KEY POINTS
- The Roundhill Memory ETF (DRAM) has returned approximately 107% since its April 2 launch, making it the best-performing thematic ETF of 2026, fueled by explosive demand for high-bandwidth memory chips in AI data centers.
- The fund has pulled in an estimated $8 billion in assets in roughly 60 trading days, with top holdings Samsung Electronics, SK Hynix, and Micron all benefiting from the AI-driven HBM supply shortage.
- Watch whether DRAM can sustain momentum as memory chip pricing cycles historically mean-revert, and whether the fund's concentrated exposure to three dominant memory makers creates outsized volatility risk.
The Roundhill Memory ETF has doubled in price since it began trading on April 2. In a year filled with strong thematic fund launches, DRAM stands alone: a 107% return in roughly two months, driven by a supply-demand imbalance in semiconductor memory that shows no sign of resolving.
The fund invests in companies deriving at least 50% of their revenue from semiconductor memory products — high-bandwidth memory, DRAM technology, NAND flash, and related manufacturing. In practice, that means heavy concentration in three names: Samsung Electronics, SK Hynix, and Micron Technology. These three firms control more than 95% of the global DRAM market and virtually all HBM production, making the ETF less a diversified basket and more a leveraged bet on the memory oligopoly.
The HBM Supercycle
The thesis behind DRAM is straightforward: every AI GPU needs HBM, HBM supply is constrained, and memory companies are printing money. Nvidia's H100, H200, and Blackwell accelerators all require multiple stacks of high-bandwidth memory to function. Each generation of GPU uses more HBM than the last. AMD's MI300X and MI400 chips follow the same pattern. The custom AI accelerators that Broadcom builds for hyperscalers also consume HBM at scale.
SK Hynix has been the primary beneficiary, having secured the majority of Nvidia's HBM3E orders. The stock has roughly tripled over the past year. Micron, which recently developed the HBM3E line specifically for industrial AI applications, has gained over 1,000% in the past twelve months — a staggering move for a company with a $200 billion market capitalization. Samsung, the largest of the three, has lagged on HBM yield rates but is catching up, and its diversified semiconductor and electronics business provides a lower-beta entry point into the same trade.
The demand picture keeps tightening. Every major cloud provider — Amazon, Google, Meta, Microsoft, Oracle — is building AI data centers at a pace that exceeds even the most aggressive forecasts from two years ago. France alone has announced over €110 billion in AI infrastructure investment. Each of those data centers needs thousands of GPUs, and each GPU needs gigabytes of HBM. The math is simple: demand is growing faster than the memory industry can add capacity.
The Flow Machine
DRAM's asset growth has been as remarkable as its price performance. The fund reportedly surpassed $6.5 billion in AUM within its first 36 trading days, a pace that would make it one of the fastest-growing ETFs ever launched if the figures hold up under independent verification. By late May, inflows had pushed total assets toward $8 billion.
The speed of accumulation reflects the difficulty investors face in getting pure-play exposure to the HBM theme through individual stocks. Buying SK Hynix requires a Korean brokerage account. Samsung trades on the Korea Exchange with complex local tax implications. Micron is the only name easily accessible to U.S. investors, but it also carries NAND flash exposure that dilutes the HBM thesis. DRAM solves the access problem by wrapping all three — plus smaller memory-adjacent names — into a single U.S.-listed vehicle.
The fund also benefits from timing. It launched at the precise moment when AI infrastructure spending shifted from "how much GPU compute can we buy?" to "can we get enough HBM to feed the GPUs we already ordered?" That supply-chain bottleneck elevated memory stocks from supporting players to lead actors in the AI narrative, and DRAM gave investors a ticker to express that view.
The Risk Nobody Wants to Discuss
Memory semiconductors are a cyclical industry with a long history of boom-bust pricing. HBM is currently in a supercycle driven by AI demand, but previous DRAM supercycles — in 2017-2018 and 2020-2021 — ended when supply caught up with demand and pricing collapsed 40-60% from peak to trough. The current cycle may be different in duration because AI demand is more structural than the PC and smartphone demand that drove prior cycles, but "this time is different" is always a dangerous phrase in commodity semiconductors.
The concentration risk is also significant. A fund where three companies represent the vast majority of assets is not really an ETF in the diversification sense — it is a basket trade with management fees. If Samsung reports a yield problem with its HBM3E production, or if SK Hynix faces a geopolitical disruption at its Icheon fab complex, the impact on DRAM would be outsized relative to broader semiconductor indices.
For traders who believe the HBM supercycle has room to run, DRAM offers efficient exposure. For those worried about cyclical mean-reversion, the fund's vertical concentration makes it the wrong place to be when the turn comes. The key signal to watch is memory pricing data from TrendForce and DRAMeXchange: as long as contract prices for HBM3E continue rising quarter over quarter, the thesis holds. The moment those prices flatten, the 107% return starts looking less like a floor and more like a ceiling.
