KEY POINTS
- The Roundhill Memory ETF (DRAM) reached nearly $10 billion in net assets just seven weeks after its April 2 launch, making it the fastest-growing ETF in history by asset accumulation.
- DRAM has returned approximately 90% since inception, driven by an AI-fueled memory chip shortage that has pushed Samsung, SK Hynix, and Micron shares sharply higher.
- Morningstar flagged valuation concerns, warning that DRAM's top holdings look overvalued — traders should monitor whether the AI memory supercycle can sustain forward earnings growth to justify current multiples.
The Roundhill Memory ETF crossed the $10 billion asset threshold in late May, just seven weeks after launching on April 2, obliterating every previous record for ETF asset gathering and doing so by a margin that is not particularly close. The fund attracted $1 billion in its first 10 trading days and $5 billion within 25 days, a pace that exceeds the early trajectories of the most successful ETF launches in history, including the original spot Bitcoin ETFs that debuted in January 2024.
DRAM has returned approximately 90% since inception, surging to roughly $52.82 per share by May 25. The fund now ranks among the top 20 most-traded ETFs by daily volume, averaging about 29 million shares per day, up from 34th position at the start of May. For context, that volume rivals established semiconductor ETFs like the VanEck Semiconductor ETF (SMH) and the iShares Semiconductor ETF (SOXX), which have existed for decades.
The Supply Crunch Thesis
DRAM's explosive growth reflects a simple but powerful thesis: the AI infrastructure buildout is consuming memory chips faster than the industry can produce them, creating a supply-demand imbalance that is driving prices and margins higher for the three companies that dominate the market.
Samsung, SK Hynix, and Micron collectively represent more than 70% of DRAM's portfolio, a concentration level that would normally raise diversification concerns but in this case reflects the reality of an industry with only three major players. High-bandwidth memory, the specialized DRAM chips used in AI accelerators like Nvidia's H100 and B200, is the tightest component in the entire AI supply chain. Every new GPU shipped requires multiple stacks of HBM, and production capacity for the latest HBM3E specification is essentially sold out through 2027.
The math is straightforward. Nvidia shipped approximately $75 billion worth of data center GPUs in Q1 alone, and each of those GPUs requires between $500 and $1,500 worth of HBM depending on configuration. That translates to tens of billions of dollars in annual memory demand from a single customer, before accounting for AMD, Google, Amazon, and other AI silicon producers who are ramping their own accelerator programs.
SK Hynix has been the primary beneficiary, holding an estimated 50% share of the HBM market and commanding premium pricing that has expanded gross margins above 50% for the first time in the company's history. Micron, which was late to HBM but has ramped aggressively, is now selling every wafer it can produce at prices that analysts estimate are 3x to 4x higher than conventional DRAM.
The Valuation Debate
Not everyone is convinced that DRAM's momentum is sustainable. Morningstar published an analysis warning that the fund's top holdings look overvalued relative to their long-term earnings power, noting that memory semiconductors have historically been brutally cyclical. Previous memory "supercycles" in 2017-2018 and 2020-2021 both ended with sharp corrections when supply caught up with demand and pricing collapsed.
The bull response is that this cycle is structurally different because HBM production requires specialized equipment and years of process development that cannot be replicated quickly. Unlike conventional DRAM, where Samsung, SK Hynix, and Micron can add capacity within 12 to 18 months by converting existing fabrication lines, HBM requires entirely new manufacturing processes that take three to four years to scale from development to mass production.
The counter-argument is that demand forecasts for AI infrastructure are themselves uncertain. If hyperscaler capex growth decelerates in 2027, as some analysts predict, the memory shortage could ease faster than the market expects, and DRAM's concentrated portfolio would amplify the downside just as effectively as it amplified the upside.
Positioning and Risk Management
For traders considering DRAM, the concentration risk cuts both ways. The fund offers pure-play exposure to the AI memory theme without the dilution that comes from broader semiconductor ETFs, which include companies with minimal AI exposure. But that purity also means there is no diversification buffer if the memory cycle turns.
The technical picture shows DRAM in a clear uptrend with no meaningful pullback since launch, a pattern that historically precedes either a continuation move higher or a sharp correction when profit-taking begins. The fund's average daily volume provides sufficient liquidity for most position sizes, but spreads can widen during volatile sessions given the relatively small number of underlying stocks.
Watch Micron's earnings in late June and SK Hynix's Q2 results in July for the next fundamental catalysts. If HBM pricing holds or increases and forward guidance projects continued supply constraints through 2027, DRAM's rally has room to extend. If either company signals inventory normalization or pricing pressure, the unwind could be swift given the fund's concentrated holdings and hot-money flow profile.
