Samsung achieves 80% manufacturing yield on HBM4 memory while Nvidia and AMD consider downgrading specifications amid supply crisis.

Samsung Electronics has achieved what the industry calls "golden yield" in the 80% range for its sixth-generation High Bandwidth Memory (HBM4), positioning itself for an aggressive supply expansion in the second half of the year. The yield, which was below 60% at the start of mass production in February, has stabilized within six months, rapidly shifting the HBM market battle from technological development to a production capacity war.

Samsung's HBM4 yield has recently reached the 80% mark—a critical inflection point in semiconductor manufacturing where defect rates are minimized, stable output is secured, and profitability becomes viable. "The initial year-end yield target was 80%, but as mass production volumes ramped up in the second half, the pace of process improvement exceeded expectations," according to industry sources.

Samsung's faster-than-expected yield stabilization stems from a "turnkey" system where its memory, foundry, and advanced packaging divisions collaborate from the design stage. The company has overcome the limitations of the Thermal Compression Non-Conductive Film (TC-NCF) process—previously considered the primary hurdle to achieving high yields—successfully controlling speed, power consumption, and heat simultaneously.

With stabilized yields in place, Samsung is accelerating its market offensive for the second half. Kim Jae-joon, Executive Vice President of Samsung's Memory Business, stated during the Q2 earnings conference call that "HBM4 revenue in the third quarter will expand more than threefold compared to the previous quarter," and that "in the second half, HBM4 revenue will comfortably exceed 60% of total HBM revenue." Samsung has reportedly set a goal to raise its HBM market share to approximately 38%—equivalent to its DRAM market share—by year-end.

Global investment bank UBS projected that while SK Hynix will maintain the top spot this year with a 48% share based on HBM bit shipments, Samsung will narrowly overtake SK Hynix's 39% share next year by capturing 41%.

SK Hynix is leveraging its experience in mass-producing every generation of HBM since the first, along with strong customer trust, as its greatest competitive advantages. Kim Ki-tae, Vice President of HBM Sales & Marketing at SK Hynix, emphasized during the Q2 conference call: "HBM4 competitiveness is complete only when it encompasses the capability for mass supply with stable yields and quality. Current mass production yields and quality are approaching levels close to the previous generation (HBM3E) products, which have entered a mature stage." Industry sources indicate SK Hynix's HBM4 yield has also climbed into the 80% range.

Alongside yield stabilization, Samsung is accelerating facility expansion, primarily at its Hwaseong and Pyeongtaek campuses, to meet explosive demand from global Big Tech firms. At the Hwaseong Campus Line 1 (H1), the company is expanding a general-purpose DRAM end-fab using idle cleanroom space. At the Pyeongtaek Campus Line 4 (P4), it is building a production line for the latest 10-nanometer-class 6th generation (1c) DRAM for HBM. Construction is also underway simultaneously for new fabs, Line 5 (P5-1) and Line 6 (P5-2), on new sites.

The true HBM supremacy battle between Samsung Electronics and SK Hynix is expected to be decided in the 7th generation HBM4E market, with mass production anticipated next year. Song Jae-hyuk, Samsung Electronics' Chief Technology Officer (CTO) and President of the Semiconductor Research Center, stated at an internal management briefing in late June: "The HBM4E reliability test yield has risen to over 70%, and the next-generation 10-nanometer-class 7th generation (D1d) DRAM process has also secured an advantage over competitors." Samsung had already sent HBM4E 12-high samples to major customers at the end of May.

Despite Samsung Electronics and SK Hynix expanding HBM4 supply, market demand still significantly outstrips supply. Consequently, global AI chip makers like NVIDIA and AMD are pivoting to a "de-spec" strategy, lowering the memory specifications for their next-generation accelerators.

NVIDIA has tested at least three prototype versions of its next-generation AI chip, "Rubin Ultra," to address the HBM shortage. While Rubin Ultra was initially planned to feature the highest-performance HBM4E 12-high stacks, NVIDIA is reportedly strongly considering prioritizing HBM4E 8-high or the previous generation HBM4 8-high and 12-high products instead.

AMD is also planning to launch its next-generation AI accelerator, the "MI400," in both HBM4 8-high and 12-high models. During a Q2 earnings conference call, AMD CEO Lisa Su hinted at this flexible adoption policy: "If high-capacity memory does not provide a clear advantage from a Total Cost of Ownership (TCO) perspective in specific workloads, we can adjust memory capacity to meet customer needs."

Downgrading memory specifications inevitably leads to a drop in chip performance. Even for the same 12-high product, HBM4 offers lower capacity and data transfer bandwidth compared to HBM4E, and lags in power efficiency, thermal control, and operational stability. Despite these trade-offs, AI chip manufacturers are making this choice because the physical production capacity limits for next-generation memory show little sign of easing.

According to JEDEC specifications, HBM4E offers approximately 38% higher bandwidth than HBM4. Market research firm TrendForce noted: "With the DRAM supply shortage expected to persist into 2027, the wafer production capacity that memory makers can allocate to HBM production is extremely limited." The analysis suggests that because the highly complex HBM4E 12-high product faces stringent qualification processes and uncertain yield ramp-up, the AI chip ecosystem has preemptively begun readjusting memory specifications.

Eugene Investment & Securities assessed NVIDIA's move as "a strategy judged to be aimed at producing more AI chips with limited production volume." If memory capacity decreases, Big Tech companies needing to run large AI models might ironically have to purchase even more AI chips.

The "de-spec" trend spreading among chip makers is expected to strengthen the position of South Korean memory suppliers. As the supply shortage intensifies, the pricing power of Samsung Electronics and SK Hynix in the HBM market could strengthen further. TrendForce predicted that "HBM suppliers will maintain pricing power through next year," adding that "the industry overall is already anticipating a significant rise in HBM prices." The trend of adopting HBM4 as an alternative for next-generation AI accelerators is also positive for both Samsung and SK Hynix.

Additionally, China's CXMT is holding its pricing floor, neutralizing Big Tech's strategy of cross-adopting cheaper Chinese components and boosting the negotiating power of South Korean memory firms. A recent example is Apple's attempt to negotiate mobile DRAM supply pricing with CXMT to reduce manufacturing costs, only for CXMT to refuse the price cut.