EWY Vs. EWH: Michael Burry Has This Backwards

The knee-jerk selloff in Korean memory names looks more like a positioning flush than a durable fundamental break. A cheaper, better coding model matters only if the market assumes AI spend is a fixed pie; in practice, lower inference costs usually expand usage, which supports total token throughput, device memory content, and eventually HBM demand even if frontier-model margins compress. The real revenue pool is still captured by the model/app layer, but the infrastructure layer often sees the second-order benefit from higher utilization.

Near term, the risk is sentiment-driven multiple compression across Korea semis rather than a cash-flow revision. That means the first move can overshoot for 1-3 weeks, but the thesis only becomes bearish if Samsung/SK Hynix commentary starts signaling weaker HBM/DRAM pricing or if hyperscaler capex is revised down over the next 1-2 quarters. If anything, a stronger Chinese model ecosystem is more likely to widen the global AI buildout than to shrink it.

The contrarian miss is that "cheap model" headlines are being read as token-deflationary for memory when the economically relevant variable is unit volume of inference, not model price per se. Over 6-18 months, efficiency gains should favor the companies that sell picks-and-shovels into a broader AI install base, while the true pressure sits on frontier-model monetization and any over-earned AI software multiples. KEP is not a clean express; its upside is indirect and gated by regulated tariff decisions, not model launches.