Memory Became the Gate on AI Capital — Reading SK hynix's $29B Listing

SK hynix filed for a $29 billion Nasdaq listing — the second-largest IPO on record after SpaceX. A company that holds one slot in the chip stack, high-bandwidth memory, is being repriced as the most expensive bottleneck in the AI cycle. The lesson isn’t about chips; it’s that the memory chips wait on has become the gate on capital. Here’s how founders should read it.

What Happened

On June 24, SK hynix filed to list American depositary receipts on the Nasdaq: 17.79 million new shares, raising $29 billion. If it clears, it would be the second-largest IPO on record after SpaceX’s recent $85.7 billion listing, with trading penciled in for around July 10 on tentative dates. The number makes sense for one reason. SK hynix produces more than half of the world’s HBM — high-bandwidth memory, the specialized RAM every AI accelerator runs alongside. Revenue for the quarter ending March 31 hit $38 billion, up 198% year over year, at a 77% net profit margin. Memory used to be a volatile commodity; now three-quarters of a quarter’s revenue drops to the bottom line. The company recently unveiled iHBM, a proprietary thermal-management architecture, and says it will pour all of the IPO proceeds into facilities to defend that lead — the first fab (Y1) at the Yongin cluster, advanced-packaging lines for AI memory, and equipment including EUV scanners. It sits at the center of a $200B-plus Korean semiconductor investment push.

What This Means for Founders

If Nvidia is the face of the AI cycle, SK hynix’s $29 billion reveals who actually controls the bottleneck behind that face. Running a model takes more than compute — it needs memory feeding data into that compute fast enough, and one company supplies more than half of that single slot. The market handing this firm the second-largest IPO valuation ever is a signal: capital pays the steepest price for the hardest-to-unblock chokepoint. For founders, this isn’t abstract macro news. First, it’s a reminder that the price and availability of the infrastructure you depend on are lashed to a single bottleneck. When HBM supply tightens, GPU instance pricing climbs, and that cost lands directly in your inference unit economics. Second, it shows where value flows. The most durable profits in the AI gold rush belong neither to the model sellers nor the app sellers, but to whoever owns the one component everyone queues to buy — a 77% margin is not an application-layer number. Third, this listing breaks the assumption that deep-tech hardware can’t win capital markets. Own a deep enough bottleneck and capital arrives in software-multiple size. But that depth is the kind nobody catches in six months; it’s defended only by years of process assets — thermal management, yield, EUV.

What You Can Do Now

First, map where the real bottleneck in your stack sits. The cost and ceiling of an AI product are usually set by the scarcest slot — here, HBM bandwidth — and if you can’t name it, you can’t control your cost structure. Second, model inference cost as a function of that bottleneck. Lay in sensitivity scenarios that assume a memory-constrained regime now, so a swing in GPU pricing doesn’t shake your runway later. Third, hunt for your own irreplaceable slot. SK hynix’s lesson is that owning the narrow passage everyone must cross lets you collect margin from every layer above it — look for whether such a passage exists in your domain. Fourth, if you can’t own depth, diversify. Betting your entire inference infrastructure on a single supplier hands your cost sheet to one company’s pricing power.