The Roundhill Memory ETF hit a 52-week high of $68.76 on Monday after Jensen Huang confirmed at his GTC Taipei keynote that Nvidia's Vera Rubin AI platform had entered full production, with Samsung Electronics, SK Hynix, and Micron named as its three HBM4 memory suppliers. The fund, which launched on April 2, 2026 at $28 per share, has more than doubled in two months as investors bet that the memory shortage constraining AI data centers is structural — not a temporary spike that will ease when chipmakers add capacity.
The timing of Monday's record underscores how tightly the ETF is coupled to its three largest holdings. When Huang confirmed that Vera Rubin required high-bandwidth memory from all three major producers, shares in Samsung, SK Hynix, and Micron moved immediately. The DRAM ETF, which concentrates roughly 73% of its portfolio in those same three names, followed.
Fastest ETF Launch in History Tops $10 Billion
The fund made ETF history before its Vera Rubin catalyst arrived. By May 15, the Roundhill Memory ETF had crossed $10 billion in assets under management in 50 days — the fastest any ETF had ever reached that milestone — eclipsing the previous record held by BlackRock's iShares Bitcoin Trust, which took 43 days. Bloomberg ETF analyst Eric Balchunas tracked the comparison and confirmed the milestone. The fund reached $6.5 billion in AUM in approximately 36 trading days, itself a record at the time.
Roundhill CEO Dave Mazza told CNBC on May 15 that the surge reflected a structural shift in how investors think about AI. "Investors are waking up to the fact that the biggest bottleneck in the AI build-out is actually memory chips," he said. The fund requires companies to derive at least 50% of their revenue from semiconductor memory products to qualify for inclusion, creating a filter that excludes the diversified chip designers that dominate broader semiconductor indexes.
What the DRAM ETF Actually Holds
The portfolio is built around nine positions. SK Hynix accounts for roughly 24–26% of net assets, Samsung Electronics for approximately 20–25%, and Micron Technology for approximately 24–27% — figures that shift at each quarterly rebalance. Kioxia, Sandisk (SNDK), Seagate Technology, Western Digital, Nanya, and Winbond make up the remainder. South Korea accounts for approximately 49% of geographic exposure and the United States for 38%, with Japan and Taiwan splitting the balance.
For U.S. investors, the access case is straightforward: neither Samsung nor SK Hynix offers American Depositary Receipts, making them difficult or costly to hold directly through most domestic brokerages. The DRAM ETF packages them into a single, liquid, U.S.-exchange-traded ticker. SK Hynix has filed confidentially with the SEC for a potential U.S. ADR listing, according to reporting from TechTimes, but that process has not yet been completed.
The underlying demand thesis rests on how AI data centers consume memory differently from conventional servers. Standard server racks need a baseline of DRAM to function. AI training clusters, by contrast, are built around GPU accelerators that require high-bandwidth memory stacked directly alongside the compute die. Each new Nvidia GPU generation demands more HBM per chip — Nvidia's B300 GPU contains eight HBM chips, each comprising 12 individual DRAM dies. That compounds demand as each new accelerator generation arrives, while the supply of HBM is constrained by the same wafer capacity that Samsung, SK Hynix, and Micron are already running at effective maximum utilization.
AI Memory Chip Shortage Explains the Rally
Micron CEO Sanjay Mehrotra said during earnings calls that HBM capacity for 2026 was fully booked before the year began, and that the company was meeting only 50% to 65% of what key customers were requesting. SK Hynix's management made similar disclosures — its entire 2026 production across DRAM, NAND, and HBM was described as essentially sold out by late 2025.
Data center consumption is the driver. Data center demand for DRAM reached roughly 50% of global consumption in 2025, up from 32% five years earlier, according to Bloomberg Intelligence. That share is expected to climb further, and the production math makes relief unlikely in the near term: one gigabyte of HBM requires roughly three times the wafer capacity of standard DDR5 DRAM to manufacture, so every AI chip produced effectively displaces three units of conventional memory output.
At Computex on Tuesday morning, SK Hynix chairman Chey Tae-won announced that the company would double its total wafer production capacity over the next five years — a significant reversal from March, when Chey had told reporters that adding capacity was not something the company had planned. Hours later, Jensen Huang visited the SK Hynix booth, picked up a marker, and wrote "Please Make More" on an HBM4E wafer on display. The two messages, from the supplier's chairman and from its most important customer, landed at the same trade show on the same afternoon.
What Are the Risks of the DRAM ETF?
The fund's structure introduces risks that a retail investor buying a basket of tech stocks might not anticipate.
The most significant is concentration. With roughly 73% of assets in three names, the fund is highly exposed to any single bad earnings report, supply-chain disruption, or demand reversal at Samsung, SK Hynix, or Micron. Morningstar equity analysts flagged every major memory stock in its covered universe as overvalued as of mid-May 2026, noting that none of the three possesses a durable economic moat and that the commodity-like nature of the memory business historically produces violent boom-bust cycles. William de Gale, portfolio manager at BlueBox Asset Management, called memory "a pretty dreadful industry" in the long run during a CNBC appearance. Morningstar researchers also warned that thematic ETF investors often chase momentum near peaks, particularly during speculative technology cycles.
The second structural risk involves how the ETF achieves its exposure. Roundhill uses total return swaps on some of its holdings in order to maintain compliance with Regulated Investment Company diversification rules for tax purposes — specifically to hold concentrated positions in non-U.S. equities that cannot be owned through depositary receipts. Swap agreements can amplify losses during corrections, introduce counterparty risk, and may behave differently from direct equity ownership during periods of market stress. The fund's SEC filings describe this explicitly as derivatives risk, a category that investors in conventional index ETFs do not typically encounter.
A third risk is cyclicality. Memory is one of the semiconductor industry's most historically cyclical segments. A supply glut — triggered by chipmakers expanding capacity faster than demand grows, or by AI data-center spending plateauing — can reverse pricing and margins quickly. Steve Brice, global chief investment officer at Standard Chartered, warned investors in May to take profits on Korean equities and diversify globally, calling peak optimism on the AI memory trade "not too far around the corner."
Wall Street noticed the ETF's success in a different way: Themes ETF Trust filed with the SEC to launch the Leverage Shares 2X Long Memory Daily ETF, which would double the daily performance of DRAM. The SEC filing itself warns that returns over periods longer than a single day may differ significantly from doubling the fund's cumulative performance due to compounding effects and volatility decay — an additional layer of risk on top of an already concentrated underlying position.
How Does High-Bandwidth Memory Fuel AI Chip Demand?
High-bandwidth memory is a chip architecture that stacks multiple DRAM dies vertically and connects them with microscopic copper pathways, bonding the stack directly alongside a GPU inside an AI server. Where standard server memory delivers data at roughly 50 to 100 gigabytes per second, HBM3E — the current production generation — delivers up to 1.2 terabytes per second per stack. That speed advantage is why Nvidia's AI accelerators require HBM to run efficiently: without it, even the fastest GPU is bottlenecked waiting for data to arrive.
Manufacturing HBM at commercial scale requires hundreds of processing steps and advanced packaging technology. As documented in prior TechTimes coverage of the memory shortage, only three companies have cleared that barrier: SK Hynix held 57% of global HBM revenue as of Q3 2025, according to Counterpoint Research, with Samsung at 22% and Micron at 21%. That oligopoly structure is what gives all three companies unusual pricing power — and what makes the DRAM ETF a concentrated bet on three linked companies rather than a diversified play on AI infrastructure broadly.
Frequently Asked Questions
What is the Roundhill Memory ETF and what does it hold?
The Roundhill Memory ETF (ticker: DRAM) is the first U.S.-listed ETF focused exclusively on memory chip manufacturers. It holds equity positions and total return swaps in companies that derive at least 50% of revenue from semiconductor memory — primarily SK Hynix, Samsung Electronics, and Micron Technology, which together account for roughly 73% of the portfolio. The remainder includes NAND flash producers Kioxia and Sandisk, and storage specialists Seagate and Western Digital.
Is the DRAM ETF a good investment for retail investors?
It depends on the investor's risk tolerance and portfolio context. The fund offers clean access to global memory leaders — including SK Hynix and Samsung, which lack U.S. depositary receipts — and has returned more than 140% since its April 2, 2026 launch. However, its concentration in three names, its use of total return swaps, and the historical cyclicality of the memory industry make it a high-conviction, high-risk bet rather than a diversified AI infrastructure play. Morningstar analysts rated every major memory stock as overvalued as of mid-May. Multiple analysts recommend treating DRAM as a small portfolio allocation rather than a core holding.
How risky is investing in the DRAM ETF compared to broader semiconductor ETFs?
DRAM is significantly more concentrated than broad semiconductor ETFs such as the VanEck Semiconductor ETF (SMH), which holds dozens of companies across chip design, manufacturing, and materials. DRAM holds nine positions, with nearly three-quarters of assets in three names. It also uses derivative instruments — total return swaps — that broad-market ETFs do not. This structure means a negative development at any single holding can move the fund sharply.
Why did the DRAM ETF grow faster than any other ETF in history?
The fund launched at a moment when AI infrastructure investment was accelerating dramatically, and when no U.S.-listed vehicle existed for investors who wanted targeted exposure to memory manufacturers rather than GPU designers. Its record growth reflects both institutional demand for the AI memory thesis and retail enthusiasm for a fund that packages hard-to-access Korean chipmakers into a familiar U.S. ticker.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
