South Korea Chip Exports Approach Half of Economy; First Trillion-Dollar Year Due in November
South Korea’s Ministry of Trade, Industry and Energy released October 2026 export data Sunday to figures that would have been implausible at any prior point in the country’s industrial history: semiconductor shipments reached $56.4 billion for the month, an increase of 259.4% from October 2025, and accounted for 48.9% of all South Korean goods exported — making chips nearly one dollar out of every two that left the country. Total exports across all categories hit $115.3 billion, a 93.7% year-on-year gain and the highest October figure in the country’s history, extending a streak of monthly records that has turned 2026 into the most extraordinary year for Korean trade in modern memory.
The driver is structural, not cyclical. South Korea produces an estimated 80–90% of the world’s High Bandwidth Memory — the specialized, stacked-DRAM architecture that AI accelerators require to perform inference at scale — and demand for that memory has grown in direct proportion to the global AI infrastructure buildout. As hyperscalers spend hundreds of billions of dollars expanding AI data center capacity, every GPU they deploy must be paired with HBM that, with very few exceptions, can only come from South Korea.
The October Numbers in Full
Among the 20 major export categories tracked by the Ministry of Trade, 18 grew year on year. The two exceptions — shipbuilding, down 43.6% to $2.64 billion, and household goods, down 13.4% — stand in stark contrast to the triple-digit growth posted by the categories most directly tied to AI infrastructure.
Computer exports posted the single highest growth rate of any category: up 496.6% to $5.85 billion. Within that figure, the most revealing sub-item is server solid-state drives. Server SSD exports surged to $5.64 billion in October 2026 from $820 million in October 2025 — a 584.3% increase in a single year — reflecting the extraordinary pace at which hyperscalers and AI cloud companies are scaling their data center storage capacity. The Ministry’s report describes this as a direct consequence of AI services expansion driving data center investment.
Automobile exports grew 18.1% to $6.55 billion, setting a record for any October on record. The calendar assisted: the Chuseok holiday fell in September this year rather than October, adding two working days compared with October 2025. Within autos, the electrification shift is visible in the export mix. Hybrid vehicle exports rose 53.1% to $2.06 billion, and electric vehicle exports climbed 41.6% to $1.04 billion, while internal combustion engine vehicle exports declined fractionally—down 0.4% to $3.45 billion.
Petroleum product exports rose 80.0% to $6.92 billion, and petrochemical exports climbed 27.5% to $4.0 billion, but in both cases the value gains were driven almost entirely by price rather than volume — actual export volumes fell 7.5% and 6.8% respectively, as elevated international oil prices pushed unit export prices higher. The divergence between value and volume growth in these sectors underscores that October’s headline surge is not uniformly broad-based.
Shipbuilding’s 43.6% decline is largely attributable to a base effect: October 2025 included the delivery of a single offshore platform valued at approximately $2.47 billion, creating a comparison this month’s more typical activity cannot match.
How HBM Works — and Why AI Can’t Run Without It
To understand why South Korea’s export statistics look the way they do, it helps to understand the physical architecture of the memory that is driving them.
Conventional DRAM — the kind found in desktop computers and smartphones — transfers data over a 64-bit interface. DDR5, the current consumer standard, delivers approximately 89 gigabytes per second of memory bandwidth. That figure is the bottleneck for AI inference: a large language model with 70 billion parameters stores approximately 140 gigabytes of model weights that must be loaded and accessed for every token the model generates. Standard DDR5 memory would leave a GPU’s compute cores idle more than 90% of the time, waiting for data.
Source: Techtimes