SEOUL–Inside Samsung Electronics’ memory fabrication plants, the metric engineers watch most closely shifted this summer. The company’s HBM4 production yield, a measure of how many of the advanced memory chips coming off its lines pass inspection, climbed to nearly 80 percent in recent weeks, according to people familiar with the company’s operations. That level, which industry engineers call the “golden yield,” arrived about half a year after the company first shipped the chips to customers.
The improvement has been rapid. When Samsung began HBM4 volume production in February, yields sat below 60 percent, a level that raised questions about whether the company could keep pace with rival SK Hynix, the dominant supplier of high-bandwidth memory to Nvidia. By August, the yield had climbed toward the target the company had set for year-end, a pace of improvement that outpaced internal projections and caught the attention of analysts who track the memory sector.
The numbers now point to a more aggressive commercial push. Samsung plans to more than triple its HBM4 revenue in the third quarter compared with the second quarter, according to people familiar with the plans. The company also aims for HBM4 to account for more than 60 percent of its total HBM revenue in the second half of the year, a shift that would mark a decisive move of its product mix toward the newest generation of memory.
By year-end, Samsung wants its share of the HBM market to approach its share of the DRAM market, roughly 38 percent, the people said. That ambition, if realized, would redraw the competitive map of a market that analysts estimate will be worth tens of billions of dollars this year, driven by demand from data-center operators building out artificial-intelligence infrastructure.
Samsung’s progress lands at a moment of unusual turbulence for its main rival. SK Hynix, the company that has led the HBM market for two years and supplies most of the chips that go into Nvidia’s AI accelerators, is locked in labor negotiations with its union over a stock-incentive program. Talks have stalled, with the union demanding a broader allocation of shares tied to the company’s record profits, according to people familiar with the discussions. The standoff has not halted production, but it has injected uncertainty into the operations of a company that is simultaneously racing to expand capacity.
For years, SK Hynix held a commanding position. It was first to mass-produce HBM3 and HBM3E, won early qualification with Nvidia, and built a reputation for delivering the highest-yielding stacks in the industry. Samsung, by contrast, spent much of 2025 fighting quality issues in its earlier HBM generations, missing qualification windows and watching customers sign long-term contracts with its rival.
The yield data now circulating in the industry suggests that gap is closing. Analysts who follow the memory sector said Samsung’s ability to push HBM4 yields to 80 percent within six months of production start would put it in a position to win qualification at additional customers and to price aggressively for capacity that Nvidia and other chip designers need urgently.
“Once you are at 80 percent yields on a chip this complex, the economics change completely,” said one semiconductor analyst who tracks the memory industry. “Samsung can now afford to sell into the market at scale, and that puts pressure on everyone else’s pricing and allocation.”
The stakes extend beyond the two Korean companies. High-bandwidth memory has become the scarcest component in the AI supply chain, with every major accelerator design requiring eight or more stacks. Prices have held at premium levels all year, and customers have signed multi-year agreements to secure supply. A second supplier with competitive yields would ease that bottleneck and, over time, pressure prices.
Samsung’s own financial results reflect the opportunity. The company’s memory division posted record operating profits in the second quarter, driven by DRAM pricing that has surged amid the memory super-cycle. HBM4, which sells at a significant premium to conventional DRAM, is expected to add further momentum in the second half.
The company is also expanding capacity to match its yield gains. Samsung has been converting portions of its Pyeongtaek and Hwaseong campuses to HBM production and has ordered additional advanced packaging equipment to assemble the stacks, according to people familiar with its plans. Packaging capacity, not just silicon yield, has been the binding constraint for HBM suppliers all year.
For SK Hynix, the response has been to accelerate its own HBM4 ramp. The company began volume shipments in the first half and has told customers it will roughly double its HBM capacity next year. Its newest fabs in Cheongju are dedicated to the technology.
The labor dispute adds an unwelcome complication. SK Hynix’s union has been vocal about sharing the spoils of the AI boom, noting that the company’s market value has more than tripled since early 2023. Management has resisted the union’s demand for a broader grant tied to future performance targets, people familiar with the talks said.
Investors have noticed the divergence. Samsung’s shares have outperformed SK Hynix’s since late July, when reports of the yield improvement first circulated.
Neither company expects the competitive balance to resolve quickly. HBM4 qualification cycles run several quarters, and customers tend to dual-source to protect supply. But the direction of travel has shifted: for the first time in two years, the industry’s conversation about who leads high-bandwidth memory includes a credible challenge to the incumbent.
What happens next depends on the numbers. If Samsung holds yields above 80 percent through the fourth quarter and delivers on its revenue targets, the HBM market will have two suppliers of scale for the first time since the AI boom began. If yields slip, SK Hynix’s lead endures. The fab floors in Pyeongtaek and Cheongju will settle the question.
