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.
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 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.
Source: techtimes