Memory chips are critical hardware components for systems that run artificial intelligence (AI) workloads, whether that processing is done in data centers, on personal computers, or on smartphones. Demand for memory currently exceeds what can be supplied by the manufacturers, and that imbalance has fueled a substantial increase in the value of the world's top three suppliers: Micron Technology, SK Hynix, and Samsung Electronics.
Those companies are the top three holdings in the Roundhill Memory ETF (DRAM +2.31%), an exchange-traded fund that launched in early April. In the short period of time since its debut, it has already delivered a return of more than 127% for investors, but I think it's time to tread with caution.
Image source: Getty Images.
Deploying AI wouldn't be possible without lots of memory chips
Graphics processing units (GPUs) are the primary chips used to handle the heavy data processing of AI training and inference workloads. High-bandwidth memory (HBM) stores information in a ready state for when the GPU needs it, which maximizes processing speeds. If the systems doing that work didn't have enough memory capacity near the processors, the result would be a bottleneck that forced GPUs to repeatedly pause while awaiting fresh data. That would create a laggy experience for anyone using AI chatbots or AI agents.
Therefore, it should come as no surprise that demand for HBM is exploding. But some AI workloads are gradually migrating to personal computers and smartphones, so manufacturers of those devices are requesting direct random access memory (DRAM) with significantly higher capacity for their products, too. This could drive the next wave of growth for the memory industry once the data center infrastructure build-out eventually slows down.
ETFs can hold hundreds or even thousands of individual stocks, but the Roundhill Memory ETF holds just 15. Micron is the largest position in the fund, but its top-five holdings alone account for 84.8% of the entire value of its portfolio.
|
Stock |
Roundhill ETF Portfolio Weighting |
|---|---|
|
1. Micron Technology |
28.2% |
|
2. SK Hynix |
27.1% |
|
3. Samsung Electronics |
17.9% |
|
4. Kioxia Holdings |
6.5% |
|
5. Sandisk |
5.1% |
Data source: Roundhill Investments. Portfolio weightings are accurate as of May 31, 2026, and are subject to change.
Micron, the only one of the big three memory suppliers based in America, says its HBM3E solution for the data center offers 50% more capacity than the competition while consuming 30% less energy. The company's new HBM4 solution is now in production, and it offers a further 60% increase in capacity alongside a 20% boost in energy efficiency. It will be a component in Nvidia's new Vera Rubin platform, which is expected to be the leading data center architecture when it ships later this year.
But Nvidia is experiencing so much demand for Vera Rubin that one memory supplier isn't enough, so it will also use HBM4 chips from SK Hynix and Samsung. Such is the state of the memory industry right now.
Other types of storage, such as solid state drives and flash drives, have their own places in the data center universe, and they, too, are in high demand right now, because AI not only consumes but also creates truckloads of data. Sandisk is one of the top suppliers of those types of storage devices, and its data center business produced a whopping 645% year-over-year revenue growth during its most recent quarter.
Don't bet the farm on the Roundhill ETF
Memory is likely to remain in high demand for the foreseeable future, but I don't think investors should chase the recent 127% gain in the Roundhill Memory ETF. Simply put, the imbalance between supply and demand is giving companies like Micron incredible pricing power, which is juicing their earnings and making their valuations look cheap -- but their growth isn't sustainable.
NYSEMKT: DRAM
Key Data Points
For example, Micron produced earnings of $8.29 per share in 2025, but that number is expected to explode to $105.28 per share in 2027, according to Wall Street's consensus forecast (provided by Yahoo! Finance). That gives Micron stock a forward P/E ratio of just 9.2, meaning that despite its recent gains, it would still have to more than triple just to trade in line with the technology-heavy Nasdaq-100 index, which has a P/E ratio of 35.2.
From that perspective, Micron looks like an absolute bargain. However, memory companies are frantically building more manufacturing capacity, which will boost supply and help stabilize prices over the next couple of years. Once production and demand are more in sync, Micron could actually see its earnings decline, and it is unlikely that it will be able to maintain its record profit margins. In that light, the stock might be more expensive today than it appears. The same goes for SK Hynix and Samsung.
And while the supply issue is slowly being resolved, cracks are simultaneously forming on the demand side as AI companies charge their customers more for their services to offset their rising costs. Anthropic recently introduced a new tokenizer that is more aggressive in calculating customers' use of its AI models. Therefore, although the company didn't technically raise any of its prices, its customers were hit with a significant increase in costs.
This was part of the reason Uber Technologies burned through its entire 2026 AI budget in just four months, and the company's chief operating officer said it was becoming harder to justify the current level of spending. This should worry investors who are piling into AI chip stocks right now.
If memory stocks eventually suffer a sharp correction, which I feel is almost inevitable, the Roundhill Memory ETF will crash alongside them. As a result, this probably isn't the best time to be buying it.
