples to major customers — twelfth-generation-stacked chips built on Samsung’s 4-nanometer logic process for the base die, delivering 3.6 terabytes per second of bandwidth per stack and a pin speed of 14 gigabits per second. HBM4E is the generation Samsung is targeting for Nvidia’s next-generation AI platform after Vera Rubin, and shipping samples to customers more than six months ahead of competitors positions Samsung for the design-win qualification race that determines memory supply allocations for each new GPU generation.
SK hynix held roughly 60-70% of Nvidia’s HBM4 supply allocation for the Vera Rubin platform, with Samsung supplying approximately 25-30%. In the first half of 2026, SK hynix drew ₩17.61 trillion (approximately $12.4 billion USD) in revenue from Nvidia — representing 13.35% of SK hynix’s total revenue for the period, down from 24% of annual revenue last year as SK hynix diversified its customer base toward hyperscalers building custom AI chips.
What $113 Billion in Six Months Reveals About the AI Supply Chain
Samsung’s combined first-half chip exports to the US and China — approximately $112.3 billion at current exchange rates — represent one of the most concentrated trade flows in the history of the semiconductor industry. A single company, headquartered in South Korea, serving two rival economic superpowers that are simultaneously its largest customers and the authors of the regulations constraining how it may serve them.
That situation was not designed. It emerged from decades of Korean memory investment, from the specific architectural requirements of large-scale AI computing, and from the fact that HBM manufacturing at yield requires a level of through-silicon via precision and die-stacking process control that only two companies in the world — Samsung and SK hynix — currently possess at production scale. The result is that the AI buildout happening in American data centers and the AI buildout happening in Chinese data centers both depend on the same Korean fabs, the same Korean engineers, and in the case of Samsung’s Xi’an operation, the same annual US government permission slip.
Memory chip prices have remained sharply elevated throughout 2026, amplifying the dollar values of every shipment in both directions. Whether those prices sustain as supply expands, and whether the December 31 annual license renewal goes smoothly, will shape whether the second half of 2026 looks anything like the first. For now, Samsung’s H1 semiannual report documents a six-month period in which the AI memory supercycle produced numbers that the semiconductor industry had not seen before — and in which the tightrope between two superpowers held.
Frequently Asked QuestionsWhy did Samsung’s chip exports to China triple while US exports merely doubled?
China’s growth was larger in percentage terms because the base was smaller — Samsung’s China exports were ₩28.8 trillion (approximately $20.3 billion USD) in the first half of 2025, compared with ₩33.5 trillion (approximately $23.6 billion USD) to the US. The absolute dollar gap between the two — China receiving about $12.7 billion more than the US in H1 2026 — reflects Samsung’s China product mix: in addition to server memory, exports to China include mobile DRAM (LPDDR), NAND flash for consumer and enterprise storage, image sensors, and display driver ICs that are less prominent in Samsung’s US-facing product portfolio. Chinese AI infrastructure companies and cloud operators also front-loaded memory procurement through early 2026 ahead of potential export control tightening.
What is the annual US export license for Samsung’s Xi’an facility, and when does it expire?
The US Bureau of Industry and Security replaced Samsung’s and SK hynix’s permanent “validated end user” pre-approval with annual export licenses in late 2025. The current annual license for Samsung’s Xi’an NAND flash facility was granted on December 30, 2025, and expires December 31, 2026. Without renewal, Samsung would need individual export licenses for each shipment of US-controlled semiconductor manufacturing equipment to the Xi’an site — a practical obstacle that would eventually degrade the facility’s production capability even if existing equipment continued to run. The renewal decision has not yet been made.
How does HBM actually work, and why can’t AI chips use cheaper memory instead?
High-bandwidth memory stacks multiple DRAM dies vertically and connects them through thousands of microscopic channels (through-silicon vias), then places the resulting stack directly next to the processor on a silicon interposer. HBM4, the current generation, delivers at least 2 terabytes per second of bandwidth per stack across a 2,048-bit interface — roughly 32 times wider than DDR5. AI accelerators like Nvidia’s Vera Rubin are physically designed around HBM configurations; their compute blocks are sized and clocked to assume HBM’s specific bandwidth. Substituting a slower memory type would leave the compute cores starved of data, reducing effective AI throughput far more than the memory price difference would justify. That non-substitutability is why Samsung and SK hynix can charge margins approaching 70-76% on HBM production.
Does the AI memory boom help or hurt the average consumer?
Both, at the same time. Samsung’s chip division is charging elevated prices for DRAM and NAND throughout the supply chain — prices that flow through to every device that uses memory. Samsung’s own semiannual report shows that mobile memory prices for its DX (Galaxy smartphone) division rose 211% year-on-year in the first half of 2026, contributing to the first operating loss in the division’s history. Those same elevated prices affect PC makers, storage device manufacturers, and ultimately retail buyers. On the other side, the AI model capabilities that HBM enables — faster inference, larger context windows, more capable agents — are delivered to consumers through cloud services. Whether the capability gains are worth the device price inflation depends on how much of each you use.