SAN FRANCISCO—South Korean chipmakers SK Hynix and Samsung Electronics each saw their shares climb more than 20 percent, their strongest single-day performance in more than a year, as investors priced in the durable economics of AI memory demand.
The gains reflect surging demand for high-bandwidth memory chips, the specialized stacks that sit atop Nvidia's flagship GPUs and make large-scale AI inference economically viable. Without HBM, the compute economics of training and running large language models collapse—it is the bandwidth bottleneck, not raw processing power, that constrains AI throughput.
SK Hynix holds an estimated 50 percent share of the global HBM market; Samsung controls roughly 40 percent. Both are primary suppliers to Nvidia, AMD and Intel.
SK Hynix has reportedly sold out its HBM3E production through the end of 2026, a backlog that gives the company two years of revenue visibility—unusual in a sector defined by boom-bust inventory cycles. Samsung is aggressively expanding HBM capacity, aiming to double output by late 2027.
The South Korean rally tracked gains in U.S. technology shares. Nvidia rose 2.9 percent to $200.75. Microsoft gained 3.0 percent to $464.72. Alphabet climbed 6.7 percent to $356.13. Amazon surged 15.3 percent to $271.58.
Both companies are committing capital to match the demand. Samsung plans to invest $15 billion this year in its memory division, with a significant portion directed toward HBM capacity. SK Hynix's capital expenditure for 2026 is projected to exceed $12 billion.
The manufacturing barriers are real. HBM production requires advanced packaging—stacking multiple DRAM dies with through-silicon vias—and precise wafer fabrication that few fabs can execute at yield. That technical difficulty keeps the market locked in a two-player structure with pricing power to match.
HBM also carries materially better unit economics than conventional DRAM. Gross margins on HBM products can exceed 40 percent, well above the commodity memory average, lifting blended margins across both companies' portfolios.
Some analysts caution that the current spending rates are not guaranteed to hold. The semiconductor industry remains cyclical, and potential overcapacity could emerge in late 2027 if enterprise AI adoption slows.
Both companies are developing HBM4, the next generation of high-bandwidth memory, with commercialization expected in 2027. HBM4 is designed to deliver higher bandwidth at lower power consumption, extending the current infrastructure upgrade cycle.
Upcoming earnings reports will clarify HBM's revenue contribution. Samsung is scheduled to report third-quarter results in Oct.; SK Hynix typically releases results around the same period.
The HBM buildout also lifts key equipment suppliers—companies providing advanced packaging tools, lithography machines and wafer fabrication equipment—extending the capital spending wave across the semiconductor supply chain.
Samsung currently trades at 1.5 times book value; SK Hynix trades at 2.1 times, a premium that reflects investor expectations for sustained HBM pricing power as AI infrastructure spending continues.
