AI + ML
Memory shortages may signal deeper challenges for AI infrastructure expansion
The memory semiconductor sector is currently experiencing a period of exceptional growth. Driven by the rapid expansion of AI datacenters, SK Hynix and Micron have seen their revenues triple over the past year, while Samsung’s have roughly doubled.
Despite this prosperity, the industry’s historical pattern of boom-and-bust cycles suggests that the current upswing may eventually reverse. The three dominant suppliers owe their present fortunes to the AI revolution, yet the structural dynamics of memory markets remain precarious.
Demand for high-bandwidth memory (HBM), DDR5, and NAND flash used in GPU servers has absorbed available production capacity, creating shortages that elevated prices across consumer electronics and AI infrastructure alike. Even entry-level smartphones have become difficult to procure.
In response, the leading memory manufacturers are committing hundreds of billions of dollars to new fabrication facilities. In June, South Korean President Lee Jae Myung announced a $576 billion investment led by SK Hynix and Samsung to expand chip output and secure AI supply chains.
Recently, Micron revealed plans to invest up to $3 billion to reinforce the U.S. semiconductor supply chain, while also increasing production at its sites in Singapore, Taiwan, and Japan.
However, expanding capacity is a slow endeavor. Semiconductor manufacturing is among the most complex and capital-intensive industries globally. Constructing a new DRAM or NAND wafer fab requires securing financing, selecting sites, obtaining permits, and installing extensive support systems for power, air handling, and ultrapure water.
After cleanroom completion, hundreds of millions of dollars in specialized lithography, transport, and testing equipment must be installed and validated. Achieving stable yields can take months, with total project timelines often spanning years. Consequently, fabs initiated today will require at least three years to become operational and longer to reach full output.
Industry analyses, including a recent IDC report, indicate that memory prices will remain elevated through at least 2028. This sustains inflated revenues for producers but imposes heavier infrastructure costs on AI startups and model developers.
Over the past four years, organizations such as OpenAI have deployed vast venture capital to build advanced models and agents. The focus has shifted from feasibility to whether returns justify continued expenditure. High memory costs erode profit margins and complicate per-token economics.
The critical uncertainty is whether memory vendors can add capacity before AI firms exhaust their funded runways. Historically, memory has been a commodity with extreme price volatility. Manufacturers leverage boom periods to finance fabs, aware that new supply may eventually depress prices.
As reported late last year, the AI surge has altered typical cycles: instead of declining through 2025–2026, prices have risen as AI systems consume available DRAM and NAND. Should anticipated AI demand falter, the resulting downturn could be severe for all stakeholders.
Eventually, easing memory costs may alleviate pressure on consumer device affordability, though current conditions remain challenging.