Tech companies have raced to build out compute capacity to fuel their AI ambitions but are now faced with a new bottleneck: memory capacity. The crunch comes as workloads shift from training models to using AI tools, and it’s driven in part by agentic AI, where a system can execute tasks independently. AI agents require greater memory to support context and continuous learning. ” Memory sits in a capacity-constrained cycle with unusually long order visibility driven by AI inference. For 2026, the risk is execution and transition, not demand,” Morgan Stanley analysts wrote in a note released late Thursday. Analysts at the investment bank see a steeper pricing climb and “favourable conditions” through 2027 as supply attempts to catch up with demand. “Multiples have expanded, but we think stock calls can still work with much higher earnings upside from here,” they wrote. “Bottlenecks are the winners – buy memory and semicap, especially EUV,” they added, referring to the abbreviation for a critical part of AI infrastructure. Here are Morgan Stanley’s 10 top stocks to play the memory bottleneck: Dynamic random-access memory – Samsung, Micron and SK Hynix Dynamic random-access memory, known as DRAM, is a specific component used in AI data centers. It saw a huge price spike in 2025 and is widely expected to rise again this year. Morgan Stanley analysts named Samsung , Micron and SK Hynix as ones to watch with higher pricing power in the DRAM space. It gave South Korea’s Samsung an upside of 18%, as it benefits from a better commodity cycle driven by AI and market share gains in high memory chips. Analysts expect an 12.2% upside from SK Hynix, also headquartered in South Korea, while U.S. company Micron was tipped for a 5% upside. Legacy Memory – Winbond “The supply-demand gap for legacy memory is widening further,” specifically for the DDR4/3, NOR and SLC/MLC NAND” computer chips, which are manufactured by the likes of Samsung, SK Hynix, and Micron, the note said. The analysts noted some forecasts expect DDR4 pricing to increase as much as 93-98% quarter-over-quarter in the first quarter of 2026. They named Taiwan’s Winbond as a top pick to take advantage of this, though analysts said Nanya Tech , Macronix , Longsys, AP Memory , GigaDevice and PSMC are also “all positioned to benefit from the current setup.” Storage – Western Digital U.S. company Western Digital could benefit from higher pricing power in HDDs and enterprise NAND — two types of memory — as greater demand means AI workloads are moved to cheaper memory stores, the note said. “A rising AI tide lifts all boats for both HDD and eSSD,” analysts wrote. They see an upside of 6% for Western Digital. Advanced packaging – Disco Japan’s Disco manufactures processing equipment and tools for the creation of computer chips, specializing in grinding and polishing equipment, which is part of the advanced packaging process. Advanced packaging refers to the actual combination of components to create chips, including high-capacity parts such as high bandwidth memory (HBM), which used in AI. Analysts said that Disco is poised to benefit from rapid growth in HBM products, giving it a 24.4% upside. Semicap – Applied Materials U.S. company Applied Materials , which supplies equipment needed to make semiconductors, is “exposed to the most visible growth drivers,” the Morgan Stanley analysts wrote. They specifically noted the build-out of DRAM capacity. Dutch chip stock ASM International was also named as a top pick, which the investment bank expects to benefit from the overall memory cycle. EUV – ASML Demand for extreme ultraviolet radiation, a niche but critical part of AI infrastructure known as EUV, is set to intensify. EUV is like a powerful laser printer that etches minuscule designs onto silicon wafers, which are then used to make semiconductors. Dutch company ASML has had a monopoly on EUV thus far and is set to benefit from increased EUV layer count, Morgan Stanley analysts said, giving it an upside of 21.80%
https://www.cnbc.com/2026/01/20/morgan-stanley-ai-memory-stock-picks.html
