Author: Deep Tide TechFlow

When a trade is crowded enough to require a separate ETF just for retail investors, the smart money has often already begun to sell.

On April 2, Roundhill Investments officially launched the world's first pure memory semiconductor ETF, code $DRAM, named directly after memory modules. The closing price on the first day was $27.76, which later rose 5% to $29.15 in after-hours trading.

It seemed very lively. But just a few hours later, BTIG released a cold research report: The launch of the DRAM ETF is precisely a sell signal for memory stocks.

Don't rush to call this sensational; this Wall Street iron law has been repeatedly validated.

The "Ingredient List" of an ETF: Three Giants Consume Three-Quarters

First, let's see what $DRAM has actually bought.

This ETF currently holds only 9 stocks, extremely concentrated. Micron Technology, Samsung Electronics, and SK Hynix each average about 25% weight, totaling nearly three-quarters of the entire fund's position. The remaining scraps are allocated to Kioxia, SanDisk, Western Digital, Seagate, and other storage companies.

With a fee rate of 0.65%, it's not cheap. There are currently no options trading. To meet the diversification requirements of RIC (Regulated Investment Company), the fund has to use Total Return Swaps to "meet the numbers," in plain terms, the holdings are too concentrated, relying on derivatives to pass compliance.

The statement from Roundhill CEO Dave Mazza is very direct: "Memory is becoming the core of the AI ecosystem." This statement has merit. HBM (High Bandwidth Memory) is indeed one of the most critical bottlenecks in today's AI infrastructure; SK Hynix's HBM market share exceeds 60%, Micron's HBM capacity is already sold out until the end of 2026, and Samsung is also catching up vigorously.

The product logic is not wrong; the problem lies in the timing.

Roundhill's "Kiss of Death": A Precisely Timed Contrarian Indicator History

BTIG brought out the history of Roundhill’s own products, and the picture is quite grim.

The classic example is Roundhill MEME ETF. This fund tracking popular retail stocks first launched in December 2021, just at the absolute peak of the Meme stock bubble. After that, the UBS MEME Index plummeted by about 80%, and the fund was forced to liquidate in November 2023. Even more absurd, it was re-launched in October 2025, at which point Meme stocks had just rebounded by 100% from their lows. What happened? After relaunching, the index fell by about 40% again.

Twice launched, twice precisely hitting the peak. If you treated the launch dates of Roundhill's products as contrarian indicators to short, your returns would likely be higher than buying it.

This is not just a problem for Roundhill alone. BTIG pointed out a broader rule: The launch of thematic ETFs often marks the "consensus peak" of a certain trade.

In October 2021, ProShares launched the first U.S. Bitcoin futures ETF ($BITO), with a trading volume exceeding $1 billion on the first day, and the entire market cheered. A month later, Bitcoin peaked at $69,000 and then plummeted by 77%.

In November 2017, ProShares launched the EMTY ETF to short physical retail, which resulted in the physical retail index rebounding by 50% in the next 9 months.

In January 2008, VanEck launched the coal ETF (KOL), after which coal stocks entered a 12-year bear market, plummeting by 99%. KOL liquidated at its lowest point in December 2020, and after liquidation, coal stocks soared by 660%.

ETF launches often mark the peak, and ETF liquidations often mark the bottom. This pattern keeps repeating, and the underlying logic is quite simple: When a theme is so hot that ETF issuers believe "retail investors will buy," the market rally has often reached its end.

After a 350% Increase, Who is Swimming Naked?

The warning signals on the data front are very clear.

Goldman Sachs' TMT memory exposure index has soared by 350% over the past year, reaching a high of nearly 400% in February, just before the DRAM ETF arrived. Micron's stock price at one point deviated from the 200-day moving average by over 150%, a deviation greater than during the 2000 tech bubble, marking an extreme level not seen in Micron's history. BTIG pointed out that if Micron merely returned to the 200-day moving average, it would mean a decline of about 30% from the current level.

The frenzy in the entire memory sector is well-documented. EWY (iShares Korean ETF) has surged by about 140% over the past year, but breaking it down, 84 percentage points of the gain came from Samsung and SK Hynix stocks. This so-called "Korean ETF" has essentially become a substitute for a memory ETF, with Samsung accounting for about 27% and SK Hynix about 20%, together making up nearly half.

And this is exactly the demand that $DRAM aims to capture. Over the past year, EWY attracted $8.3 billion, with many investors buying the Korean ETF primarily to bet on memory. Roundhill accurately targeted this demand gap.

However, "precisely capturing demand" and "precisely hitting the top" are often only distinguishable in hindsight.

The "Other Side" of the Super Cycle

Fairly speaking, the bullish logic is also strong enough.

Bank of America defines 2026 as a "super cycle similar to the 1990s," predicting a 51% growth in global DRAM revenue and a 45% growth in NAND. Goldman Sachs estimates that the HBM market will reach $54.6 billion by 2026, representing a year-on-year growth of 58%. WSTS predicts that the global semiconductor market will grow by over 25% in 2026, approaching $975 billion.

Micron's fiscal year 2025 data center revenues surged by 137% to $20.7 billion, HBM capacity is completely sold out until 2026, and capital expenditure plans amount to $20 billion (a year-on-year increase of 45%). SK Hynix maintains over 50% market share in HBM3E and is the preferred supplier for Nvidia and Google custom chips.

These are real industry trends unrelated to speculation. The demand for memory from AI is structural; each generation of GPUs requires exponentially more HBM, with the H100 needing 80GB, and by the time of the GB300 NVL72 architecture, 17.3TB is required.

Therefore, the core contradiction is clear: the memory industry is certainly a good business, but is there still a good price for a good business?

A comparison: when BITO launched in October 2021, the long-term outlook for Bitcoin was correct, and BTC indeed reached new highs after the approval of the spot ETF in 2024. However, if you bought on the day BITO launched, you would first have to endure a 77% drawdown before waiting three years to break even.

The industrial trend may be right, but the trade can be wrong. Timing is everything.

Deep Tide's Perspective: It's Not the Bell Tolls, but It Definitely Rings Alarms

Our judgment: The launch of the DRAM ETF is not necessarily linked to a peak collapse in the memory industry, but it absolutely cannot be seen as a signal of "safe bet." It is more like an extremely precise thermometer of sentiment. When an industry is hot enough to require the issuance of a dedicated ETF to satisfy the appetite of retail investors, it at least indicates three things:

First, the Easy Money phase has ended. Most of the 350% increase in memory stocks over the past year is valuation expansion, rather than earning catch-up. Moving forward, memory stocks need to prove that the current prices are reasonable with actual performance growth, leaving very little room for error.

Second, the "thematic ETF trap" is a cause for high alert. Roundhill's historical performance is the best teaching material. When an investment theme is packaged as a zero-threshold retail product, it often means that institutions have already been reducing their positions while retail investors are picking up the shares. It’s too strong to call it a conspiracy theory; this is the natural ecology of the capital market, where the incentives of product issuers dictate that they will always chase trends and never anticipate turning points.

Third, the real risk lies in pricing; the industry fundamentals are not really a concern. Micron's 150% deviation from the 200-day moving average is more exaggerated than during the tech bubble period. Even if the demand for memory doubles, a technical correction of 30% is entirely within a reasonable range.

History does not simply repeat itself, but it always rhymes. After BITO launched, Bitcoin plummeted by 77%, and the MEME ETF nailed the peak twice; can $DRAM break this spell?

The only thing we can be certain of is one thing: When everyone thinks the same deal is "impossible to lose," that is the most dangerous time.