The Memory AI Trade

Episode Summary

Executive Summary: The episode examines the explosive rise of AI-related memory chips and the Roundhill DRAM ETF, explaining why high-bandwidth memory is becoming a critical bottleneck and investment theme. It compares pure-play memory suppliers, explores supply constraints and fab build timelines, and extends the trade to semiconductor equipment makers as a secondary way to play the AI infrastructure buildout.

Main Topics: The rise of the DRAM memory ETF (Priority: 5/5): Hosts discuss Roundhill’s DRAM ETF, which rapidly attracted enormous trading volume and assets because it offers direct exposure to the main memory suppliers benefiting from AI demand. Why high-bandwidth memory matters for AI (Priority: 5/5): Jake Silverman explains that AI models and data centers need much more memory, especially HBM, because larger context windows and model parameters require storage and fast access in limited physical space. Company-by-company comparison of the memory leaders (Priority: 4/5): The conversation contrasts Samsung, SK Hynix, and Micron, emphasizing that Samsung is diversified, while SK Hynix has led HBM and Micron has been catching up. Memory as a commodity and the cycle risk (Priority: 5/5): The panel frames memory as a highly cyclical, commodity-like industry where prices can rise sharply during supply shortages but eventually face boom-bust dynamics. Supply constraints and new fab construction (Priority: 4/5): They discuss how shifting capacity toward HBM and building new fabs takes years, which supports current pricing power but also creates future oversupply risk. Second-order trade: semiconductor equipment (Priority: 3/5): The episode identifies chip equipment makers such as ASML, Lam Research, Applied Materials, and KLA as another way to benefit from memory capacity expansion.

Key Arguments: The DRAM ETF succeeded because it bundled the most direct memory exposure available to U.S. investors, especially since key Korean names lack ADRs. AI growth is driving a structural increase in memory demand because models require more storage for parameters, context windows, and inference-related cache. SK Hynix has been the leader in high-bandwidth memory, with Micron improving but still behind; Samsung is less of a pure-play because it contains multiple businesses. Memory remains commodity-like, so current pricing strength is driven by a supply-demand imbalance rather than a permanent re-rating. HBM uses significantly more wafer area than traditional DRAM, reducing supply for standard memory products and tightening the market further. New fabs take years to build and equip, so near-term supply is constrained even as companies announce expansions. The memory trade extends to equipment suppliers, but their revenue exposure to memory is indirect because many sell more heavily to TSMC, Intel, and Samsung’s foundry business rather than memory alone. Broad market ETFs like SPY largely miss this trade today, which is why thematic ETFs can provide earlier exposure to a fast-growing niche.

Data Points: DRAM ETF trading volume: $200 million on day one - Used to illustrate the extraordinary launch demand for Roundhill’s memory ETF. DRAM ETF assets: Over $1 billion - The ETF quickly crossed the billion-dollar threshold shortly after launch. Time to $1 billion for DRAM: 10 days - Shown as a major outlier versus other thematic ETFs. Time for top four thematic ETFs to reach $1 billion: Over 1,000 days to over 3,000 days - Compared with DRAM’s speed, highlighting the launch as exceptional. Memory industry size: $75 billion - Current scale of the memory market discussed in the episode. Projected memory industry size: $135 billion in two years - Described as the expected doubling of the market. HBM wafer space requirement: 3 to 4 times more space per wafer - Explains why HBM consumes more capacity than traditional memory products. Cycle duration: 7 to 10 quarters - Historical length of memory upcycles referenced by the analyst. Expected greenfield timing: Second half of 2027 to 2028 - When new fabs could begin affecting supply materially. Micron expansion: Boise and New York - Examples of announced capacity expansion by Micron. Pacer Blue Star Engineering the Future ETF assets: $5 million - The alternative equipment-focused ETF had limited assets despite relevant holdings. Pacer Blue Star Engineering the Future ETF fee: 49 basis points - Reported as the ETF’s expense ratio. BULD performance: Up 64% since launch - Given as the return since the ETF’s inception four years earlier. DRAM ETF portfolio concentration: 67% in three stocks - SK Hynix, Samsung, and Micron together dominate the portfolio. SMH Micron weight: About 5% - Used to show that widely used semiconductor ETFs have limited memory exposure.

Pivotal Quotes: "It's almost like downstream from AI." — Eric Beltranas: Describing memory chips and their role in the broader AI investment stack. "Memory is usually considered the boring aspect of semiconductors, but it's gotten a lot more exciting." — Jake Silverman: Summarizing why memory has become a major AI investing theme. "So basically, you need memory to interact with the model." — Jake Silverman: Explaining the technical reason memory demand rises with AI usage and larger context windows.

Implications: The episode suggests AI infrastructure demand is broadening beyond GPUs into memory and equipment. Investors looking for pure exposure may prefer niche thematic ETFs, but the trade is cyclical and could reverse if new fabs and Chinese capacity create oversupply.