Prediction: This Will Be the Next AI Stock to Split (Hint: It's Not Micron or Sandisk)

Roundhill Memory ETF (DRAM) gained ~78% from ~$26 at launch (Apr 2) to >$81 in under two months, reaching nearly $10B in net assets within 45 days. The run is attributed to HBM demand from AI accelerators outpacing supply, with holdings concentrated in SK Hynix (~24.5%), Samsung (~24.5%), and Micron (~26% including ~10% in leveraged shares). The article flags potential share split later this year due to the rapid price/momentum, while noting concentration and cyclical memory risk.

Analysis

The market is increasingly pricing memory as a scarce input to AI capex rather than a cyclical commodity, which supports continued multiple expansion in MU and SSNLF, with SNDK as a lower-quality but still levered beta expression. The immediate winner is the memory complex itself: when supply is tight, pricing power accrues faster to the few vendors with HBM qualification and advanced packaging capacity, while downstream buyers absorb the cost through higher server BOMs rather than cutting orders.

Second-order effect: the real margin pressure is likely to show up in AI platform economics, not in headline chip demand. NVDA can likely pass through some of it near term, but hyperscalers and server OEMs face a quieter squeeze as memory content per accelerator rises; that can slow unit growth or shift procurement toward fewer, higher-ROI deployments. In other words, the trade is not just long memory versus short semis — it is long scarce components versus the broader AI capex stack.

The near-term catalyst path is mostly flow-driven: ETF demand and retail momentum can persist for weeks, especially if the fund splits and widens accessibility. The structural risk is 6-18 months out, when high pricing induces incremental supply, new capacity comes onstream, and the same scarcity narrative turns into margin normalization. What breaks the thesis is not a single soft quarter, but evidence that HBM lead times are shortening or that MU/SSNLF guide capex and bit-growth materially higher than demand, which would cap the scarcity premium.

Contrarian view: consensus is treating this as a clean AI winner, but memory is usually the first place cycle excess gets exposed. The move may still have room technically, but the asymmetry is worse for late buyers than for existing holders because the fund is highly concentrated and the underlying names are already discounting an aggressive earnings trajectory.

Key Decisions for Investors

• Long MU on pullbacks over the next 2-4 weeks; best risk/reward is if the market is overreacting to near-term AI demand and underestimating HBM pricing power. Falsify if MU guide mentions flattening DRAM/HBM margins or slower-than-expected capacity conversion.
• Pair trade: long MU / short a broad semis proxy (SOXX or SMH) for 1-3 months. Thesis is that memory outperforms the rest of semis as the market rewards scarcity over general AI exposure; trim if SOXX re-accelerates on foundry/logic upside.
• Long SSNLF as a lagging beneficiary if Korean memory catches the same flow-driven rerating. Use it as a lower-beta way to express the same theme; invalidated if Korean won strength or capex commentary suggests supply is ramping faster than expected.
• Avoid chasing SNDK after strength; treat it as a higher-volatility trading vehicle rather than the cleanest HBM expression. Better only on a 10-15% pullback, since the bullish case is more fragile if NAND is the main driver rather than HBM scarcity.
• Watch NVDA into earnings as a hedge, not a core long: if memory costs keep rising faster than pricing, the market may start haircutting gross margin assumptions. A modest short-dated call spread is only attractive if implied vol is cheap and memory data keeps tightening.