THE OLIGOPOLY
Samsung, SK Hynix, and Micron produce roughly 95% of the world's DRAM. Four decades of sub-10nm lithography investment and multibillion-dollar fab costs eliminated every other competitor, leaving three companies to allocate global memory supply.
WHY HBM IS DIFFERENT
High-bandwidth memory stacks DRAM dies vertically and bonds them with through-silicon vias, sitting millimeters from the GPU on a shared interposer. An Nvidia H100 ships with 80GB of HBM3; a B200 ships with 192GB. Memory is now a larger share of an AI accelerator's bill of materials than the logic die itself.
THE PRODUCT IS THE PROCESS
A CPU design can be ported between foundries. DRAM cannot — each company's memory cells are inseparable from their proprietary fabrication steps. There is no TSMC-for-DRAM that newcomers can outsource to, which is why the oligopoly has been stable since the 1990s shakeout.
THE CYCLE THAT WASN'T
DRAM has historically been viciously cyclical: oversupply crashes prices every 2-3 years, forcing one player to take a quarter of losses before the cycle resets. AI demand has compressed that cycle — Samsung's guidance that 2027 will be tighter than 2026 implies the bust phase has been deferred, possibly structurally.
THE SAMSUNG–HYNIX GAP
SK Hynix shipped HBM3E to Nvidia first and held a roughly one-generation lead through 2024. Samsung has been qualifying its HBM3E with Nvidia for over a year — the trillion-dollar valuation depends partly on that qualification finally clearing for HBM4.
WHY KOREA
South Korea's chaebol model — vertically integrated conglomerates with patient capital and state-backed financing — was purpose-built for the memory business. Samsung Electronics alone accounts for roughly 20% of Korea's exports and a fifth of the KOSPI's market cap. The country's macro story moves with one company's HBM yields.