SK Hynix listed on US markets around July 10, 2026, in what's reported as the largest American Depositary Receipt offering in history — raising roughly $28-29 billion. The money funds new chip fabrication plants in South Korea to meet surging AI infrastructure demand. Behind the headline number is a lesson about where AI's real constraint sits.
It's not just GPUs — it's memory
Everyone talks about GPU shortages, but the deeper bottleneck is high-bandwidth memory (HBM) — the fast memory stacked next to AI accelerators. Every frontier training cluster and high-end inference deployment needs it, and Nvidia's entire data-center roadmap is gated on HBM supply. SK Hynix beat rivals to each successive HBM generation and locked in Nvidia qualification wins, turning memory into a near-sole-source business with the pricing power to match.
The AI boom is often pictured as a GPU race. Just as much, it's a memory race — and memory is where a handful of firms hold the choke point.
Why it matters
A record listing to fund more fabs tells you how much capital the industry believes AI infrastructure will keep demanding. It also underlines a strategic reality: AI capability at scale depends on a physical supply chain — chips, memory, power — controlled by a few players. That shapes who can build, what it costs, and where the leverage lies.
The takeaway
For anyone building on AI, the memory story is a reminder that "compute" is a physical, contested resource. Prices and availability of inference are downstream of HBM supply. It's a good moment to appreciate that the abstractions we build on rest on fabs, memory stacks, and the companies racing to supply them.
