DRAM
Long the Roundhill Memory ETF (DRAM) on dips because the memory cycle has room to run, driven by a supply crisis expected to last until 2028 and upcoming catalysts like the SK Hynix IPO.
Near-term caution on AI memory semis points to downside for the memory chip basket
Memory ASP deflation is coming, not the Korean capex boom the bulls expect, pressuring memory maker revenues
Real-time AI companions need massive memory bandwidth; Wan Streamer proves the demand wave is real, lifting memory chip stocks
Buying the Roundhill Memory ETF (DRAM) to front-run a soft inflation print that could unlock a memory/semi rally.
Memory ETF rallies as Micron's blowout quarter confirms DRAM supercycle; author sizing in before macro catalyst.
Author personally adds to DRAM calls; memory ETF is the primary alpha vehicle ahead of Friday's cool inflation print.
AI memory ETF DRAM is the author's all-in conviction play on the 2025-2030 AI stock supercycle.
Micron earnings catalyst sets up DRAM ETF for continuation; tight 2.2% stop defines cheap invalidation.
Cloud hyperscalers building custom AI ASICs (Google, Amazon, Microsoft) are expanding HBM4 demand beyond just Nvidia, lifting the entire memory chip sector.
CXMT's wafer ramp cannot close the DRAM supply gap, keeping the undersupply cycle intact through 2028 and lifting memory prices
CXMT's $55B 2026 DRAM revenues confirm AI memory supercycle powering the entire sector; DRAM ETF captures the basket
The AI capital cycle is a structural feedback loop locking capital into memory semiconductors regardless of rates or macro
Memory and semiconductors continue higher throughout Q3 on renewed ETF inflows and capital repositioning after the July 4th seasonal reset
All four Korea memory export categories (DRAM, NAND, HBM) are in simultaneous all-time-high pricing, confirming a broad memory supercycle that lifts the entire memory chip basket.
Local AI runs into a memory wall, not a compute wall, so on-device inference demand pulls DRAM/NAND higher.
AI's scarcity bottleneck is memory; as the constraint resolves, the crowded premium in memory names de-rates.
Cheaper-model shift raises total inference volume, and every token still runs on memory the bottleneck.
AI memory is sold out into 2028; the makers that own the scarce HBM capture the scarcity rent and re-rate.
Memory bandwidth, not GPU FLOPS, is the binding AI constraint, keeping HBM sold out and memory pricing tight.