South Korea’s $950 Billion AI Push Is Really A Chip Grab
eneration HBM. SK Telecom also outlined a 2-gigawatt data center that’s meant to use Nvidia’s Vera Rubin chips and SK Hynix’s HBM4, due online in 2027. Samsung, meanwhile, said it will work with Broadcom on a next-generation AI accelerator tied to Samsung’s HBM, plus sub-2-nanometer foundry production and advanced packaging. The common thread: US AI firms are trying to lock in memory and manufacturing allocation now, because demand is outrunning supply, and SK Chair Chey Tae-won says even five-year forecasts for memory could end up “too low.”
Why should I care?
For markets: Nvidia’s $500 billion SK tie-up could shift HBM from cyclical pricing to allocation pricing.
When customers sign multi-year offtake deals, demand that might normally hit the spot market gets pulled into contracted volumes. That can soften the usual boom-bust swings in commodity memory, and it puts more weight on who has guaranteed supply, at what contract terms, and whether suppliers hit delivery milestones. For SK Hynix and Samsung Electronics, that raises the stakes on execution: HBM profitability may depend less on quarterly DRAM price moves and more on long-run allocation and capacity ramp-ups. It also lifts the competitive bar for Micron, since catching up isn’t just about technology, but about securing enough manufacturing slots early enough to meet contract commitments.
