WHAT HBM ACTUALLY IS
High-bandwidth memory stacks DRAM dies vertically and connects them with through-silicon vias — copper wires drilled straight through the silicon. The result: roughly 10× the bandwidth of standard DRAM, which is what AI training chips need to keep their tensor cores fed.
THE OLIGOPOLY
Three companies make all of the world's HBM. SK Hynix took an early lead with HBM3, Samsung is racing to qualify HBM3E with Nvidia, and Micron is the distant third. The barrier is not design — it is yield on stacks where a single defective die ruins the whole package.
THE INDEX ILLUSION
Cap-weighted indices reward concentration. When two stocks are 60% of an index, the index is no longer a country bet — it is a leveraged bet on those two stocks. Korea's Kospi, Taiwan's Taiex (TSMC), and the Nasdaq-100 (Mag 7) all share this trait.
THE HYPERSCALER CYCLE
Memory is the most violently cyclical part of semiconductors. Capex commitments by four buyers — Microsoft, Google, Meta, Amazon — set demand for an entire industry. When their capex slows, DRAM and HBM prices have historically halved within two quarters.
THE KOREA DISCOUNT
Korean equities have historically traded at a structural discount to global peers — a function of chaebol governance, cross-shareholdings, and limited shareholder rights. The current rally is partly a re-rating, partly a bet that this time the cycle won't crack.
WHY HYPERSCALER CAPEX MATTERS
AI training capex is concentrated in a handful of buyers, and their announcements move memory prices more than any macro indicator. Watch their guidance, not the Federal Reserve.