Every AI model you've used this year — every chat reply, every generated image, every agent run — passed through high-bandwidth memory on its way to you, and there's a decent chance that memory was made by SK Hynix. On July 10 the Korean chipmaker began trading on the Nasdaq under the ticker SKHY, in an American depositary receipt offering worth roughly $29 billion — one of the largest listings of the year, and a milestone with a simple meaning: the AI buildout's least glamorous bottleneck now has a ticker American money can buy directly.
Why memory is the choke point
The public conversation about compute fixates on GPUs, but the dirty secret of the accelerator business is that the processor spends much of its time waiting for memory. High-bandwidth memory — HBM, the vertically stacked chips bonded next to the GPU die — determines how fast a model can actually think, and it is chronically, structurally scarce. SK Hynix has spent two years as the leading supplier of it, with demand so far ahead of supply that capacity through next year has been effectively spoken for. A Nasdaq listing doesn't change the physics, but it changes the capital: this is a war chest raised in the currency of the customers driving the shortage.
Why list in New York at all? Because that's where the customers — and the multiple — live. Korean-listed technology companies have traded for years at a persistent discount to US peers, and SK Hynix's revenue base has migrated decisively toward American hyperscalers building AI capacity. An ADR listing puts the stock in front of the index funds and AI-thematic capital that already own Nvidia and Broadcom, denominated in the currency its biggest customers spend. There's a strategic layer too: the next memory generation, HBM4, demands capital expenditure at a scale that makes even a profitable memory giant think hard about funding sources. Raising in the deepest capital market on earth, at AI-era multiples rather than memory-cycle multiples, is how you finance a bet that big without betting the company.
GPUs get the keynotes. Memory gets the margins.
The week the supply chain went global
The listing is the headline, but two smaller moves this week sketch the same map. FuriosaAI — the Korean chip startup whose RNGD accelerators made their name on efficiency — deployed servers at an Equinix data center in Lisbon, its first European infrastructure foothold, selling inference capacity on the continent where data-residency rules increasingly demand local compute. And OpenAI's announcement that it will serve its flagship Sol model on Cerebras wafer-scale hardware at up to 750 tokens per second put a spotlight on the challenger-silicon thesis: that inference, unlike training, doesn't have to belong to the incumbent GPU stack. Three stories, one pattern — the compute layer is diversifying by geography, by vendor, and now by shareholder.
What to watch from here: whether SKHY's debut pricing holds through its first earnings as a US-listed name (memory is famously cyclical, and public markets have short memories about that), whether the HBM shortage eases as new capacity comes online next year, and whether the challenger-inference bets — Cerebras, Furiosa, and their cohort — convert marquee deployments into recurring volume. The picks-and-shovels story of this boom is no longer a private-market rumor; as of this week it trades in daylight.
- SK Hynix listed on Nasdaq as SKHY in a ~$29 billion ADR offering.
- HBM memory, not GPUs, decides how fast models think — and it's chronically scarce.
- New York gives it AI-era multiples and the currency its hyperscaler customers spend.
- Same week: FuriosaAI landed in Lisbon; OpenAI put Sol on Cerebras.
- Caveat: memory is famously cyclical, and public markets have short memories about that.