Memory Chip Makers Invest $525B, But Relief Won't Come Until 2028

AI expansion hits a memory wall

The artificial intelligence boom is running headlong into a fundamental constraint: there aren't enough memory chips to go around. DRAM and NAND flash memory—the technologies that store data for AI workloads—have been in increasingly short supply for the past two years, with manufacturers reporting they're sold out through at least 2027.

Prices have surged as a result, creating a serious bottleneck for new AI data centers and consumer applications. Now the industry's largest players are responding with massive investments in new production capacity, though the relief won't arrive quickly.

Why it matters

Memory shortages represent a hard infrastructure limit on AI deployment, not a software problem that can be patched or optimized away. Organizations planning AI initiatives need to factor in extended lead times and elevated costs for memory components through 2027. The multi-year lag between investment announcements and actual production also signals that current supply-demand imbalances will shape the competitive landscape—favoring companies with existing capacity allocations and long-term supplier relationships.

Record investments across Asia and the US

In a joint venture, Samsung and SK Hynix announced they will invest a combined 800 trillion won ($518 billion) to build four new memory manufacturing facilities and a high-bandwidth memory (HBM) packaging hub in Korea's Southwest Region. Together, the two companies produce roughly two-thirds of the world's DRAM chips. SK Hynix's chair noted that building up the company's current manufacturing capability in Gyeonggi Province took nine years, suggesting a similar timeline for the new sites.

SK Hynix is also establishing its first US production facility in West Lafayette, Indiana—a $4 billion advanced packaging and R&D site focused on HBM. Mass production there is scheduled for late 2028.

Meanwhile, SanDisk and Kioxia have begun production of their 10th-generation 3D NAND flash memory at a facility in Kitakami, Japan. The 332-layer technology delivers 1Tb TLC memory density and 4.8 Gb/s interface speeds, representing a 33 percent performance improvement over their eighth-generation product. The manufacturing partnership between SanDisk and Kioxia has been extended through 2034.

Micron announced $3 billion in US supply chain investments, including a $500 million commitment to GlobalWafers for a 300mm silicon wafer facility in Sherman, Texas, backed by a 10-year supply agreement. Micron has fabrication facilities under construction in Idaho and New York.

The timeline problem

Despite these investments totaling more than $525 billion, the industry faces a stark reality: building semiconductor fabs takes years. Micron CEO Sanjay Mehrotra warned that meaningful large-scale new capacity is unlikely before 2028 due to lengthy construction timelines. The company is currently meeting only 50 to 67 percent of core customer demand.

Industry observers expect significant new manufacturing capacity won't come online until late 2027 at the earliest, with 2028 more realistic for most facilities. Until then, memory supply will remain the limiting factor for AI infrastructure expansion.

These details were first reported by Thomas Coughlin in Forbes.