Wall Street has spent much of the artificial intelligence boom focused on processors and the enormous data centers being built around them. The next constraint, however, may come from a less visible part of the AI infrastructure chain. Memory chips are becoming increasingly difficult to secure as spending on AI servers accelerates, potentially setting up one of the tightest supply environments the industry has seen.
That puts Samsung Electronics and SK Hynix back in focus after shares of the two South Korean chipmakers fell about 38% from their highs over the past three months. Samsung Electronics is one of the world’s largest producers of DRAM and NAND flash memory, while SK Hynix is a leading supplier of high bandwidth memory used in AI servers. Their ability to expand output increasingly matters to the broader AI buildout because modern servers require far more than processors alone.
Large cloud operators are increasing spending as AI services, model hosting, agentic AI and token based businesses begin generating more revenue. KB Securities, a major South Korean brokerage, estimates global AI infrastructure investment could rise 60% from a year earlier to $1.3 trillion in 2027. That expansion is expected to direct a much larger share of capital toward memory chips, making supply from companies such as Samsung and SK Hynix increasingly important to the pace of the global data center buildout.
KB Securities estimates memory semiconductors will account for 14% of AI infrastructure spending in 2025, 40% in 2026 and 57% in 2027. TrendForce, a global technology market research firm, projects an even higher 68% share for 2027.
The problem is that available supply is already tightening. Memory inventories at Samsung Electronics and SK Hynix have fallen below 10 days in the third quarter, according to KB Securities. The brokerage said next year could bring the tightest memory supply conditions on record, raising the possibility that manufacturers could run short of chips available for sale.
The squeeze is no longer limited to high bandwidth memory. AI servers consume HBM alongside server grade DDR5 memory and enterprise solid state drives, pushing demand higher across several categories at once. KB Securities expects growth in DRAM and NAND demand next year to exceed supply growth by more than 10 percentage points on a bit basis.
The transition to HBM4 could make conventional memory even tighter. KB Securities estimates HBM4 requires roughly three times as much wafer production capacity as standard DRAM. As manufacturers dedicate more capacity to advanced AI memory, less remains available for conventional DRAM.
That creates an unusual gap between industry conditions and stock performance. Samsung Electronics and SK Hynix are entering a period in which memory could become increasingly scarce, yet their shares have fallen sharply enough that KB Securities estimates they trade at roughly three times projected 2027 earnings.
The brokerage expects both companies to set record earnings over the next three years while continuing significant shareholder returns. The key question is whether AI infrastructure spending remains strong enough to keep memory demand ahead of supply.
If that happens, the semiconductor shortage tied to AI may no longer be defined mainly by processors. Memory could become the next major bottleneck, giving Samsung Electronics and SK Hynix a more central role in determining how quickly the global AI infrastructure boom can expand.
