Samsung unveils next-generation HBM reshaping AI memory landscape
Samsung Electronics has unveiled "zHBM," a three-dimensional memory architecture that stacks high-bandwidth memory (HBM) on top of graphics processing units (GPUs), the world's first technology to resolve the data bottleneck in AI chips. The innovation was unveiled at FMS 2026 in the United States. When mass-produced, zHBM is expected to deliver up to eight times the performance of eighth-generation HBM5 by eliminating the performance penalty previously caused by physical distance between GPU and HBM in conventional AI accelerators. Kim Kyung-ryun, executive vice president of Samsung Electronics' DRAM Development Division, stated: "Samsung is back," emphasizing the company's technological breakthrough in overcoming the Memory Wall. Samsung also unveiled complementary architectures: z-NAND-O, optimized for on-device AI, and V10 BV-NAND, a tenth-generation V-NAND stacked more than 400 layers.
Beyond HBM, Samsung and SK hynix are accelerating development of processing-in-memory (PIM), which integrates memory and computing functions, and Compute Express Link (CXL), which boosts connection speed. Samsung has unveiled the industry's first low-power double data rate (LPDDR)5X-PIM and plans detailed presentation at next month's Hot Chips conference. Competing efforts from U.S. Micron, Japan's Kioxia, Intel, and Qualcomm, alongside China's CXMT, are intensifying the race to shift the HBM paradigm.
The scale of big tech's data center infrastructure boom is creating hidden financial liabilities. The five largest U.S. tech companies—Alphabet, Microsoft, Amazon, Meta, and Oracle—have signed long-term data center lease contracts containing $1.09 trillion in unrecognized debt, roughly four times the $285 billion already reflected in financial statements. Under accounting standards, lease debt is recognized only when facilities begin operation, leaving data centers not yet operational disclosed only in footnotes. Oracle faces the largest burden, with $260 billion in unstarted leases—approximately seven times its currently recognized lease debt.
China's technology companies are pursuing an export strategy centered on platform expansion rather than product sales alone. As export price indices fell from 114.5 in 2022 to 99 in 2024, Chinese manufacturers are shifting profit models by exporting industrial platforms. Construction equipment maker XCMG, for example, has exported its industrial internet platform "Hanyun" to Kazakhstan and Brazil alongside machinery sales, generating recurring revenue through remote management services. China's Ministry of Industry and Information Technology reports that industries applying industrial internet have increased added value by 28.4% over four years, suggesting the nation aims to expand its leadership in global technology standards through platform exports.
The United States is implementing export controls to exclude China from advanced supply chains including AI data center components, robots, solar materials, and semiconductors. The Federal Communications Commission is preparing legislation to ban imports of Chinese-made optical transceivers, while the Trump administration is reviewing price floors and tariffs on solar panels and polysilicon. Recent measures have included import bans on foreign-made power inverters, humanoid robots, and tariffs on products involving forced labor. China has responded by strictly controlling exports of drone components and related technologies to the U.S.
Global semiconductor researcher Yole Group forecasts that Korea will maintain its position as the top player in the memory chip market through 2031, citing aggressive investment from Samsung Electronics and SK hynix alongside Korean government plans to double wafer production within five years. The competitive landscape is shifting amid U.S.-China competition: U.S. memory chip consumption rose from 37% in 2021 to 52 percent last year, while China's share fell from 35% to 29%. Nvidia's dominance in HBM is narrowing as major technology companies expand adoption of proprietary application-specific integrated circuits (ASICs), with Nvidia's market share expected to decline to approximately 58% this year.