The first thing you notice when you walk through the server rooms of a large AI data center is the heat, which is caused by cooling systems constantly battling the thermal output of thousands of GPUs processing data in parallel. The memory that sits next to those CPUs is less obvious but possibly more important: High Bandwidth Memory, or HBM, which is stacked in layers and so closely attached to the chip that the distance between them is measured in microns. An AI model can transfer data quickly enough to prevent the GPU from being idle thanks to HBM. Additionally, as of 2026, it is one of the most restricted parts of the whole AI supply chain, and the businesses who produce it are reporting revenue growth that tends to draw significant investment interest.
Unlike most technology sectors, the AI memory industry is structurally concentrated. The three businesses that can produce advanced HBM on a large scale are SK Hynix, Samsung Electronics, and Micron Technology. That’s all. The three incumbents are enjoying pricing leverage that is unavailable in the majority of commodity semiconductor markets because the engineering requirements and capital expenditure required to develop the manufacturing facilities are so high that new entrants are not a credible short-term threat. When it comes to increasing HBM3E capacity, SK Hynix has been the most aggressive, and as a result, its margins have increased. Over the past 18 months, investors who were prepared to put up with the volatility associated with direct exposure to chip companies have benefited.
The Roundhill Memory ETF, ticker DRAM, provides diversified exposure to HBM, DRAM, and NAND producers without requiring a view on which of the three IDMs will capture the largest margin in any given quarter, making it an option for investors who are less comfortable selecting individual names in a sector that can swing 30% on a single earnings call. It’s a lower-maintenance method to take part in what some analysts are referring to as a memory supercycle, but the term carries some risk because supercycles come to an end and the memory market has historically had periods of great shortage interspersed with periods of equally high excess. The basic economics of the chip industry are unaffected by the AI demand driver. In this case, position size may be more important than choosing a specific name.
The equipment and supply chain layers are included in the investment case in addition to the chip manufacturers. Advanced packaging cannot be manufactured without ASML’s lithography machines. In order to integrate HBM with GPU dies at the density required for AI workloads, TSMC’s packaging capabilities are crucial. Large AI clusters put demands on interconnect infrastructure for data throughput, which benefits Corning’s optical fiber and signal management businesses like Broadcom and Marvell farther down the value chain. Companies that profit from the AI memory build-out but do not bear the direct commodity price risk as IDMs do are the picks and shovels.
It’s difficult to ignore the fact that when the AI workload itself changes, so does the investment discourse surrounding AI memory. Large models require a lot of memory bandwidth to train, while inference, or using those models in production at the edge, demands less power, tighter integration, and a smaller form factor. Businesses like SiMa.ai and Semidynamics are working toward that architectural reality. The direction of the demand is unknown, but it is still unclear if they become significant public market investments or are absorbed into bigger players before that occurs. The memory bottleneck that characterized AI infrastructure in 2024 and 2025 remains. Investment strategies around it are becoming increasingly diverse.
