This article first appeared on GuruFocus.
The artificial intelligence boom is beginning to expose a critical pressure point across the semiconductor industry: memory chips. Market research firm IDC has described the current shortage as a crisis like no other, as AI development accelerates and technology companies prepare to spend roughly $650 billion on computing infrastructure in 2026, about 80% higher than the previous year’s record. Executives across the industry have already started flagging the issue. Leaders at companies including Apple (NASDAQ:AAPL), Alphabet (NASDAQ:GOOG) and Tesla (NASDAQ:TSLA) have discussed how tight memory supply could influence profitability and AI timelines, while Tesla Chief Executive Officer Elon Musk raised the possibility on a late-January earnings call that the company could potentially explore producing its own memory chips.
The strain reflects how central memory has become to modern AI systems. Memory chips do not perform calculations themselves but store and feed data to processors, making them essential for everything from smartphones to large-scale data centers. AI servers increasingly rely on high-bandwidth memory, or HBM, a technology that stacks multiple layers of memory vertically and places them closer to processors to accelerate data movement. Compared with conventional DDR5 memory, an HBM3 chip can transfer data roughly ten times faster, a speed advantage that may become critical as AI models grow larger and demand enormous volumes of data to be processed without bottlenecks.
Demand is rising much faster than supply can adjust. Data-center usage already accounted for around 50% of global DRAM consumption in 2025, up from 32% five years earlier, according to Bloomberg Intelligence, and projections suggest AI servers could represent more than 60% of consumption by 2030. Yet the global memory industry is dominated by only three manufacturers Samsung Electronics (SSNLF), SK Hynix (HXSCL) and Micron Technology (NASDAQ:MU) and expanding production is both expensive and time-consuming. High-bandwidth memory is particularly challenging to manufacture because it requires stacking extremely thin silicon dies and drilling thousands of microscopic channels through them with precision, meaning capacity additions may take years to deliver meaningful supply.
