Q&A Series: The memory ETF that surged around 150% — and the AI thesis behind it

Artificial intelligence has emerged as one of the dominant drivers of technology investment in recent years, fuelling an unprecedented wave of spending on data centres and computing infrastructure. Alongside GPUs and semiconductor designers, memory chips have become one of the market's key areas of focus, reflecting their increasingly important role in the AI stack. Once viewed as a cyclical commodity business, memory semiconductors are increasingly being recognised as a strategic bottleneck, with demand for high-bandwidth memory (HBM) and other advanced memory technologies supported by a more durable backdrop than traditional PC and smartphone cycles.

For investors seeking to capitalise on this trend, the available options have historically been imperfect. Broad semiconductor ETFs provide memory exposure only as a secondary allocation, with most assets concentrated in chip designers, foundries, and equipment makers. Against this backdrop, Roundhill Investments launched the Roundhill Memory ETF (BATS: DRAM) on 2 April 2026, the first-ever ETF focused exclusively on the global memory industry. In this article, we share insights from Roundhill Investments on the memory industry's outlook, the fund's construction, and where it fits within the ongoing AI infrastructure buildout.

Introducing the Roundhill Memory ETF (DRAM)

The Roundhill Memory ETF (BATS: DRAM) is an actively managed fund that provides targeted exposure to companies operating across the memory semiconductor industry. The fund invests in businesses deriving at least 50% of their revenues or profits from the development or manufacture of memory products, including high-bandwidth memory (HBM), DRAM, NAND flash, NOR flash, hard disk drives, and specialty memory. With 18 holdings, the portfolio is anchored by industry leaders SK Hynix, Samsung Electronics, and Micron Technology, which together currently account for around 70% of assets.

Table 1: Key information about the ETF

Figure 1: DRAM ETF exposure concentrated in leading memory manufacturers

1. Roundhill has described memory as a ‘secular growth story tied to the multi-decade buildout of AI infrastructure.’ Given that memory has historically been one of the most cyclical sectors in tech, what gives you the conviction that this time represents a genuine structural shift rather than a particularly strong upcycle?

The scepticism is warranted as memory has a well-earned reputation for boom and bust. The Windows 95 cycle in the mid-1990s, the iPhone upgrade cycle in the mid-2010s, both followed the same playbook: demand spikes, manufacturers overbuild, prices crater. That history is real and we are not dismissing it.

What has changed is the nature of the demand driver. Prior cycles were tied to consumer device upgrades, which are episodic by nature. A new phone needs more memory, the cycle runs its course, manufacturers are left holding excess capacity. AI infrastructure demand does not work that way. Once a system is deployed, it requires ongoing inference, retraining, monitoring, and data retention. The demand becomes recurring rather than episodic, and it does not have a natural endpoint the way a consumer upgrade cycle does.

The supply picture reinforces the structural argument. Even the companies that dominate HBM production acknowledge the shortage runs through 2030, and the math explains why. Building a new fab takes at least three years. A greenfield site takes more than five. The equipment required to produce leading-edge memory is itself constrained, with ASML entering 2026 carrying a backlog larger than its entire projected annual sales. Capital commitment alone cannot solve a problem that is fundamentally about time. 

2. Could you explain DRAM’s index methodology, including how companies are selected for inclusion and what criteria would exclude a company from being classified as a “memory company”?

DRAM is actively managed rather than index-based, though its investment framework is rigorous and consistently applied. The defining criterion is revenue derivation. A company qualifies as a memory company if at least 50% of its revenues or profits are attributable to the development or manufacturing of semiconductor memory products, spanning high bandwidth memory, DRAM, NAND flash and solid-state drives, NOR flash, hard disk drives, and specialty and embedded memory.

For the Nvidia example, the issue is that Nvidia is so widely recognized as a dominant AI company that calling it a non-qualifier may confuse readers who associate it closely with the memory trade. A cleaner example might be:

The 50% threshold is the mechanism that produces genuine pure-play exposure. Taiwan Semiconductor Manufacturing Company is a useful illustration. TSMC fabricates memory chips on behalf of other companies but generates its revenues from foundry services rather than memory products. It is infrastructure for the memory industry, not a participant in it, and it would not qualify. The same logic applies to a diversified semiconductor company with a meaningful memory segment alongside logic, analog, or power businesses. If memory does not constitute at least half the business, the company is excluded. The standard ensures that every holding is genuinely in the memory business rather than adjacent to it.

The active management structure allows Roundhill to make timely judgments about emerging memory categories and new entrants as the industry evolves. Holdings that migrate below the revenue threshold are removed at quarterly rebalancing, which keeps the portfolio aligned with the investment mandate over time.

3. How does DRAM differ from a broad semiconductor ETF such as VanEck Semiconductor ETF (SMH)? What exposure or advantages does DRAM offer that investors may not obtain through a more established semiconductor fund?

Broad semiconductor ETFs are well-constructed products built to deliver chip sector exposure. They were not designed to deliver precise memory exposure, and that distinction matters significantly for investors who believe memory is the critical bottleneck in AI infrastructure.

The easiest way to illustrate the difference is through holdings. A broad semiconductor ETF would hold Nvidia, Broadcom, TSMC, and Qualcomm, but none of them are memory companies. When SK Hynix reports a record quarter because HBM supply is fully committed and pricing is elevated, that dynamic does not flow through to a broad semiconductor ETF in any material way. 

The global dimension is equally important and easy to underestimate. Samsung and SK Hynix are Korean-listed companies that dominate global HBM supply. There is another way for obtaining concentrated, targeted exposure to them through a U.S. semiconductor ETF. DRAM was purpose-built to solve exactly that problem.

4. DRAM launched on 2 April 2026 and surpassed USD 1 billion in assets within its first 10 trading days. What has driven this strong demand, and has it come primarily from retail investors, institutions, or both? Does such rapid asset growth create any portfolio management challenges?

The demand reflected a market that had already formed a well-developed thesis and was searching for the right vehicle to act on it. Memory stocks had been among the strongest performing areas of the market for over a year. Analysts across the sell side had been upgrading the space broadly. 

The interest came from both retail and institutional allocators, which reflects where the memory thesis sits in terms of accessibility and analytical depth. The fundamental argument is intuitive enough for individual investors to engage with directly while being rigorous enough for institutional portfolio managers to construct a position around with confidence.

5. DRAM’s 52-week trading range spans from USD 26.14 to USD 70.15, reflecting significant price volatility. How should retail investors approach entry timing when considering an investment in the fund?

The volatility in DRAM reflects the volatility of the underlying memory stocks, and that is inseparable from the thesis. Memory stocks reprice dramatically as AI infrastructure spending expectations shift, earnings revisions arrive in large increments, and supply and demand signals update in real time. That volatility is not incidental to the opportunity. The magnitude of the potential return comes directly from the magnitude of the repricing underway, and investors who want the former need to be genuinely comfortable with the latter.

On valuation, the picture is more nuanced than the price range alone suggests. Despite the magnitude of recent stock moves, the median next-twelve-month price to earnings ratio across DRAM holdings sits at just 8.37x, a compelling figure relative to broader technology peers. The median estimated earnings per share growth for the current fiscal year across the portfolio is 632%. Those numbers do not describe a universe that has already fully priced its opportunity.

6. SK Hynix, Samsung, and Micron account for roughly 73% of the portfolio. Should investors view DRAM largely as a concentrated exposure to these three companies, and how does the fund manage concentration risk? Is this weighting likely to remain at current levels?

The concentration reflects the structure of the industry rather than a portfolio construction choice. The global HBM market is controlled by three companies. An ETF truly committed to pure-play memory exposure will naturally be heavily weighted in those three names because the industry itself is structured that way. Spreading the weight artificially to appear more diversified would simply mean diluting the precision that makes the fund useful in the first place.

DRAM is not a three-stock product, however. The remainder of the portfolio provides meaningful exposure across the full memory stack. SanDisk and Kioxia anchor the NAND and enterprise SSD segment. Western Digital and Seagate cover the storage layer, where revenues track hyperscale data growth more than short-term chip pricing dynamics. Specialty and embedded memory names including GigaDevice, Winbond, and Nanya add geographic and product diversification within the memory complex across end markets with different demand characteristics.

The concentration in the top three is likely to remain elevated given that the industry structure is not changing materially in the near term. The more interesting question over time is how the relative weighting among Samsung, SK Hynix, and Micron shifts as HBM market share, NAND dynamics, and individual company earnings trajectories develop.

7. The fund gains exposure to Micron through total return swaps rather than direct share ownership. Can you explain the rationale behind this structure and what it means for investors in terms of portfolio exposure and investment risk?

Total return swaps are used selectively alongside those direct holdings and serve a specific structural purpose: they provide the flexibility to manage regulatory obligations without compromising the intended exposure.

The reason swaps come into play is that the fund must maintain compliance with Regulated Investment Company diversification tests under US tax law. In a portfolio this concentrated, a handful of positions can approach those thresholds quickly. The ability to hold a combination of direct shares and swap exposure across certain names gives the portfolio the tools to remain within the required parameters while preserving the full economic participation an investor is seeking. The return profile from the investor’s perspective is equivalent whether exposure is held directly or through a swap.

8. If AI infrastructure spending slows due to regulatory pressures, weaker hyperscaler capital expenditure, or a broader technology downturn, what is the key bear case for DRAM? Are there any holdings or segments within the portfolio that may prove more resilient under such a scenario?

The most straightforward risk is a pullback in AI capital expenditure. If hyperscalers materially reduce data centre spending, HBM demand softens, pricing follows, and the earnings estimates currently underpinning these stock prices get revised sharply lower. The top three holdings represent nearly three quarters of the fund and are the most direct beneficiaries of the AI memory thesis. They would bear the brunt of that scenario.

The more nuanced and arguably more credible bear case is a structural shift in how AI generates revenue. The current infrastructure buildout is premised on a set of assumptions about AI monetization that have not yet been fully tested at scale. If the economics of deploying AI prove more difficult than anticipated, if enterprise adoption is slower, if competition compresses pricing on AI services faster than expected, or if hyperscalers begin questioning the return on the capital they are deploying, the investment cycle could decelerate materially without any single dramatic catalyst. That kind of gradual reassessment of growth assumptions can be just as damaging to highly valued stocks as an outright demand shock, and it is harder to see coming.