On July 10, SK Hynix CEO Kwak Noh-jung stood before reporters and made a statement that would have sounded absurd two years ago. The world is heading into the worst memory supply shortage in history, he said, and demand will outstrip supply well beyond 2030. His company's response: 100 trillion won, roughly $64 billion, to build new fabrication and advanced packaging plants in South Korea. One facility, M17, will not even begin operations until the first half of 2029.
The market barely blinked. SK Hynix shares had already surged 13% the day before after pricing the largest-ever US IPO by a foreign company at $26.5 billion on the Nasdaq. Reporting this remarkable achievement, CNN called it, "A once-obscure chip maker has landed the largest US listing by a foreign company". Samsung, meanwhile, had just posted a preliminary Q2 operating profit of 89.4 trillion won ($58 billion), an 18-fold year-over-year leap that Bloomberg described as a "record profit" that still "failed to impress" investors. When a 1,700% earnings surge counts as a disappointment, something fundamental has shifted.
That something is high-bandwidth memory, or HBM, the specialized stack of DRAM chips that sits beside every Nvidia GPU and is now the single most consequential bottleneck in the global AI supply chain. TSMC's CoWoS advanced packaging capacity, which bonds HBM to logic dies, is fully booked through 2026 with lead times of 40 to 52 weeks. Nvidia books more than 50% of that capacity for its Blackwell and upcoming Vera Rubin platforms. Micron has sold out its entire 2026 HBM output under fixed-price contracts and moved HBM4 volume production forward by a quarter. SK Hynix, the market leader, finalized its 2026 allocation months ago and expects tightness to extend through 2027 at minimum.
The numbers are staggering. Data centers are projected to consume 70% of all memory chips produced globally in 2026, up from a fraction of that just three years ago. Gartner forecasts that DRAM prices will rise 125% and NAND flash prices will surge 234% for the full year 2026. TrendForce reported that NAND contract prices alone jumped 70 to 75% quarter-over-quarter in Q2 2026, building on a 90% surge in Q1. The global NAND market hit a record $46 billion in Q1 2026, a 3.5x year-over-year increase. SemiAnalysis, the chip industry research firm, has called this a "once-in-four-deades" memory supercycle.
What makes this cycle structurally different from every memory boom and bust before it is that AI does not follow consumer electronics demand patterns. Smartphone and PC sales are cyclical and price-elastic. AI infrastructure buildout is not. Alphabet, Amazon, Meta, and Microsoft are collectively spending $650 billion to $725 billion on capex in 2026 alone. Google raised its projection by $5 billion in its latest earnings call. Meta added $10 billion. None of these companies have indicated any intention to slow down . The IEA projects data center electricity consumption will nearly double from 485 TWh in 2025 to 950 TWh by 2030, and Reuters reported July 7 that US power demand will hit successive record highs in 2026 and 2027 as AI use accelerates.
The cascading consequences are already visible. AMD expects DDR5 prices to remain elevated until 2028 because manufacturing capacity has pivoted toward HBM, leaving conventional DRAM undersupplied. Gaming GPUs are caught in the crossfire, as shared packaging capacity and memory allocation get absorbed by AI orders. The HBM total addressable market is projected to reach $100 billion by 2028, up from roughly $35 billion in 2025, but supply cannot scale fast enough to meet that trajectory. New fabrication plants require three to five years from groundbreaking to volume production. SK Hynix's M17 fab, announced this week, will not come online for nearly three years. Samsung and SK Hynix together control two-thirds of global memory output, and their existing plants in Gyeonggi Province have "hit the wall on power, water, and room to grow," according to a Reuters report, prompting a $518 billion investment in four new southwest fabrication sites.
This is where the story stops being about semiconductors and starts being about the limits of physical infrastructure. The U.S. Energy Information Administration's July 7 projection of back-to-back record power demand years was driven primarily by data center load growth. Former Google CEO Eric Schmidt testified before Congress that data centers will need 29 gigawatts of additional power by 2027 and 67 more gigawatts by 2030. The World Economic Forum noted in April 2026 that AI-driven electricity demand is renewing interest in small modular nuclear reactors and raising energy security concerns in the Strait of Hormuz. A single next-generation AI training cluster can consume more electricity than a mid-size American city. Regarding the ever-increasing energy demand and the specter of an energy crisis, Mindplex had been covering the story extensively, with the following as good examples: Giulio's report on 'The pursuit of fusion energy', Linsey's report titled 'AI Must Earn ''Social Permission'' For Its Vast Energy Use', and GizmoGuru's alarming call, 'Generative AI’s Hidden Environmental Toll'.There is no Moore's Law for power grids!
The computing industry is beginning to respond. Silicon photonics, which replaces electrical interconnects with light-based communication, entered what industry analysts call a "breakout era" in 2026, with Tower Semiconductor tripling its silicon photonics capacity and TSMC integrating co-packaged optics into its roadmap. Neuromorphic and photonic computing research is accelerating as an alternative to traditional von Neumann architectures. The Beyond Silicon 2026 conference in Bavaria showcased quantum, neuromorphic, and biological computing as potential post-silicon paradigms. These technologies are years or decades from commercial scale, but their development trajectories have sharpened considerably precisely because the current path looks unsustainable.
So what comes next? Three things are likely, and none of them are comfortable.
- First, memory prices will remain elevated through at least 2028, with HBM commanding premium margins that crowd out every other segment. Samsung's warning that shortages will intensify in 2027 and Micron's admission that meaningful output increases will not arrive until 2028 both point to a multi-year crunch.
- Second, the shortage will spread. When 70% of global memory output feeds data centers, everything else, from smartphones to automobiles to industrial equipment, competes for the remaining 30%!
- Third, the geopolitical dimensions will intensify. China's CXMT is scaling to roughly 350,000 wafers per month by end of 2026, challenging the Samsung-SK Hynix-Micron oligopoly, while Beijing simultaneously weighs export restrictions on its own advanced AI models. The US-Chip dispute over AI accelerators is about to become a US-China-Korea dispute over memory.
SK Hynix's $64 billion bet is the clearest signal yet that the memory industry believes this is not a cycle. It is a structural transformation. The company that just raised $26.5 billion on Wall Street is not building capacity for a boom that will bust. It is building capacity for a world in which memory, not logic, is the scarcest resource in computing.
The chip the world forgot is the chip the world now cannot live without.