Why the world's AI giants depend on Korea: HBM, explained
AI's biggest names keep courting Korea's chipmakers. The reason is HBM — the stacked memory Nvidia's AI chips can't run without, and which SK Hynix dominates.
목차
On October 30, 2025, three of the most powerful people in tech crowded into a fried-chicken shop in Seoul. Nvidia’s Jensen Huang, Samsung Electronics chairman Lee Jae-yong, and Hyundai Motor Group chair Chung Euisun ordered chicken and beer — “chimaek” — at a Kkanbu Chicken outlet in the Samseong-dong district. Huang, who had suggested the outing to sample Korea’s chimaek culture, called the food “KFC,” for Korean Fried Chicken (Fortune, Kyunghyang Shinmun).
The dinner was friendly. The reason Huang keeps coming is business. Korea makes a component the entire AI industry now depends on, and one Korean company makes most of it: HBM.
- ~63%
- SK Hynix 2025 HBM shareby revenue, one industry count
- ~27%
- of SK Hynix H1 2025 revenuefrom Nvidia alone
- ~70%
- of Nvidia's HBM4 early ordersreportedly to SK Hynix
- Sold out
- SK Hynix memory outputthrough 2026
What is HBM, and why can’t AI run without it?
HBM stands for High Bandwidth Memory. It is ordinary DRAM — the same kind of memory in a PC — stacked into a tall tower and placed right beside the processor on an AI accelerator. The point of stacking it that close is speed. Training and running a large language model is not just a compute problem; it is a memory-bandwidth problem. The chip spends much of its time moving model weights from memory to the processing cores, and if that pipe is too narrow, the expensive GPU sits idle. HBM is the wide pipe (Introl).
The numbers climb fast with each Nvidia generation.
| Nvidia GPU | Memory | Capacity | Bandwidth |
|---|---|---|---|
| H100 (2022) | HBM3 | 80 GB | 3.35 TB/s |
| B200 (2024) | HBM3E | 192 GB | 8 TB/s |
(Sources: NVIDIA, Introl.) Every one of those gigabytes is HBM, and Nvidia does not make it. It buys it — mostly from Korea.
Why Korea holds the cards
SK Hynix, the smaller of Korea’s two memory giants by overall size, is the clear leader in HBM. One industry tally put its 2025 HBM revenue share around 63% (Astute Group, citing IDC). It became the effective sole supplier of the 12-layer HBM3E used in Nvidia’s GB300 (TrendForce), and Nvidia alone drove about 27% of its revenue in the first half of 2025 (TrendForce). By late 2025 the company said it had sold out its DRAM, NAND and HBM output through 2026 (Notebookcheck). For the next generation, HBM4, it reportedly locked up around 70% of Nvidia’s early orders for the “Vera Rubin” platform (Seeking Alpha).
Source: TrendForce-cited forecasts (projection, not final)
| SK Hynix | 54% |
|---|---|
| Samsung | 28% |
| Micron (US) | 18% |
Samsung, the world’s largest memory maker overall, has been the surprise laggard in HBM and spent 2025 trying to qualify its parts with Nvidia; US-based Micron is the third player. The exact ranking for second place shifts from quarter to quarter and source to source, so treat any single figure as a snapshot. What is not in dispute is the top of the table: on the one memory that AI cannot do without, a South Korean firm sets the pace.
Why the chicken dinners keep happening
This is the bargaining power behind the courtship. And it is about to play out again — in the United States this time. On July 24, 2026, President Lee Jae-myung begins a multi-country trip whose first stop is an AI summit in San Francisco, where he and the chairs of Samsung, SK, Hyundai and Naver are due to meet Huang, OpenAI’s Sam Altman, Anthropic’s Dario Amodei and Broadcom’s Hock Tan (Korea Herald, Korea Times). As of this writing the meetings have not happened and no deals have been announced; the president is expected to host a dinner for the group, and MOUs are anticipated rather than confirmed.
The traffic runs both ways across the Pacific. SK Hynix is building a roughly $3.87 billion advanced-packaging plant for AI memory in Indiana, due to open in 2028, and Samsung is building a foundry in Taylor, Texas. Washington wants more of this supply chain on US soil, which is part of why Korean chips have become a diplomatic asset, not just a commercial one.
Here’s the reality check, though. Korea’s grip is real but narrow. It rests on a single product category, and the moat is not permanent: Samsung is pushing hard to close the HBM4 gap, Micron gives American buyers a domestic alternative that policy may favor, and Nvidia has every reason to want more than one supplier. An edge built on being the only good option lasts exactly as long as that stays true. For now, it holds — which is why, when the world’s AI powers want to talk memory, they still fly to Seoul, or invite Seoul to dinner.
FAQ
- What is HBM and why does AI need it?
- HBM (High Bandwidth Memory) is DRAM stacked into a tall chip placed right next to a GPU. AI models are limited by how fast data reaches the processor, not just raw compute, so HBM's very high bandwidth is essential to Nvidia's AI chips.
- Which company makes the most HBM?
- SK Hynix of South Korea leads — about 63% of the 2025 HBM market by revenue by one industry estimate — and is Nvidia's main supplier. Samsung and US-based Micron compete for second.
- Did Jensen Huang really eat Korean fried chicken with Samsung's chairman?
- Yes. On October 30, 2025, Nvidia's Jensen Huang, Samsung's Lee Jae-yong and Hyundai's Chung Euisun had fried chicken and beer ('chimaek') at a Kkanbu Chicken outlet in Seoul. Huang had proposed it to try Korea's chimaek culture.
- What is 'chimaek'?
- Chimaek is Korea's fried-chicken-and-beer pairing — a national social ritual whose name blends chikin (chicken) and maekju (beer).
